microsoft
/
CNTK
зеркало из
1
0
Форкнуть 0
Код Задачи Пакеты Проекты Релизы Вики Активность

Merge branch 'master' of https://github.com/Microsoft/CNTK into dongyu/rollbackfix

This commit is contained in:
Dong Yu 2016-05-19 10:50:58 -07:00
Родитель 872c9fcdb6 76c4bd8717
Коммит 2f2ca021c9
226 изменённых файлов: 66985 добавлений и 30233 удалений
Показать статистику Скачать Patch файл Скачать Diff файл Показать все файлы Свернуть все файлы

4
.gitattributes поставляемый
Просмотреть файл

@ -40,6 +40,10 @@ run-test-common text eol=lf
run-timit-test-common text eol=lf
make_binary_drop_linux text eol=lf
# Used from Unix / Cygwin 'md5sum -c', needs to have LF line endings:
Tests/EndToEndTests/Examples/Speech/TIMIT/WriteBottleneck/expected_output_md5sum.*.txt eol=lf
Tests/EndToEndTests/Examples/Speech/TIMIT/WriteScaledLogLike/expected_output_md5sum.*.txt eol=lf
Makefile text
*.sln text
*.vcxproj text

230
CNTK.sln
Просмотреть файл

@ -976,38 +976,148 @@ Project("{8BC9CEB8-8B4A-11D0-8D11-00A0C91BC942}") = "CompositeDataReader", "Sour
{F0A9637C-20DA-42F0-83D4-23B4704DE602} = {F0A9637C-20DA-42F0-83D4-23B4704DE602}
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "CNTKTextFormatReader", "CNTKTextFormatReader", "{B1110F99-A307-4745-B464-7FD75951645A}"
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "CNTKTextFormatReader", "CNTKTextFormatReader", "{99FAAACE-C360-43CF-B706-20621F164484}"
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "ParallelTraining", "ParallelTraining", "{369656B8-DDFD-412E-901B-DFEBCC31ABE0}"
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Examples", "Examples", "{629761D1-7A05-409A-B62B-FC1CCC0D6EED}"
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Image", "Image", "{D4302516-C77F-4FAF-82FB-18DB39F5A53B}"
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "ParallelTraining", "ParallelTraining", "{06BE675D-80DD-419A-8E00-26953EF11F25}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\SimpleMultiGPU.cntk = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\SimpleMultiGPU.cntk
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Speech", "Speech", "{CCEFD1F9-E843-43E0-B127-EF73EF90582D}"
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Speech", "Speech", "{5642F047-490B-4ABD-8113-8563C872B39F}"
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "NoQuantization", "NoQuantization", "{751BEA27-2187-4BE5-82E7-A3668CFCE7A9}"
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Image", "Image", "{2B6CCAB6-A92A-483C-9FDB-8412FA4DC42F}"
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "DoublePrecision", "DoublePrecision", "{4084DC9D-0A53-4029-9C86-92AF243C2E09}"
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Other", "Other", "{225F5A3A-7CAF-4C71-9143-3AD2AC4D47A3}"
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "MNIST", "MNIST", "{EBD36FD9-FE5B-420E-A572-DC6117300DB3}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\baseline.cpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\baseline.cpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\baseline.gpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\baseline.gpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\baseline.windows.cpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\baseline.windows.cpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\baseline.windows.gpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\baseline.windows.gpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\run-test = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\run-test
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\testcases.yml = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\testcases.yml
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\run-test-common = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\run-test-common
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "SinglePrecision", "SinglePrecision", "{ACB91DED-FB23-4FF0-A1A1-EBE56B783EFC}"
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Miscellaneous", "Miscellaneous", "{08D284FA-2914-4B35-A89C-896DBA2B4484}"
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "CIFAR-10", "CIFAR-10", "{95FAC6A0-6AE7-4947-9DFD-498FE71311AD}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\baseline.cpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\baseline.cpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\baseline.gpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\baseline.gpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\baseline.windows.cpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\baseline.windows.cpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\baseline.windows.gpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\baseline.windows.gpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\run-test = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\run-test
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\testcases.yml = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\testcases.yml
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\run-test-common = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\run-test-common
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Simple", "Simple", "{BCA65A0C-D93B-4F90-81B1-73048DE04DF1}"
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Config", "Config", "{A877E526-89C1-422E-9F90-4DDE84135A36}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\Config\01_Conv.cntk = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\Config\01_Conv.cntk
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\Config\02_BatchNormConv.cntk = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\Config\02_BatchNormConv.cntk
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\Config\05_ConvLocal.cntk = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\Config\05_ConvLocal.cntk
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "01_Convolution", "01_Convolution", "{071D8449-D080-4141-869D-600CC3C2A0BE}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\01_Convolution\baseline.linux.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\01_Convolution\baseline.linux.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\01_Convolution\baseline.windows.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\01_Convolution\baseline.windows.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\01_Convolution\README.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\01_Convolution\README.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\01_Convolution\run-test = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\01_Convolution\run-test
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\01_Convolution\testcases.yml = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\01_Convolution\testcases.yml
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "02_BatchNormConv", "02_BatchNormConv", "{D3A74C52-BC74-4DA3-BE93-8F4241D54EE0}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\02_BatchNormConv\baseline.linux.gpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\02_BatchNormConv\baseline.linux.gpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\02_BatchNormConv\baseline.windows.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\02_BatchNormConv\baseline.windows.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\02_BatchNormConv\README.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\02_BatchNormConv\README.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\02_BatchNormConv\run-test = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\02_BatchNormConv\run-test
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\02_BatchNormConv\testcases.yml = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\02_BatchNormConv\testcases.yml
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "05_ConvLocal", "05_ConvLocal", "{EC466625-BC66-41DF-B55A-EB28AFABE24E}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\05_ConvLocal\baseline.linux.gpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\05_ConvLocal\baseline.linux.gpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\05_ConvLocal\baseline.windows.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\05_ConvLocal\baseline.windows.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\05_ConvLocal\README.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\05_ConvLocal\README.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\05_ConvLocal\run-test = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\05_ConvLocal\run-test
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\05_ConvLocal\testcases.yml = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10\05_ConvLocal\testcases.yml
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "01_OneHidden", "01_OneHidden", "{34D578DB-0101-45C4-9DF0-37DE9AB87C65}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\01_OneHidden\baseline.linux.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\01_OneHidden\baseline.linux.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\01_OneHidden\baseline.windows.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\01_OneHidden\baseline.windows.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\01_OneHidden\README.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\01_OneHidden\README.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\01_OneHidden\run-test = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\01_OneHidden\run-test
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\01_OneHidden\testcases.yml = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\01_OneHidden\testcases.yml
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "02_Convolution", "02_Convolution", "{1FE04815-E02E-498C-B276-6D058D46D754}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\02_Convolution\baseline.linux.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\02_Convolution\baseline.linux.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\02_Convolution\baseline.windows.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\02_Convolution\baseline.windows.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\02_Convolution\README.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\02_Convolution\README.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\02_Convolution\run-test = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\02_Convolution\run-test
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\02_Convolution\testcases.yml = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\02_Convolution\testcases.yml
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "03_ConvBatchNorm", "03_ConvBatchNorm", "{2A125ED5-9C8A-4BDF-A200-862104289608}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\03_ConvBatchNorm\baseline.linux.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\03_ConvBatchNorm\baseline.linux.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\03_ConvBatchNorm\baseline.windows.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\03_ConvBatchNorm\baseline.windows.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\03_ConvBatchNorm\README.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\03_ConvBatchNorm\README.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\03_ConvBatchNorm\run-test = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\03_ConvBatchNorm\run-test
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\03_ConvBatchNorm\testcases.yml = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\03_ConvBatchNorm\testcases.yml
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Config", "Config", "{E9207003-B860-4D57-B2CA-09AF52FF191F}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\Config\01_OneHidden.cntk = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\Config\01_OneHidden.cntk
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\Config\01_OneHidden.ndl = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\Config\01_OneHidden.ndl
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\Config\02_Convolution.cntk = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\Config\02_Convolution.cntk
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\Config\02_Convolution.ndl = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\Config\02_Convolution.ndl
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\Config\03_ConvBatchNorm.cntk = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\Config\03_ConvBatchNorm.cntk
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\Config\03_ConvBatchNorm.ndl = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\Config\03_ConvBatchNorm.ndl
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\Config\Macros.ndl = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Image\MNIST\Config\Macros.ndl
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Simple2d", "Simple2d", "{50420947-E502-40B4-8739-2C0BADD93BEE}"
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "MultiGpu", "MultiGpu", "{935E5A95-888D-4922-AB5A-E9C11D65E974}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\MultiGpu\baseline.linux.cpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\MultiGpu\baseline.linux.cpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\MultiGpu\baseline.linux.gpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\MultiGpu\baseline.linux.gpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\MultiGpu\baseline.windows.cpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\MultiGpu\baseline.windows.cpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\MultiGpu\baseline.windows.gpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\MultiGpu\baseline.windows.gpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\MultiGpu\README.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\MultiGpu\README.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\MultiGpu\run-test = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\MultiGpu\run-test
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\MultiGpu\testcases.yml = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\MultiGpu\testcases.yml
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Simple", "Simple", "{773313DD-69DD-463F-ADC9-E8A902A5223C}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\Simple\baseline.linux.cpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\Simple\baseline.linux.cpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\Simple\baseline.linux.gpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\Simple\baseline.linux.gpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\Simple\baseline.windows.cpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\Simple\baseline.windows.cpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\Simple\baseline.windows.gpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\Simple\baseline.windows.gpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\Simple\README.txt = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\Simple\README.txt
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\Simple\run-test = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\Simple\run-test
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\Simple\testcases.yml = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\Simple\testcases.yml
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Config", "Config", "{C8E2EF3B-CCBF-4BDD-8127-2252626FB22B}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\Config\Multigpu.cntk = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\Config\Multigpu.cntk
Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\Config\Simple.cntk = Tests\EndToEndTests\CNTKTextFormatReader\Examples\Other\Simple2d\Config\Simple.cntk
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "QuickE2E", "QuickE2E", "{A4F79A83-DE30-40FA-88F4-86304C89AC7F}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\baseline.linux.txt = Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\baseline.linux.txt
Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\baseline.windows.txt = Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\baseline.windows.txt
Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\Image_QuickE2E.cntk = Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\Image_QuickE2E.cntk
Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\README.txt = Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\README.txt
Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\run-test = Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\run-test
Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\testcases.yml = Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\testcases.yml
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Simple", "Simple", "{CC47AF62-2558-455F-81CB-36901AF033B0}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\Speech\Simple\baseline.linux.cpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\Speech\Simple\baseline.linux.cpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\Speech\Simple\baseline.linux.gpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\Speech\Simple\baseline.linux.gpu.txt
@ -1019,6 +1129,33 @@ Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Simple", "Simple", "{BCA65A
Tests\EndToEndTests\CNTKTextFormatReader\Speech\Simple\testcases.yml = Tests\EndToEndTests\CNTKTextFormatReader\Speech\Simple\testcases.yml
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "NoQuantization", "NoQuantization", "{1BA5209D-3EB6-48E7-BE8A-0622315070C0}"
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Data", "Data", "{AA14A8DB-669D-447B-A97F-8B726BF30188}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\Data\SimpleDataTrain.txt = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\Data\SimpleDataTrain.txt
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "SinglePrecision", "SinglePrecision", "{CA248859-AA91-47D6-AC05-3542AB27E290}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\baseline.cpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\baseline.cpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\baseline.gpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\baseline.gpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\baseline.windows.cpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\baseline.windows.cpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\baseline.windows.gpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\baseline.windows.gpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\run-test = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\run-test
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\testcases.yml = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\SinglePrecision\testcases.yml
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "DoublePrecision", "DoublePrecision", "{8B6E9318-5ED0-49BF-945B-072E0D90A886}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\baseline.cpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\baseline.cpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\baseline.gpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\baseline.gpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\baseline.windows.cpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\baseline.windows.cpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\baseline.windows.gpu.txt = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\baseline.windows.gpu.txt
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\run-test = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\run-test
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\testcases.yml = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\NoQuantization\DoublePrecision\testcases.yml
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "SparseDSSM", "SparseDSSM", "{1FB54750-B668-4AC3-966F-ED504020AC06}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\Text\SparseDSSM\baseline.cpu.txt = Tests\EndToEndTests\Text\SparseDSSM\baseline.cpu.txt
@ -1031,23 +1168,6 @@ Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "SparseDSSM", "SparseDSSM",
Tests\EndToEndTests\Text\SparseDSSM\testcases.yml = Tests\EndToEndTests\Text\SparseDSSM\testcases.yml
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Image", "Image", "{59980D6E-1732-4809-B17C-6EF4B4F5CF8B}"
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "QuickE2E", "QuickE2E", "{6FB9C411-C19A-403A-94C2-F7DF393F7612}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\baseline.linux.txt = Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\baseline.linux.txt
Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\baseline.windows.txt = Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\baseline.windows.txt
Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\Image_QuickE2E.cntk = Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\Image_QuickE2E.cntk
Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\README.txt = Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\README.txt
Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\run-test = Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\run-test
Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\testcases.yml = Tests\EndToEndTests\CNTKTextFormatReader\Image\QuickE2E\testcases.yml
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "Data", "Data", "{C30742B5-4DBE-4D80-B429-901856E4043D}"
ProjectSection(SolutionItems) = preProject
Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\Data\SimpleDataTrain.txt = Tests\EndToEndTests\CNTKTextFormatReader\ParallelTraining\Data\SimpleDataTrain.txt
EndProjectSection
EndProject
Project("{2150E333-8FDC-42A3-9474-1A3956D46DE8}") = "WriteCommand", "WriteCommand", "{3E9BD61F-1F0A-4966-BE17-803AEFD1DFA4}"
ProjectSection(SolutionItems) = preProject
tests\endtoendtests\Speech\DNN\WriteCommand\baseline.cpu.txt = tests\endtoendtests\Speech\DNN\WriteCommand\baseline.cpu.txt
@ -1526,19 +1646,37 @@ Global
{EC780385-7580-4D15-914B-1D878A295CBC} = {E53E63A0-FAA9-4416-9AD1-08A8FB87FEE1}
{D11F76CC-DB6D-4CB4-B3B7-AB139DE2F5FA} = {E53E63A0-FAA9-4416-9AD1-08A8FB87FEE1}
{181664AC-4C95-4798-A923-09B879215B33} = {8656B71D-E24C-4AC2-8BE4-C07B415A3E15}
{99FAAACE-C360-43CF-B706-20621F164484} = {6E565B48-1923-49CE-9787-9BBB9D96F4C5}
{629761D1-7A05-409A-B62B-FC1CCC0D6EED} = {99FAAACE-C360-43CF-B706-20621F164484}
{D4302516-C77F-4FAF-82FB-18DB39F5A53B} = {99FAAACE-C360-43CF-B706-20621F164484}
{06BE675D-80DD-419A-8E00-26953EF11F25} = {99FAAACE-C360-43CF-B706-20621F164484}
{5642F047-490B-4ABD-8113-8563C872B39F} = {99FAAACE-C360-43CF-B706-20621F164484}
{2B6CCAB6-A92A-483C-9FDB-8412FA4DC42F} = {629761D1-7A05-409A-B62B-FC1CCC0D6EED}
{225F5A3A-7CAF-4C71-9143-3AD2AC4D47A3} = {629761D1-7A05-409A-B62B-FC1CCC0D6EED}
{EBD36FD9-FE5B-420E-A572-DC6117300DB3} = {2B6CCAB6-A92A-483C-9FDB-8412FA4DC42F}
{08D284FA-2914-4B35-A89C-896DBA2B4484} = {2B6CCAB6-A92A-483C-9FDB-8412FA4DC42F}
{95FAC6A0-6AE7-4947-9DFD-498FE71311AD} = {08D284FA-2914-4B35-A89C-896DBA2B4484}
{A877E526-89C1-422E-9F90-4DDE84135A36} = {95FAC6A0-6AE7-4947-9DFD-498FE71311AD}
{071D8449-D080-4141-869D-600CC3C2A0BE} = {95FAC6A0-6AE7-4947-9DFD-498FE71311AD}
{D3A74C52-BC74-4DA3-BE93-8F4241D54EE0} = {95FAC6A0-6AE7-4947-9DFD-498FE71311AD}
{EC466625-BC66-41DF-B55A-EB28AFABE24E} = {95FAC6A0-6AE7-4947-9DFD-498FE71311AD}
{34D578DB-0101-45C4-9DF0-37DE9AB87C65} = {EBD36FD9-FE5B-420E-A572-DC6117300DB3}
{1FE04815-E02E-498C-B276-6D058D46D754} = {EBD36FD9-FE5B-420E-A572-DC6117300DB3}
{2A125ED5-9C8A-4BDF-A200-862104289608} = {EBD36FD9-FE5B-420E-A572-DC6117300DB3}
{E9207003-B860-4D57-B2CA-09AF52FF191F} = {EBD36FD9-FE5B-420E-A572-DC6117300DB3}
{50420947-E502-40B4-8739-2C0BADD93BEE} = {225F5A3A-7CAF-4C71-9143-3AD2AC4D47A3}
{935E5A95-888D-4922-AB5A-E9C11D65E974} = {50420947-E502-40B4-8739-2C0BADD93BEE}
{773313DD-69DD-463F-ADC9-E8A902A5223C} = {50420947-E502-40B4-8739-2C0BADD93BEE}
{C8E2EF3B-CCBF-4BDD-8127-2252626FB22B} = {50420947-E502-40B4-8739-2C0BADD93BEE}
{A4F79A83-DE30-40FA-88F4-86304C89AC7F} = {D4302516-C77F-4FAF-82FB-18DB39F5A53B}
{CC47AF62-2558-455F-81CB-36901AF033B0} = {5642F047-490B-4ABD-8113-8563C872B39F}
{1BA5209D-3EB6-48E7-BE8A-0622315070C0} = {06BE675D-80DD-419A-8E00-26953EF11F25}
{AA14A8DB-669D-447B-A97F-8B726BF30188} = {06BE675D-80DD-419A-8E00-26953EF11F25}
{CA248859-AA91-47D6-AC05-3542AB27E290} = {1BA5209D-3EB6-48E7-BE8A-0622315070C0}
{8B6E9318-5ED0-49BF-945B-072E0D90A886} = {1BA5209D-3EB6-48E7-BE8A-0622315070C0}
{86883653-8A61-4038-81A0-2379FAE4200A} = {DD043083-71A4-409A-AA91-F9C548DCF7EC}
{7B7A563D-AA8E-4660-A805-D50235A02120} = {33EBFE78-A1A8-4961-8938-92A271941F94}
{B1110F99-A307-4745-B464-7FD75951645A} = {6E565B48-1923-49CE-9787-9BBB9D96F4C5}
{369656B8-DDFD-412E-901B-DFEBCC31ABE0} = {B1110F99-A307-4745-B464-7FD75951645A}
{CCEFD1F9-E843-43E0-B127-EF73EF90582D} = {B1110F99-A307-4745-B464-7FD75951645A}
{751BEA27-2187-4BE5-82E7-A3668CFCE7A9} = {369656B8-DDFD-412E-901B-DFEBCC31ABE0}
{4084DC9D-0A53-4029-9C86-92AF243C2E09} = {751BEA27-2187-4BE5-82E7-A3668CFCE7A9}
{ACB91DED-FB23-4FF0-A1A1-EBE56B783EFC} = {751BEA27-2187-4BE5-82E7-A3668CFCE7A9}
{BCA65A0C-D93B-4F90-81B1-73048DE04DF1} = {CCEFD1F9-E843-43E0-B127-EF73EF90582D}
{1FB54750-B668-4AC3-966F-ED504020AC06} = {8656B71D-E24C-4AC2-8BE4-C07B415A3E15}
{59980D6E-1732-4809-B17C-6EF4B4F5CF8B} = {B1110F99-A307-4745-B464-7FD75951645A}
{6FB9C411-C19A-403A-94C2-F7DF393F7612} = {59980D6E-1732-4809-B17C-6EF4B4F5CF8B}
{C30742B5-4DBE-4D80-B429-901856E4043D} = {369656B8-DDFD-412E-901B-DFEBCC31ABE0}
{3E9BD61F-1F0A-4966-BE17-803AEFD1DFA4} = {6994C86D-A672-4254-824A-51F4DFEB807F}
{5560DDD4-1E6E-4F41-B9BD-F52A19DF0B31} = {6994C86D-A672-4254-824A-51F4DFEB807F}
{9834E864-A8CD-4D28-A3C9-F79FE0F421AE} = {6994C86D-A672-4254-824A-51F4DFEB807F}

3
Examples/Image/MNIST/Config/02_Convolution.cntk
Просмотреть файл

@ -26,7 +26,8 @@ ndlMacros = "$ConfigDir$/Macros.ndl"
traceLevel=1
numMBsToShowResult=500
prefetch=true
# Note: turn off prefetching; known to crash UCIFastReader occasionally.
prefetch=false
# If set to true, always initialize the network on CPU, making initialization consistent across CPU and GPU targets (for testing).
initOnCPUOnly=true

3
Examples/Image/Miscellaneous/CIFAR-10/01_Conv.cntk
Просмотреть файл

@ -16,7 +16,8 @@ imageLayout = "cudnn"
# If set to true, always initialize the network on CPU, making initialization consistent across CPU and GPU targets (for testing).
initOnCPUOnly=true
prefetch = "true"
# Note: turn off prefetching; known to crash UCIFastReader occasionally.
prefetch = "false"
command = Train:Test

3
Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv.cntk
Просмотреть файл

@ -16,7 +16,8 @@ imageLayout = "cudnn"
# If set to true, always initialize the network on CPU, making initialization consistent across CPU and GPU targets (for testing).
initOnCPUOnly=true
prefetch = "true"
# Note: turn off prefetching; known to crash UCIFastReader occasionally.
prefetch = "false"
command = Train:Test

3
Examples/Image/Miscellaneous/CIFAR-10/05_ConvLocal.cntk
Просмотреть файл

@ -13,7 +13,8 @@ imageLayout = "cudnn"
# override the above as follows when running on CPU:
# deviceId = -1
prefetch = "true"
# Note: turn off prefetching; known to crash UCIFastReader occasionally.
prefetch = "false"
command = Train:Test

16
Examples/Image/Miscellaneous/ImageNet/ResNet/README.md
Просмотреть файл

@ -1,12 +1,22 @@
# CNTK example: ImageNet ResNet
**Disclaimer: network configurations and experiment settings in this this folder try to follow those published in the [ResNet paper](http://arxiv.org/abs/1512.03385) as close as possible. However, these samples are NOT endorsed or verified by the researchers who published the original work. It is NOT guaranteed that you get the same (or even close) results as those in the paper.**
## Overview
|Data: |The ILSVRC2012 dataset (http://www.image-net.org/challenges/LSVRC/2012/) of images.
|:---------|:---
|Purpose |This example demonstrates usage of the NDL (Network Description Language) to define networks similar to ResNet.
|Network |NDLNetworkBuilder, deep convolutional networks resembling ResNet networks.
|Network |NDLNetworkBuilder, deep convolutional residual networks (ResNet).
|Training |Stochastic gradient descent with momentum.
## Details
The network configurations and experiment settings in this this folder resemble the ones in the original [ResNet paper](http://arxiv.org/abs/1512.03385) with few minor changes inspired by [this work](https://github.com/facebook/fb.resnet.torch).
The following table contains results as well as links to pre-trained models that can be used in various applications.
| Network | Top-1 error | Top-5 error | Model
| ------------- | ----------- | ----------- | ----------
| ResNet-18 | 29.57 | 10.41 | [Download](https://www.cntk.ai/resnet/ResNet_18.model)
| ResNet-34 | 27.31 | 8.97 | [Download](https://www.cntk.ai/resnet/ResNet_34.model)
| ResNet-50 | 24.74 | 7.56 | [Download](https://www.cntk.ai/resnet/ResNet_50.model)
## Notes
This work is an implementation of ResNets in CNTK. If you are interested in the original implementation of ResNet, follow [this link](https://github.com/KaimingHe/deep-residual-networks).

603
Examples/Other/Simple2d/Data/SimpleDataTest_cntk_text.txt Normal file
Просмотреть файл

@ -0,0 +1,603 @@
|labels 0 1 |features -1 -1
|labels 0 1 |features -1 -0.99
|labels 0 1 |features -1 -0.98
|labels 0 1 |features -1 -0.97
|labels 0 1 |features -1 -0.96
|labels 0 1 |features -1 -0.95
|labels 0 1 |features -1 -0.94
|labels 0 1 |features -1 -0.93
|labels 0 1 |features -1 -0.92
|labels 0 1 |features -1 -0.91
|labels 0 1 |features -1 -0.9
|labels 0 1 |features -1 -0.89
|labels 0 1 |features -1 -0.88
|labels 0 1 |features -1 -0.87
|labels 0 1 |features -1 -0.86
|labels 0 1 |features -1 -0.85
|labels 0 1 |features -1 -0.84
|labels 0 1 |features -1 -0.83
|labels 0 1 |features -1 -0.82
|labels 0 1 |features -1 -0.81
|labels 0 1 |features -1 -0.8
|labels 0 1 |features -1 -0.79
|labels 0 1 |features -1 -0.78
|labels 0 1 |features -1 -0.77
|labels 0 1 |features -1 -0.76
|labels 0 1 |features -1 -0.75
|labels 0 1 |features -1 -0.74
|labels 0 1 |features -1 -0.73
|labels 0 1 |features -1 -0.72
|labels 0 1 |features -1 -0.71
|labels 0 1 |features -1 -0.7
|labels 0 1 |features -1 -0.69
|labels 0 1 |features -1 -0.68
|labels 0 1 |features -1 -0.67
|labels 0 1 |features -1 -0.66
|labels 0 1 |features -1 -0.65
|labels 0 1 |features -1 -0.64
|labels 0 1 |features -1 -0.63
|labels 0 1 |features -1 -0.62
|labels 0 1 |features -1 -0.61
|labels 0 1 |features -1 -0.6
|labels 0 1 |features -1 -0.59
|labels 0 1 |features -1 -0.58
|labels 0 1 |features -1 -0.57
|labels 0 1 |features -1 -0.56
|labels 0 1 |features -1 -0.55
|labels 0 1 |features -1 -0.54
|labels 0 1 |features -1 -0.53
|labels 0 1 |features -1 -0.52
|labels 0 1 |features -1 -0.51
|labels 0 1 |features -1 -0.5
|labels 0 1 |features -1 -0.49
|labels 0 1 |features -1 -0.48
|labels 0 1 |features -1 -0.47
|labels 0 1 |features -1 -0.46
|labels 0 1 |features -1 -0.45
|labels 0 1 |features -1 -0.44
|labels 0 1 |features -1 -0.43
|labels 0 1 |features -1 -0.42
|labels 0 1 |features -1 -0.41
|labels 0 1 |features -1 -0.4
|labels 0 1 |features -1 -0.39
|labels 0 1 |features -1 -0.38
|labels 0 1 |features -1 -0.37
|labels 0 1 |features -1 -0.36
|labels 0 1 |features -1 -0.35
|labels 0 1 |features -1 -0.34
|labels 0 1 |features -1 -0.33
|labels 0 1 |features -1 -0.32
|labels 0 1 |features -1 -0.31
|labels 0 1 |features -1 -0.3
|labels 0 1 |features -1 -0.29
|labels 0 1 |features -1 -0.28
|labels 0 1 |features -1 -0.27
|labels 0 1 |features -1 -0.26
|labels 0 1 |features -1 -0.25
|labels 0 1 |features -1 -0.24
|labels 0 1 |features -1 -0.23
|labels 0 1 |features -1 -0.22
|labels 0 1 |features -1 -0.21
|labels 0 1 |features -1 -0.2
|labels 0 1 |features -1 -0.19
|labels 0 1 |features -1 -0.18
|labels 0 1 |features -1 -0.17
|labels 0 1 |features -1 -0.16
|labels 0 1 |features -1 -0.15
|labels 0 1 |features -1 -0.14
|labels 0 1 |features -1 -0.13
|labels 0 1 |features -1 -0.12
|labels 0 1 |features -1 -0.11
|labels 0 1 |features -1 -0.1
|labels 1 0 |features -1 -0.09
|labels 1 0 |features -1 -0.08
|labels 1 0 |features -1 -0.07
|labels 1 0 |features -1 -0.06
|labels 1 0 |features -1 -0.05
|labels 1 0 |features -1 -0.04
|labels 1 0 |features -1 -0.03
|labels 1 0 |features -1 -0.02
|labels 1 0 |features -1 -0.01
|labels 1 0 |features -1 0
|labels 1 0 |features -1 0.01
|labels 1 0 |features -1 0.02
|labels 1 0 |features -1 0.03
|labels 1 0 |features -1 0.04
|labels 1 0 |features -1 0.05
|labels 1 0 |features -1 0.06
|labels 1 0 |features -1 0.07
|labels 1 0 |features -1 0.08
|labels 1 0 |features -1 0.09
|labels 1 0 |features -1 0.1
|labels 1 0 |features -1 0.11
|labels 1 0 |features -1 0.12
|labels 1 0 |features -1 0.13
|labels 1 0 |features -1 0.14
|labels 1 0 |features -1 0.15
|labels 1 0 |features -1 0.16
|labels 1 0 |features -1 0.17
|labels 1 0 |features -1 0.18
|labels 1 0 |features -1 0.19
|labels 1 0 |features -1 0.2
|labels 1 0 |features -1 0.21
|labels 1 0 |features -1 0.22
|labels 1 0 |features -1 0.23
|labels 1 0 |features -1 0.24
|labels 1 0 |features -1 0.25
|labels 1 0 |features -1 0.26
|labels 1 0 |features -1 0.27
|labels 1 0 |features -1 0.28
|labels 1 0 |features -1 0.29
|labels 1 0 |features -1 0.3
|labels 1 0 |features -1 0.31
|labels 1 0 |features -1 0.32
|labels 1 0 |features -1 0.33
|labels 1 0 |features -1 0.34
|labels 1 0 |features -1 0.35
|labels 1 0 |features -1 0.36
|labels 1 0 |features -1 0.37
|labels 1 0 |features -1 0.38
|labels 1 0 |features -1 0.39
|labels 1 0 |features -1 0.4
|labels 1 0 |features -1 0.41
|labels 1 0 |features -1 0.42
|labels 1 0 |features -1 0.43
|labels 1 0 |features -1 0.44
|labels 1 0 |features -1 0.45
|labels 1 0 |features -1 0.46
|labels 1 0 |features -1 0.47
|labels 1 0 |features -1 0.48
|labels 1 0 |features -1 0.49
|labels 1 0 |features -1 0.5
|labels 1 0 |features -1 0.51
|labels 1 0 |features -1 0.52
|labels 1 0 |features -1 0.53
|labels 1 0 |features -1 0.54
|labels 1 0 |features -1 0.55
|labels 1 0 |features -1 0.56
|labels 1 0 |features -1 0.57
|labels 1 0 |features -1 0.58
|labels 1 0 |features -1 0.59
|labels 1 0 |features -1 0.6
|labels 1 0 |features -1 0.61
|labels 1 0 |features -1 0.62
|labels 1 0 |features -1 0.63
|labels 1 0 |features -1 0.64
|labels 1 0 |features -1 0.65
|labels 1 0 |features -1 0.66
|labels 1 0 |features -1 0.67
|labels 1 0 |features -1 0.68
|labels 1 0 |features -1 0.69
|labels 1 0 |features -1 0.7
|labels 1 0 |features -1 0.71
|labels 1 0 |features -1 0.72
|labels 1 0 |features -1 0.73
|labels 1 0 |features -1 0.74
|labels 1 0 |features -1 0.75
|labels 1 0 |features -1 0.76
|labels 1 0 |features -1 0.77
|labels 1 0 |features -1 0.78
|labels 1 0 |features -1 0.79
|labels 1 0 |features -1 0.8
|labels 1 0 |features -1 0.81
|labels 1 0 |features -1 0.82
|labels 1 0 |features -1 0.83
|labels 1 0 |features -1 0.84
|labels 1 0 |features -1 0.85
|labels 1 0 |features -1 0.86
|labels 1 0 |features -1 0.87
|labels 1 0 |features -1 0.88
|labels 1 0 |features -1 0.89
|labels 1 0 |features -1 0.9
|labels 1 0 |features -1 0.91
|labels 1 0 |features -1 0.92
|labels 1 0 |features -1 0.93
|labels 1 0 |features -1 0.94
|labels 1 0 |features -1 0.95
|labels 1 0 |features -1 0.96
|labels 1 0 |features -1 0.97
|labels 1 0 |features -1 0.98
|labels 1 0 |features -1 0.99
|labels 1 0 |features -1 0
|labels 0 1 |features 0 -1
|labels 0 1 |features 0 -0.99
|labels 0 1 |features 0 -0.98
|labels 0 1 |features 0 -0.97
|labels 0 1 |features 0 -0.96
|labels 0 1 |features 0 -0.95
|labels 0 1 |features 0 -0.94
|labels 0 1 |features 0 -0.93
|labels 0 1 |features 0 -0.92
|labels 0 1 |features 0 -0.91
|labels 0 1 |features 0 -0.9
|labels 0 1 |features 0 -0.89
|labels 0 1 |features 0 -0.88
|labels 0 1 |features 0 -0.87
|labels 0 1 |features 0 -0.86
|labels 0 1 |features 0 -0.85
|labels 0 1 |features 0 -0.84
|labels 0 1 |features 0 -0.83
|labels 0 1 |features 0 -0.82
|labels 0 1 |features 0 -0.81
|labels 0 1 |features 0 -0.8
|labels 0 1 |features 0 -0.79
|labels 0 1 |features 0 -0.78
|labels 0 1 |features 0 -0.77
|labels 0 1 |features 0 -0.76
|labels 0 1 |features 0 -0.75
|labels 0 1 |features 0 -0.74
|labels 0 1 |features 0 -0.73
|labels 0 1 |features 0 -0.72
|labels 0 1 |features 0 -0.71
|labels 0 1 |features 0 -0.7
|labels 0 1 |features 0 -0.69
|labels 0 1 |features 0 -0.68
|labels 0 1 |features 0 -0.67
|labels 0 1 |features 0 -0.66
|labels 0 1 |features 0 -0.65
|labels 0 1 |features 0 -0.64
|labels 0 1 |features 0 -0.63
|labels 0 1 |features 0 -0.62
|labels 0 1 |features 0 -0.61
|labels 0 1 |features 0 -0.6
|labels 0 1 |features 0 -0.59
|labels 0 1 |features 0 -0.58
|labels 0 1 |features 0 -0.57
|labels 0 1 |features 0 -0.56
|labels 0 1 |features 0 -0.55
|labels 0 1 |features 0 -0.54
|labels 0 1 |features 0 -0.53
|labels 0 1 |features 0 -0.52
|labels 0 1 |features 0 -0.51
|labels 0 1 |features 0 -0.5
|labels 0 1 |features 0 -0.49
|labels 0 1 |features 0 -0.48
|labels 0 1 |features 0 -0.47
|labels 0 1 |features 0 -0.46
|labels 0 1 |features 0 -0.45
|labels 0 1 |features 0 -0.44
|labels 0 1 |features 0 -0.43
|labels 0 1 |features 0 -0.42
|labels 0 1 |features 0 -0.41
|labels 0 1 |features 0 -0.4
|labels 0 1 |features 0 -0.39
|labels 0 1 |features 0 -0.38
|labels 0 1 |features 0 -0.37
|labels 0 1 |features 0 -0.36
|labels 0 1 |features 0 -0.35
|labels 0 1 |features 0 -0.34
|labels 0 1 |features 0 -0.33
|labels 0 1 |features 0 -0.32
|labels 0 1 |features 0 -0.31
|labels 0 1 |features 0 -0.3
|labels 0 1 |features 0 -0.29
|labels 0 1 |features 0 -0.28
|labels 0 1 |features 0 -0.27
|labels 0 1 |features 0 -0.26
|labels 0 1 |features 0 -0.25
|labels 0 1 |features 0 -0.24
|labels 0 1 |features 0 -0.23
|labels 0 1 |features 0 -0.22
|labels 0 1 |features 0 -0.21
|labels 0 1 |features 0 -0.2
|labels 0 1 |features 0 -0.19
|labels 0 1 |features 0 -0.18
|labels 0 1 |features 0 -0.17
|labels 0 1 |features 0 -0.16
|labels 0 1 |features 0 -0.15
|labels 0 1 |features 0 -0.14
|labels 0 1 |features 0 -0.13
|labels 0 1 |features 0 -0.12
|labels 0 1 |features 0 -0.11
|labels 0 1 |features 0 -0.1
|labels 1 0 |features 0 -0.09
|labels 1 0 |features 0 -0.08
|labels 1 0 |features 0 -0.07
|labels 1 0 |features 0 -0.06
|labels 1 0 |features 0 -0.05
|labels 1 0 |features 0 -0.04
|labels 1 0 |features 0 -0.03
|labels 1 0 |features 0 -0.02
|labels 1 0 |features 0 -0.01
|labels 1 0 |features 0 0
|labels 1 0 |features 0 0.01
|labels 1 0 |features 0 0.02
|labels 1 0 |features 0 0.03
|labels 1 0 |features 0 0.04
|labels 1 0 |features 0 0.05
|labels 1 0 |features 0 0.06
|labels 1 0 |features 0 0.07
|labels 1 0 |features 0 0.08
|labels 1 0 |features 0 0.09
|labels 1 0 |features 0 0.1
|labels 1 0 |features 0 0.11
|labels 1 0 |features 0 0.12
|labels 1 0 |features 0 0.13
|labels 1 0 |features 0 0.14
|labels 1 0 |features 0 0.15
|labels 1 0 |features 0 0.16
|labels 1 0 |features 0 0.17
|labels 1 0 |features 0 0.18
|labels 1 0 |features 0 0.19
|labels 1 0 |features 0 0.2
|labels 1 0 |features 0 0.21
|labels 1 0 |features 0 0.22
|labels 1 0 |features 0 0.23
|labels 1 0 |features 0 0.24
|labels 1 0 |features 0 0.25
|labels 1 0 |features 0 0.26
|labels 1 0 |features 0 0.27
|labels 1 0 |features 0 0.28
|labels 1 0 |features 0 0.29
|labels 1 0 |features 0 0.3
|labels 1 0 |features 0 0.31
|labels 1 0 |features 0 0.32
|labels 1 0 |features 0 0.33
|labels 1 0 |features 0 0.34
|labels 1 0 |features 0 0.35
|labels 1 0 |features 0 0.36
|labels 1 0 |features 0 0.37
|labels 1 0 |features 0 0.38
|labels 1 0 |features 0 0.39
|labels 1 0 |features 0 0.4
|labels 1 0 |features 0 0.41
|labels 1 0 |features 0 0.42
|labels 1 0 |features 0 0.43
|labels 1 0 |features 0 0.44
|labels 1 0 |features 0 0.45
|labels 1 0 |features 0 0.46
|labels 1 0 |features 0 0.47
|labels 1 0 |features 0 0.48
|labels 1 0 |features 0 0.49
|labels 1 0 |features 0 0.5
|labels 1 0 |features 0 0.51
|labels 1 0 |features 0 0.52
|labels 1 0 |features 0 0.53
|labels 1 0 |features 0 0.54
|labels 1 0 |features 0 0.55
|labels 1 0 |features 0 0.56
|labels 1 0 |features 0 0.57
|labels 1 0 |features 0 0.58
|labels 1 0 |features 0 0.59
|labels 1 0 |features 0 0.6
|labels 1 0 |features 0 0.61
|labels 1 0 |features 0 0.62
|labels 1 0 |features 0 0.63
|labels 1 0 |features 0 0.64
|labels 1 0 |features 0 0.65
|labels 1 0 |features 0 0.66
|labels 1 0 |features 0 0.67
|labels 1 0 |features 0 0.68
|labels 1 0 |features 0 0.69
|labels 1 0 |features 0 0.7
|labels 1 0 |features 0 0.71
|labels 1 0 |features 0 0.72
|labels 1 0 |features 0 0.73
|labels 1 0 |features 0 0.74
|labels 1 0 |features 0 0.75
|labels 1 0 |features 0 0.76
|labels 1 0 |features 0 0.77
|labels 1 0 |features 0 0.78
|labels 1 0 |features 0 0.79
|labels 1 0 |features 0 0.8
|labels 1 0 |features 0 0.81
|labels 1 0 |features 0 0.82
|labels 1 0 |features 0 0.83
|labels 1 0 |features 0 0.84
|labels 1 0 |features 0 0.85
|labels 1 0 |features 0 0.86
|labels 1 0 |features 0 0.87
|labels 1 0 |features 0 0.88
|labels 1 0 |features 0 0.89
|labels 1 0 |features 0 0.9
|labels 1 0 |features 0 0.91
|labels 1 0 |features 0 0.92
|labels 1 0 |features 0 0.93
|labels 1 0 |features 0 0.94
|labels 1 0 |features 0 0.95
|labels 1 0 |features 0 0.96
|labels 1 0 |features 0 0.97
|labels 1 0 |features 0 0.98
|labels 1 0 |features 0 0.99
|labels 1 0 |features 0 1
|labels 0 1 |features 1 -1
|labels 0 1 |features 1 -0.99
|labels 0 1 |features 1 -0.98
|labels 0 1 |features 1 -0.97
|labels 0 1 |features 1 -0.96
|labels 0 1 |features 1 -0.95
|labels 0 1 |features 1 -0.94
|labels 0 1 |features 1 -0.93
|labels 0 1 |features 1 -0.92
|labels 0 1 |features 1 -0.91
|labels 0 1 |features 1 -0.9
|labels 0 1 |features 1 -0.89
|labels 0 1 |features 1 -0.88
|labels 0 1 |features 1 -0.87
|labels 0 1 |features 1 -0.86
|labels 0 1 |features 1 -0.85
|labels 0 1 |features 1 -0.84
|labels 0 1 |features 1 -0.83
|labels 0 1 |features 1 -0.82
|labels 0 1 |features 1 -0.81
|labels 0 1 |features 1 -0.8
|labels 0 1 |features 1 -0.79
|labels 0 1 |features 1 -0.78
|labels 0 1 |features 1 -0.77
|labels 0 1 |features 1 -0.76
|labels 0 1 |features 1 -0.75
|labels 0 1 |features 1 -0.74
|labels 0 1 |features 1 -0.73
|labels 0 1 |features 1 -0.72
|labels 0 1 |features 1 -0.71
|labels 0 1 |features 1 -0.7
|labels 0 1 |features 1 -0.69
|labels 0 1 |features 1 -0.68
|labels 0 1 |features 1 -0.67
|labels 0 1 |features 1 -0.66
|labels 0 1 |features 1 -0.65
|labels 0 1 |features 1 -0.64
|labels 0 1 |features 1 -0.63
|labels 0 1 |features 1 -0.62
|labels 0 1 |features 1 -0.61
|labels 0 1 |features 1 -0.6
|labels 0 1 |features 1 -0.59
|labels 0 1 |features 1 -0.58
|labels 0 1 |features 1 -0.57
|labels 0 1 |features 1 -0.56
|labels 0 1 |features 1 -0.55
|labels 0 1 |features 1 -0.54
|labels 0 1 |features 1 -0.53
|labels 0 1 |features 1 -0.52
|labels 0 1 |features 1 -0.51
|labels 0 1 |features 1 -0.5
|labels 0 1 |features 1 -0.49
|labels 0 1 |features 1 -0.48
|labels 0 1 |features 1 -0.47
|labels 0 1 |features 1 -0.46
|labels 0 1 |features 1 -0.45
|labels 0 1 |features 1 -0.44
|labels 0 1 |features 1 -0.43
|labels 0 1 |features 1 -0.42
|labels 0 1 |features 1 -0.41
|labels 0 1 |features 1 -0.4
|labels 0 1 |features 1 -0.39
|labels 0 1 |features 1 -0.38
|labels 0 1 |features 1 -0.37
|labels 0 1 |features 1 -0.36
|labels 0 1 |features 1 -0.35
|labels 0 1 |features 1 -0.34
|labels 0 1 |features 1 -0.33
|labels 0 1 |features 1 -0.32
|labels 0 1 |features 1 -0.31
|labels 0 1 |features 1 -0.3
|labels 0 1 |features 1 -0.29
|labels 0 1 |features 1 -0.28
|labels 0 1 |features 1 -0.27
|labels 0 1 |features 1 -0.26
|labels 0 1 |features 1 -0.25
|labels 0 1 |features 1 -0.24
|labels 0 1 |features 1 -0.23
|labels 0 1 |features 1 -0.22
|labels 0 1 |features 1 -0.21
|labels 0 1 |features 1 -0.2
|labels 0 1 |features 1 -0.19
|labels 0 1 |features 1 -0.18
|labels 0 1 |features 1 -0.17
|labels 0 1 |features 1 -0.16
|labels 0 1 |features 1 -0.15
|labels 0 1 |features 1 -0.14
|labels 0 1 |features 1 -0.13
|labels 0 1 |features 1 -0.12
|labels 0 1 |features 1 -0.11
|labels 0 1 |features 1 -0.1
|labels 0 1 |features 1 -0.09
|labels 0 1 |features 1 -0.08
|labels 0 1 |features 1 -0.07
|labels 0 1 |features 1 -0.06
|labels 0 1 |features 1 -0.05
|labels 0 1 |features 1 -0.04
|labels 0 1 |features 1 -0.03
|labels 0 1 |features 1 -0.02
|labels 0 1 |features 1 -0.01
|labels 1 0 |features 1 0
|labels 1 0 |features 1 0.01
|labels 1 0 |features 1 0.02
|labels 1 0 |features 1 0.03
|labels 1 0 |features 1 0.04
|labels 1 0 |features 1 0.05
|labels 1 0 |features 1 0.06
|labels 1 0 |features 1 0.07
|labels 1 0 |features 1 0.08
|labels 1 0 |features 1 0.09
|labels 1 0 |features 1 0.1
|labels 1 0 |features 1 0.11
|labels 1 0 |features 1 0.12
|labels 1 0 |features 1 0.13
|labels 1 0 |features 1 0.14
|labels 1 0 |features 1 0.15
|labels 1 0 |features 1 0.16
|labels 1 0 |features 1 0.17
|labels 1 0 |features 1 0.18
|labels 1 0 |features 1 0.19
|labels 1 0 |features 1 0.2
|labels 1 0 |features 1 0.21
|labels 1 0 |features 1 0.22
|labels 1 0 |features 1 0.23
|labels 1 0 |features 1 0.24
|labels 1 0 |features 1 0.25
|labels 1 0 |features 1 0.26
|labels 1 0 |features 1 0.27
|labels 1 0 |features 1 0.28
|labels 1 0 |features 1 0.29
|labels 1 0 |features 1 0.3
|labels 1 0 |features 1 0.31
|labels 1 0 |features 1 0.32
|labels 1 0 |features 1 0.33
|labels 1 0 |features 1 0.34
|labels 1 0 |features 1 0.35
|labels 1 0 |features 1 0.36
|labels 1 0 |features 1 0.37
|labels 1 0 |features 1 0.38
|labels 1 0 |features 1 0.39
|labels 1 0 |features 1 0.4
|labels 1 0 |features 1 0.41
|labels 1 0 |features 1 0.42
|labels 1 0 |features 1 0.43
|labels 1 0 |features 1 0.44
|labels 1 0 |features 1 0.45
|labels 1 0 |features 1 0.46
|labels 1 0 |features 1 0.47
|labels 1 0 |features 1 0.48
|labels 1 0 |features 1 0.49
|labels 1 0 |features 1 0.5
|labels 1 0 |features 1 0.51
|labels 1 0 |features 1 0.52
|labels 1 0 |features 1 0.53
|labels 1 0 |features 1 0.54
|labels 1 0 |features 1 0.55
|labels 1 0 |features 1 0.56
|labels 1 0 |features 1 0.57
|labels 1 0 |features 1 0.58
|labels 1 0 |features 1 0.59
|labels 1 0 |features 1 0.6
|labels 1 0 |features 1 0.61
|labels 1 0 |features 1 0.62
|labels 1 0 |features 1 0.63
|labels 1 0 |features 1 0.64
|labels 1 0 |features 1 0.65
|labels 1 0 |features 1 0.66
|labels 1 0 |features 1 0.67
|labels 1 0 |features 1 0.68
|labels 1 0 |features 1 0.69
|labels 1 0 |features 1 0.7
|labels 1 0 |features 1 0.71
|labels 1 0 |features 1 0.72
|labels 1 0 |features 1 0.73
|labels 1 0 |features 1 0.74
|labels 1 0 |features 1 0.75
|labels 1 0 |features 1 0.76
|labels 1 0 |features 1 0.77
|labels 1 0 |features 1 0.78
|labels 1 0 |features 1 0.79
|labels 1 0 |features 1 0.8
|labels 1 0 |features 1 0.81
|labels 1 0 |features 1 0.82
|labels 1 0 |features 1 0.83
|labels 1 0 |features 1 0.84
|labels 1 0 |features 1 0.85
|labels 1 0 |features 1 0.86
|labels 1 0 |features 1 0.87
|labels 1 0 |features 1 0.88
|labels 1 0 |features 1 0.89
|labels 1 0 |features 1 0.9
|labels 1 0 |features 1 0.91
|labels 1 0 |features 1 0.92
|labels 1 0 |features 1 0.93
|labels 1 0 |features 1 0.94
|labels 1 0 |features 1 0.95
|labels 1 0 |features 1 0.96
|labels 1 0 |features 1 0.97
|labels 1 0 |features 1 0.98
|labels 1 0 |features 1 0.99
|labels 1 0 |features 1 1

10000
Examples/Other/Simple2d/Data/SimpleDataTrain_cntk_text.txt Normal file
Просмотреть файл

Разница между файлами не показана из-за своего большого размера Загрузить разницу

1
Makefile
Просмотреть файл

@ -247,6 +247,7 @@ READER_SRC =\
$(SOURCEDIR)/Readers/ReaderLib/TruncatedBpttPacker.cpp \
$(SOURCEDIR)/Readers/ReaderLib/PackerBase.cpp \
$(SOURCEDIR)/Readers/ReaderLib/FramePacker.cpp \
$(SOURCEDIR)/Readers/ReaderLib/ChunkCache.cpp \
COMMON_SRC =\
$(SOURCEDIR)/Common/Config.cpp \

6
README.md
Просмотреть файл

@ -1,6 +1,10 @@
# CNTK
## Latest news
*2016-05-16.* An example illustrating [Using CNTK with ResNet](https://github.com/Microsoft/CNTK/tree/master/Examples/Image/Miscellaneous/ImageNet/ResNet) is added to the codebase. The example contains some pre-trained models that can be used in various applications.
*2016-05-16.* CNTK Wiki now has [FAQ Page](https://github.com/Microsoft/CNTK/wiki/CNTK-FAQ)
*2016-05-05.* CNTK now supports *BlockMomentum* Stochastic Gradient Descent (SGD) algorithm.
See the details in the [Multiple GPUs and machines Wiki section](https://github.com/Microsoft/CNTK/wiki/Multiple-GPUs-and-machines)
@ -10,8 +14,6 @@ See the description in the [Image Reader Wiki section](https://github.com/Micros
*2016-04-25.* V 1.1 Binary release
CNTK v.1.1 binaries are on the [CNTK Releases page](https://github.com/Microsoft/CNTK/releases/tag/v1.1)
*2016-04-12.* CNTK is available as [Azure Virtual Machines](https://github.com/Microsoft/CNTK/wiki/CNTK-on-Azure) and [Docker Containers](https://github.com/Microsoft/CNTK/wiki/CNTK-Docker-Containers)
See [all news](https://github.com/Microsoft/CNTK/wiki/News).
## What is CNTK

2
Source/ComputationNetworkLib/ComputationNetwork.cpp
Просмотреть файл

@ -498,7 +498,7 @@ template <class ElemType>
fprintf(stderr, "Setting dropout rate to %.8g.\n", dropoutRate);
// TODO: Change this to use an interface that is independent of <ElemType>.
if (dropoutNodes.size() == 0 && dropoutRate > 0)
fprintf(stderr, "WARNING: there is no dropout node.\n");
fprintf(stderr, "WARNING: Attempting to set dropout rate, but there is no dropout node in the network.\n");
}
// Each dropout node gets a distinct seed. The actual seed for each dropout node is computed as follows:

2
Source/ComputationNetworkLib/ReshapingNodes.h
Просмотреть файл

@ -656,6 +656,7 @@ public:
WhereNode(DEVICEID_TYPE deviceId, const wstring& name) :
Base(deviceId, name)
{
MarkValueNonSharable();
}
virtual void /*ComputationNodeNonLooping::*/ ForwardPropNonLooping() override;
@ -696,6 +697,7 @@ public:
PackedIndexNode(DEVICEID_TYPE deviceId, const wstring& name) :
Base(deviceId, name)
{
MarkValueNonSharable();
}
virtual void /*ComputationNodeNonLooping::*/ ForwardPropNonLooping() override;

6
Source/Math/CPUMatrix.cpp
Просмотреть файл

@ -1382,7 +1382,6 @@ void CPUMatrix<ElemType>::RequireSize(const size_t numRows, const size_t numCols
// Resize() -- change matrix size
// This function is cheap if the matrix size does not change.
// Current content is not preserved.
// BUGBUG: There is code that relies on zero initialization (without, we get subtle variations of output). That is wrong--we should initialize to QNaN and see where it fails.
// If growOnly is true, resize will not reallocate memory if the current memory is large enough (i.e., will not shrink).
// If this object does not own its memory then new memory cannot be allocated (one can still shrink and/or reshape).
template <class ElemType>
@ -1411,8 +1410,9 @@ void CPUMatrix<ElemType>::Resize(const size_t numRows, const size_t numCols, boo
}
// success
m_numRows = numRows;
m_numCols = numCols;
m_sliceViewOffset = 0;
m_numRows = numRows;
m_numCols = numCols;
}
// allocated by the callee but should be deleted by the caller

1
Source/Math/CPUSparseMatrix.cpp
Просмотреть файл

@ -1384,6 +1384,7 @@ template void CPUSparseMatrix<char>::SetValue(CPUSparseMatrix<char> const&);
template char* CPUSparseMatrix<char>::Data() const;
template char* CPUSparseMatrix<char>::Data();
template void CPUSparseMatrix<char>::Reset(void);
template void CPUSparseMatrix<char>::Resize(const size_t, const size_t, const size_t, const bool);
template void CPUSparseMatrix<char>::RequireSizeAndAllocate(const size_t, const size_t, const size_t, const bool, bool);
template void CPUSparseMatrix<char>::RequireSizeAndAllocate(const size_t, const size_t, const size_t, const MatrixFormat, const bool, bool);
template CPUSparseMatrix<char>::~CPUSparseMatrix();

40
Source/Math/GPUMatrix.cu
Просмотреть файл

@ -1467,29 +1467,27 @@ void GPUMatrix<ElemType>::Resize(const size_t numRows, const size_t numCols, boo
if (GetNumRows() == numRows && GetNumCols() == numCols)
return;
size_t numElements = numRows * numCols;
if (numElements > GetSizeAllocated() || // grow allocation
(!growOnly && numElements != GetSizeAllocated())) // shrink allocation if not growOnly
{
// reallocate buffer if numElements > 0
ElemType* pArray = nullptr;
if (numElements > 0)
pArray = TracingGPUMemoryAllocator::Allocate<ElemType>(GetComputeDeviceId(), numRows, numCols);
// If the buffer exists, free it
if (Buffer())
TracingGPUMemoryAllocator::Free<ElemType>(GetComputeDeviceId(), Buffer());
SetBuffer(pArray, numElements * sizeof(ElemType));
SetSizeAllocated(numElements);
}
// success
m_sliceViewOffset = 0;
m_numRows = numRows;
m_numCols = numCols;
size_t numElements = GetNumElements();
if (numElements > GetSizeAllocated() || (!growOnly && numElements != GetSizeAllocated()))
{
if (IsEmpty())
{
SetSizeAllocated(0);
SetBuffer(nullptr, 0);
}
else
{
if (Buffer())
{
TracingGPUMemoryAllocator::Free<ElemType>(GetComputeDeviceId(), Buffer());
}
SetSizeAllocated(numElements);
SetBuffer(TracingGPUMemoryAllocator::Allocate<ElemType>(GetComputeDeviceId(), m_numRows, m_numCols), numElements * sizeof(ElemType));
CUDA_CALL(cudaMemset(Buffer(), 0, sizeof(ElemType) * GetSizeAllocated()));
}
}
m_sliceViewOffset = 0;
}
template <class ElemType>

81
Source/Math/Matrix.cpp
Просмотреть файл

@ -1223,32 +1223,32 @@ void Matrix<ElemType>::AssignValuesOf(const Matrix<ElemType>& deepCopyFrom)
DISPATCH_MATRIX_ON_FLAG(this, this,
{
// Set CPUMatrix from:
DISPATCH_MATRIX_ON_FLAG(&deepCopyFrom, &deepCopyFrom,
DISPATCH_MATRIX_ON_FLAG(&deepCopyFrom, nullptr,
{ m_CPUMatrix->SetValue(*deepCopyFrom.m_CPUMatrix); },
NOT_IMPLEMENTED,//{ m_CPUMatrix->SetValue(*deepCopyFrom.m_GPUMatrix); },
NOT_IMPLEMENTED,//{ m_CPUMatrix->SetValue(*deepCopyFrom.m_CPUSparseMatrix); },
NOT_IMPLEMENTED);//{ m_CPUMatrix->SetValue(*deepCopyFrom.m_GPUSparseMatrix); });
{ LogicError("AssignValuesOf: Assigning a GPUMatrix to a CPUMatrix is not yet implemented."); },//{ m_CPUMatrix->SetValue(deepCopyFrom.m_GPUMatrix->GetNumRows(), deepCopyFrom.m_GPUMatrix->GetNumCols(), deepCopyFrom.m_GPUMatrix->CopyToArray()); },// //{ m_CPUMatrix->SetValue(*deepCopyFrom.m_GPUMatrix); },
{ LogicError("AssignValuesOf: Assigning a CPUSparseMatrix to a CPUMatrix is not yet implemented."); },//{ m_CPUMatrix->SetValue(*deepCopyFrom.m_CPUSparseMatrix); },
{ LogicError("AssignValuesOf: Assigning a GPUSparseMatrix to a CPUMatrix is not yet implemented."); });//{ m_CPUMatrix->SetValue(*deepCopyFrom.m_GPUSparseMatrix); });
},
{
// Set GPUMatrix from:
DISPATCH_MATRIX_ON_FLAG(&deepCopyFrom, &deepCopyFrom,
NOT_IMPLEMENTED,//{ m_GPUMatrix->SetValue(*deepCopyFrom.m_CPUMatrix); },
DISPATCH_MATRIX_ON_FLAG(&deepCopyFrom, nullptr,
{ LogicError("AssignValuesOf: Assigning a GPUMatrix to a CPUMatrix is not yet implemented."); },//{ m_GPUMatrix->SetValue(deepCopyFrom.m_CPUMatrix->GetNumRows(), deepCopyFrom.m_CPUMatrix->GetNumCols(), m_GPUMatrix->GetComputeDeviceId(), deepCopyFrom.m_CPUMatrix->Data()); },////{ m_CPUMatrix->SetValue(*deepCopyFrom.m_GPUMatrix); },
{ m_GPUMatrix->SetValue(*deepCopyFrom.m_GPUMatrix); },
NOT_IMPLEMENTED,//{ m_GPUMatrix->SetValue(*deepCopyFrom.m_CPUSparseMatrix); },
NOT_IMPLEMENTED);//{ m_GPUMatrix->SetValue(*deepCopyFrom.m_GPUSparseMatrix); });
{ LogicError("AssignValuesOf: Assigning a CPUSparseMatrix to a GPUMatrix is not yet implemented."); },//{ m_GPUMatrix->SetValue(*deepCopyFrom.m_CPUSparseMatrix); },
{ LogicError("AssignValuesOf: Assigning a GPUSparseMatrix to a GPUMatrix is not yet implemented."); });//{ m_GPUMatrix->SetValue(*deepCopyFrom.m_GPUSparseMatrix); });
},
{
// Set CPUSparseMatrix from:
DISPATCH_MATRIX_ON_FLAG(&deepCopyFrom, &deepCopyFrom,
NOT_IMPLEMENTED,//{ m_CPUSparseMatrix->SetValue(*deepCopyFrom.m_CPUMatrix); },
NOT_IMPLEMENTED,//{ m_CPUSparseMatrix->SetValue(*deepCopyFrom.m_GPUMatrix); },
DISPATCH_MATRIX_ON_FLAG(&deepCopyFrom, nullptr,
{ LogicError("AssignValuesOf: Assigning a CPUMatrix to a CPUSparseMatrix is not yet implemented."); },//{ m_CPUSparseMatrix->SetValue(*deepCopyFrom.m_CPUMatrix); },
{ LogicError("AssignValuesOf: Assigning a GPUMatrix to a CPUSparseMatrix is not yet implemented."); },//{ m_CPUSparseMatrix->SetValue(*deepCopyFrom.m_GPUMatrix); },
{ m_CPUSparseMatrix->SetValue(*deepCopyFrom.m_CPUSparseMatrix); },
NOT_IMPLEMENTED);//{ m_CPUSparseMatrix->SetValue(*deepCopyFrom.m_GPUSparseMatrix); });
{ LogicError("AssignValuesOf: Assigning a GPUSparseMatrix to a CPUSparseMatrix is not yet implemented."); });//{ m_CPUSparseMatrix->SetValue(*deepCopyFrom.m_GPUSparseMatrix); });
},
{
// Set GPUSparseMatrix from:
DISPATCH_MATRIX_ON_FLAG(&deepCopyFrom, &deepCopyFrom,
NOT_IMPLEMENTED,//{ m_GPUSparseMatrix->SetValue(*deepCopyFrom.m_CPUMatrix); },
DISPATCH_MATRIX_ON_FLAG(&deepCopyFrom, nullptr,
{ LogicError("AssignValuesOf: Assigning a CPUMatrix to a GPUSparseMatrix is not yet implemented."); },//{ m_GPUSparseMatrix->SetValue(*deepCopyFrom.m_CPUMatrix); },
{ m_GPUSparseMatrix->SetValue(*deepCopyFrom.m_GPUMatrix); },
{ m_GPUSparseMatrix->SetValue(*deepCopyFrom.m_CPUSparseMatrix); },
{ m_GPUSparseMatrix->SetValue(*deepCopyFrom.m_GPUSparseMatrix); });
@ -3610,13 +3610,17 @@ void Matrix<ElemType>::_transferFromDeviceToDevice(int from_id, int to_id, bool
if (!m_CPUSparseMatrix)
LogicError("Can't move from CPU because I'm not there!");
if (!m_GPUSparseMatrix)
m_GPUSparseMatrix = make_shared<GPUSparseMatrix<ElemType>>(to_id, m_CPUSparseMatrix->GetFormat());
else
m_GPUSparseMatrix->ChangeDeviceTo(to_id);
if (m_CPUSparseMatrix->GetNumElements() != 0 && !emptyTransfer)
if (emptyTransfer)
{
if (m_GPUSparseMatrix && m_GPUSparseMatrix->GetComputeDeviceId() == to_id)
m_GPUSparseMatrix->Resize(m_CPUSparseMatrix->GetNumRows(), m_CPUSparseMatrix->GetNumCols(), m_CPUSparseMatrix->NzCount());
else
m_GPUSparseMatrix = make_shared<GPUSparseMatrix<ElemType>>(m_CPUSparseMatrix->GetNumRows(), m_CPUSparseMatrix->GetNumCols(), m_CPUSparseMatrix->NzCount(), to_id, m_CPUSparseMatrix->GetFormat());
}
else
{
if (!m_GPUSparseMatrix || m_GPUSparseMatrix->GetComputeDeviceId() != to_id)
m_GPUSparseMatrix = make_shared<GPUSparseMatrix<ElemType>>(to_id);
m_GPUSparseMatrix->SetValue(*m_CPUSparseMatrix);
}
@ -3640,10 +3644,10 @@ void Matrix<ElemType>::_transferFromDeviceToDevice(int from_id, int to_id, bool
if (!m_CPUSparseMatrix)
m_CPUSparseMatrix = make_shared<CPUSparseMatrix<ElemType>>(m_GPUSparseMatrix->GetFormat());
if (m_GPUSparseMatrix->GetNumElements() != 0 && !emptyTransfer)
{
if (emptyTransfer)
m_CPUSparseMatrix->Resize(m_GPUSparseMatrix->GetNumRows(), m_GPUSparseMatrix->GetNumCols(), m_GPUSparseMatrix->NzCount(), true);
else
m_GPUSparseMatrix->CopyToCPUSparseMatrix(*m_CPUSparseMatrix);
}
if (isBeingMoved)
{
@ -3668,13 +3672,19 @@ void Matrix<ElemType>::_transferFromDeviceToDevice(int from_id, int to_id, bool
{
if (!m_CPUMatrix)
LogicError("Can't move from CPU because I'm not there!");
if (m_CPUMatrix->GetNumElements() != 0 && !emptyTransfer)
if (emptyTransfer)
{
m_GPUMatrix = make_shared<GPUMatrix<ElemType>>(m_CPUMatrix->GetNumRows(), m_CPUMatrix->GetNumCols(), to_id, m_CPUMatrix->Buffer(), matrixFlagNormal);
if (m_GPUMatrix && m_GPUMatrix->GetComputeDeviceId() == to_id)
m_GPUMatrix->Resize(m_CPUMatrix->GetNumRows(), m_CPUMatrix->GetNumCols());
else
m_GPUMatrix = make_shared<GPUMatrix<ElemType>>(m_CPUMatrix->GetNumRows(), m_CPUMatrix->GetNumCols(), to_id);
}
else
{
m_GPUMatrix = make_shared<GPUMatrix<ElemType>>(to_id);
if (m_GPUMatrix && m_GPUMatrix->GetComputeDeviceId() == to_id)
m_GPUMatrix->SetValue(m_CPUMatrix->GetNumRows(), m_CPUMatrix->GetNumCols(), to_id, m_CPUMatrix->Data());
else
m_GPUMatrix = make_shared<GPUMatrix<ElemType>>(m_CPUMatrix->GetNumRows(), m_CPUMatrix->GetNumCols(), to_id, m_CPUMatrix->Data());
}
if (isBeingMoved)
{
@ -3682,9 +3692,7 @@ void Matrix<ElemType>::_transferFromDeviceToDevice(int from_id, int to_id, bool
m_CPUMatrix = nullptr;
}
else
{
SetDataLocation(BOTH, DENSE);
}
}
else // from GPU
{
@ -3693,15 +3701,22 @@ void Matrix<ElemType>::_transferFromDeviceToDevice(int from_id, int to_id, bool
if (to_id < 0) // to CPU
{
if (m_GPUMatrix->GetNumElements() != 0 && !emptyTransfer)
if (emptyTransfer)
{
ElemType* arr = m_GPUMatrix->CopyToArray(); // TODO: unnecessary allocation/copy; why not make this a vector that we move over as an rvalue ref?
m_CPUMatrix = make_shared<CPUMatrix<ElemType>>(m_GPUMatrix->GetNumRows(), m_GPUMatrix->GetNumCols(), arr, matrixFlagNormal);
delete[] arr;
if (m_CPUMatrix)
m_CPUMatrix->Resize(m_GPUMatrix->GetNumRows(), m_GPUMatrix->GetNumCols());
else
m_CPUMatrix = make_shared<CPUMatrix<ElemType>>(m_GPUMatrix->GetNumRows(), m_GPUMatrix->GetNumCols());
}
else
{
m_CPUMatrix = make_shared<CPUMatrix<ElemType>>();
ElemType* arr = m_GPUMatrix->CopyToArray(); // TODO: unnecessary allocation/copy; why not make this a vector that we move over as an rvalue ref?
if (m_CPUMatrix)
m_CPUMatrix->SetValue(m_GPUMatrix->GetNumRows(), m_GPUMatrix->GetNumCols(), arr);
else
m_CPUMatrix = make_shared<CPUMatrix<ElemType>>(m_GPUMatrix->GetNumRows(), m_GPUMatrix->GetNumCols(), arr, matrixFlagNormal);
delete[] arr;
}
if (isBeingMoved)

34
Source/Readers/CNTKTextFormatReader/CNTKTextFormatReader.cpp
Просмотреть файл

@ -7,13 +7,18 @@
#include "CNTKTextFormatReader.h"
#include "Config.h"
#include "TextConfigHelper.h"
#include "ChunkCache.h"
#include "BlockRandomizer.h"
#include "NoRandomizer.h"
#include "TextParser.h"
#include "SequencePacker.h"
#include "FramePacker.h"
namespace Microsoft { namespace MSR { namespace CNTK {
// TODO: This class should go away eventually.
// TODO: The composition of packer + randomizer + different deserializers in a generic manner is done in the CompositeDataReader.
// TODO: Currently preserving this for backward compatibility with current configs.
CNTKTextFormatReader::CNTKTextFormatReader(MemoryProviderPtr provider,
const ConfigParameters& config) :
m_provider(provider)
@ -31,28 +36,37 @@ CNTKTextFormatReader::CNTKTextFormatReader(MemoryProviderPtr provider,
m_deserializer = shared_ptr<IDataDeserializer>(new TextParser<double>(configHelper));
}
TransformerPtr randomizer;
if (configHelper.ShouldKeepDataInMemory())
{
m_deserializer = shared_ptr<IDataDeserializer>(new ChunkCache(m_deserializer));
}
size_t window = configHelper.GetRandomizationWindow();
if (window > 0)
{
// Verbosity is a general config parameter, not specific to the text format reader.
int verbosity = config(L"verbosity", 2);
randomizer = make_shared<BlockRandomizer>(verbosity, window, m_deserializer);
m_randomizer = make_shared<BlockRandomizer>(verbosity, window, m_deserializer);
}
else
{
randomizer = std::make_shared<NoRandomizer>(m_deserializer);
m_randomizer = std::make_shared<NoRandomizer>(m_deserializer);
}
randomizer->Initialize(nullptr, config);
m_transformer = randomizer;
// TODO: add "frameMode" config paramter
if (configHelper.IsInFrameMode())
{
m_packer = std::make_shared<FramePacker>(
m_provider,
m_randomizer,
GetStreamDescriptions());
}
else
{
m_packer = std::make_shared<SequencePacker>(
m_provider,
m_transformer,
m_randomizer,
GetStreamDescriptions());
}
}
catch (const std::runtime_error& e)
{
@ -72,7 +86,7 @@ void CNTKTextFormatReader::StartEpoch(const EpochConfiguration& config)
RuntimeError("Epoch size cannot be 0.");
}
m_transformer->StartEpoch(config);
m_randomizer->StartEpoch(config);
m_packer->StartEpoch(config);
}

5
Source/Readers/CNTKTextFormatReader/CNTKTextFormatReader.h
Просмотреть файл

@ -8,6 +8,7 @@
#include "TextParser.h"
#include "Reader.h"
#include "Packer.h"
#include "SequenceEnumerator.h"
namespace Microsoft { namespace MSR { namespace CNTK {
@ -32,8 +33,8 @@ public:
private:
IDataDeserializerPtr m_deserializer;
// A head transformer in a list of transformers.
TransformerPtr m_transformer;
// Randomizer.
SequenceEnumeratorPtr m_randomizer;
// Packer.
PackerPtr m_packer;

28
Source/Readers/CNTKTextFormatReader/Exports.cpp
Просмотреть файл

@ -11,7 +11,7 @@
#include "ReaderShim.h"
#include "CNTKTextFormatReader.h"
#include "HeapMemoryProvider.h"
#include "CudaMemoryProvider.h"
#include "StringUtil.h"
namespace Microsoft { namespace MSR { namespace CNTK {
@ -32,4 +32,30 @@ extern "C" DATAREADER_API void GetReaderD(IDataReader** preader)
*preader = new ReaderShim<double>(factory);
}
// TODO: Not safe from the ABI perspective. Will be uglified to make the interface ABI.
// A factory method for creating text deserializers.
extern "C" DATAREADER_API bool CreateDeserializer(IDataDeserializer** deserializer, const std::wstring& type, const ConfigParameters& deserializerConfig, CorpusDescriptorPtr corpus, bool)
{
string precision = deserializerConfig.Find("precision", "float");
if (!AreEqualIgnoreCase(precision, "float") && !AreEqualIgnoreCase(precision, "double"))
{
InvalidArgument("Unsupported precision '%s'", precision.c_str());
}
// TODO: Remove type from the parser. Current implementation does not support streams of different types.
if (type == L"CNTKTextFormatDeserializer")
{
if (precision == "float")
*deserializer = new TextParser<float>(corpus, TextConfigHelper(deserializerConfig));
else // double
*deserializer = new TextParser<double>(corpus, TextConfigHelper(deserializerConfig));
}
else
InvalidArgument("Unknown deserializer type '%ls'", type.c_str());
// Deserializer created.
return true;
}
}}}

18
Source/Readers/CNTKTextFormatReader/Indexer.cpp
Просмотреть файл

@ -113,7 +113,7 @@ void Indexer::BuildFromLines()
}
void Indexer::Build()
void Indexer::Build(CorpusDescriptorPtr corpus)
{
if (!m_chunks.empty())
{
@ -174,7 +174,8 @@ void Indexer::Build()
{
// found a new sequence, which starts at the [offset] bytes into the file
sd.m_byteSize = offset - sd.m_fileOffsetBytes;
AddSequence(sd);
AddSequenceIfIncluded(corpus, sd);
sd = {};
sd.m_id = id;
sd.m_fileOffsetBytes = offset;
@ -184,9 +185,20 @@ void Indexer::Build()
// calculate the byte size for the last sequence
sd.m_byteSize = m_fileOffsetEnd - sd.m_fileOffsetBytes;
AddSequence(sd);
AddSequenceIfIncluded(corpus, sd);
}
void Indexer::AddSequenceIfIncluded(CorpusDescriptorPtr corpus, SequenceDescriptor& sd)
{
auto& stringRegistry = corpus->GetStringRegistry();
auto key = msra::strfun::utf16(std::to_string(sd.m_id));
if (corpus->IsIncluded(key))
{
sd.m_key.m_sequence = stringRegistry[key];
sd.m_key.m_sample = 0;
AddSequence(sd);
}
}
void Indexer::SkipLine()
{

6
Source/Readers/CNTKTextFormatReader/Indexer.h
Просмотреть файл

@ -8,6 +8,7 @@
#include <stdint.h>
#include <vector>
#include "Descriptors.h"
#include "CorpusDescriptor.h"
namespace Microsoft { namespace MSR { namespace CNTK {
@ -23,7 +24,7 @@ public:
// Reads the input file, building and index of chunks and corresponding
// sequences.
void Build();
void Build(CorpusDescriptorPtr corpus);
// Returns input data index (chunk and sequence metadata)
const Index& GetIndex() const { return m_chunks; }
@ -59,6 +60,9 @@ private:
// (except when a sequence size is greater than the maximum chunk size)
void AddSequence(SequenceDescriptor& sd);
// Same function as above but with check that the sequence is included in the corpus descriptor.
void AddSequenceIfIncluded(CorpusDescriptorPtr corpus, SequenceDescriptor& sd);
// fills up the buffer with data from file, all previously buffered data
// will be overwritten.
void RefillBuffer();

39
Source/Readers/CNTKTextFormatReader/TextConfigHelper.cpp
Просмотреть файл

@ -4,6 +4,7 @@
//
#include "stdafx.h"
#define __STDC_FORMAT_MACROS
#include <inttypes.h>
#include <limits>
#include "TextConfigHelper.h"
@ -116,32 +117,36 @@ TextConfigHelper::TextConfigHelper(const ConfigParameters& config)
m_filepath = msra::strfun::utf16(config(L"file"));
if (config.Exists(L"randomize"))
// EvalActions inserts randomize = "none" into the reader config in DoWriteOutoput.
wstring randomizeString = config(L"randomize", wstring());
if (!_wcsicmp(randomizeString.c_str(), L"none"))
{
wstring randomizeString = config.CanBeString(L"randomize") ? config(L"randomize") : wstring();
if (!_wcsicmp(randomizeString.c_str(), L"none"))
{
m_randomizationWindow = randomizeNone;
}
else if (!_wcsicmp(randomizeString.c_str(), L"auto"))
{
m_randomizationWindow = randomizeAuto;
}
else
{
m_randomizationWindow = config(L"randomize");
}
m_randomizationWindow = randomizeNone;
}
else
{
m_randomizationWindow = randomizeAuto;
bool randomize = config(L"randomize", true);
if (!randomize)
{
m_randomizationWindow = randomizeNone;
}
else if (config.Exists(L"randomizationWindow"))
{
m_randomizationWindow = config(L"randomizationWindow");
}
else
{
m_randomizationWindow = randomizeAuto;
}
}
m_skipSequenceIds = config(L"skipSequenceIds", false);
m_maxErrors = config(L"maxErrors", 0);
m_traceLevel = config(L"traceLevel", 0);
m_traceLevel = config(L"traceLevel", 1);
m_chunkSizeBytes = config(L"chunkSizeInBytes", 32 * 1024 * 1024); // 32 MB by default
m_chunkCacheSize = config(L"numChunksToCache", 32); // 32 * 32 MB = 1 GB of memory in total
m_keepDataInMemory = config(L"keepDataInMemory", false);
m_frameMode = config(L"frameMode", false);
}
}}}

7
Source/Readers/CNTKTextFormatReader/TextConfigHelper.h
Просмотреть файл

@ -35,7 +35,9 @@ public:
size_t GetChunkSize() const { return m_chunkSizeBytes; }
unsigned int GetNumChunksToCache() const { return m_chunkCacheSize; }
bool ShouldKeepDataInMemory() const { return m_keepDataInMemory; }
bool IsInFrameMode() const { return m_frameMode; }
ElementType GetElementType() const { return m_elementType; }
@ -50,7 +52,8 @@ private:
unsigned int m_maxErrors;
unsigned int m_traceLevel;
size_t m_chunkSizeBytes; // chunks size in bytes
unsigned int m_chunkCacheSize; // number of chunks to keep in the memory
bool m_keepDataInMemory; // if true the whole dataset is kept in memory
bool m_frameMode; // if true, the maximum expected sequence length in the dataset is one sample.
};
} } }

216
Source/Readers/CNTKTextFormatReader/TextParser.cpp
Просмотреть файл

@ -16,12 +16,6 @@
namespace Microsoft { namespace MSR { namespace CNTK {
inline bool isDelimiter(char c)
{
return c == VALUE_DELIMITER || c == NAME_PREFIX || c == COLUMN_DELIMITER ||
c == INDEX_DELIMITER || c == ROW_DELIMITER || c == CARRIAGE_RETURN;
}
enum State
{
Init = 0,
@ -48,10 +42,6 @@ public:
// chunk id (copied from the descriptor)
size_t m_id;
// Keeps track of how many times GetSequence was called.
// When this counter value reaches the number of sequences in
// the this chunk, it can be safely unloaded.
size_t m_sequenceRequestCount;
// a non-owned pointer to the parser that created this chunk
TextParser* m_parser;
@ -66,18 +56,22 @@ struct TextParser<ElemType>::StreamInfo
};
template <class ElemType>
TextParser<ElemType>::TextParser(const TextConfigHelper& helper) :
TextParser<ElemType>::TextParser(const TextConfigHelper& helper) : TextParser(std::make_shared<CorpusDescriptor>(), helper)
{}
template <class ElemType>
TextParser<ElemType>::TextParser(CorpusDescriptorPtr corpus, const TextConfigHelper& helper) :
TextParser(helper.GetFilePath(), helper.GetStreams())
{
SetTraceLevel(helper.GetTraceLevel());
SetMaxAllowedErrors(helper.GetMaxAllowedErrors());
SetChunkCacheSize(helper.GetNumChunksToCache());
SetChunkSize(helper.GetChunkSize());
SetSkipSequenceIds(helper.ShouldSkipSequenceIds());
Initialize();
Initialize(corpus);
}
template <class ElemType>
TextParser<ElemType>::TextParser(const std::wstring& filename, const vector<StreamDescriptor>& streams) :
m_filename(filename),
@ -91,7 +85,6 @@ TextParser<ElemType>::TextParser(const std::wstring& filename, const vector<Stre
m_bufferEnd(nullptr),
m_pos(nullptr),
m_chunkSizeBytes(0),
m_chunkCacheSize(0),
m_traceLevel(TraceLevel::Error),
m_hadWarnings(false),
m_numAllowedErrors(0),
@ -145,29 +138,36 @@ void TextParser<ElemType>::PrintWarningNotification()
}
template <class ElemType>
void TextParser<ElemType>::Initialize()
void TextParser<ElemType>::Initialize(CorpusDescriptorPtr corpus)
{
if (m_indexer != nullptr)
{
return;
}
attempt(m_numRetries, [this]()
attempt(m_numRetries, [this, corpus]()
{
m_file = fopenOrDie(m_filename, L"rbS");
});
if (m_file == nullptr)
{
m_file = fopenOrDie(m_filename, L"rbS");
}
else if (ferror(m_file) != 0)
{
fclose(m_file);
m_file = fopenOrDie(m_filename, L"rbS");
}
if (funicode(m_file))
{
// Retrying won't help here, the file is UTF-16 encoded.
m_numRetries = 0;
RuntimeError("Found a UTF-16 BOM at the beginning of the input file (%ls). "
"UTF-16 encoding is currently not supported.", m_filename.c_str());
}
if (funicode(m_file))
{
RuntimeError("Found a UTF-16 BOM at the beginning of the input file (%ls). "
"UTF-16 encoding is currently not supported.", m_filename.c_str());
}
m_indexer = make_unique<Indexer>(m_file, m_skipSequenceIds, m_chunkSizeBytes);
m_indexer = make_unique<Indexer>(m_file, m_skipSequenceIds, m_chunkSizeBytes);
attempt(m_numRetries, [this]()
{
m_indexer->Build();
m_indexer->Build(corpus);
});
// it's still possible that the actual input data does not have sequence id column.
@ -232,7 +232,6 @@ TextParser<ElemType>::TextDataChunk::TextDataChunk(const ChunkDescriptor& descri
m_parser(parser)
{
m_id = descriptor.m_id;
m_sequenceRequestCount = 0;
}
template <class ElemType>
@ -240,7 +239,6 @@ void TextParser<ElemType>::TextDataChunk::GetSequence(size_t sequenceId, std::ve
{
auto it = m_sequenceMap.find(sequenceId);
assert(it != m_sequenceMap.end());
++m_sequenceRequestCount;
result.reserve(m_parser->m_streamInfos.size());
const auto& sequenceData = it->second;
for (size_t j = 0; j < m_parser->m_streamInfos.size(); ++j)
@ -278,50 +276,20 @@ void TextParser<ElemType>::TextDataChunk::GetSequence(size_t sequenceId, std::ve
template <class ElemType>
ChunkPtr TextParser<ElemType>::GetChunk(size_t chunkId)
{
ChunkPtr chunk;
auto it = m_chunkCache.find(chunkId);
if (it != m_chunkCache.end())
const auto& chunkDescriptor = m_indexer->GetIndex()[chunkId];
auto textChunk = make_shared<TextDataChunk>(chunkDescriptor, this);
attempt(m_numRetries, [this, &textChunk, &chunkDescriptor]()
{
chunk = it->second;
}
else
{
const auto& chunkDescriptor = m_indexer->GetIndex()[chunkId];
auto textChunk = make_shared<TextDataChunk>(chunkDescriptor, this);
attempt(m_numRetries, [this, &textChunk, &chunkDescriptor]()
if (ferror(m_file) != 0)
{
LoadChunk(textChunk, chunkDescriptor);
});
if (m_chunkCacheSize > 0 && m_chunkCache.size() == m_chunkCacheSize)
{
size_t candidateId = SIZE_MAX;
size_t minNumSequencesLeft = SIZE_MAX;
for (const auto& it : m_chunkCache)
{
const auto& chunk = *(it.second.get());
size_t numSequencesUsed = 0;
numSequencesUsed += chunk.m_sequenceRequestCount;
size_t numSequencesLeft = chunk.m_sequenceMap.size() - numSequencesUsed;
if (numSequencesLeft < minNumSequencesLeft)
{
minNumSequencesLeft = numSequencesLeft;
candidateId = it.first;
}
}
assert(candidateId != SIZE_MAX);
m_chunkCache.erase(candidateId);
fclose(m_file);
m_file = fopenOrDie(m_filename, L"rbS");
}
LoadChunk(textChunk, chunkDescriptor);
});
if (m_chunkCacheSize > 0)
{
m_chunkCache[chunkId] = textChunk;
}
chunk = textChunk;
}
return chunk;
return textChunk;
}
template <class ElemType>
@ -456,7 +424,7 @@ typename TextParser<ElemType>::SequenceBuffer TextParser<ElemType>::LoadSequence
fprintf(stderr,
"WARNING: Exhausted all input"
" expected for the current sequence (id = %" PRIu64 ") %ls,"
" but only read %" PRId64 " out of %" PRId64 " expected rows.\n",
" but only read %" PRIu64 " out of %" PRIu64 " expected rows.\n",
sequenceDsc.m_id, GetFileInfo().c_str(), numRowsRead, expectedRowCount);
}
break;
@ -466,7 +434,7 @@ typename TextParser<ElemType>::SequenceBuffer TextParser<ElemType>::LoadSequence
// Double check if there are empty input streams.
// TODO this handling needs to be graceful, but currently CNTK complains when we return empty sequences.
bool hasEmptyInputs = false, hasDuplicateInputs = false;
size_t maxInputLength = 0;
for (size_t i = 0; i < sequence.size(); ++i)
{
if (sequence[i]->m_numberOfSamples == 0)
@ -484,11 +452,12 @@ typename TextParser<ElemType>::SequenceBuffer TextParser<ElemType>::LoadSequence
{
fprintf(stderr,
"WARNING: Input ('%ls') contains more samples than expected"
" (%" PRId64 " vs. %" PRId64 ") for sequence (id = %" PRIu64 ") %ls.\n",
" (%" PRIu64 " vs. %" PRIu64 ") for sequence (id = %" PRIu64 ") %ls.\n",
m_streams[i]->m_name.c_str(), sequence[i]->m_numberOfSamples, expectedRowCount,
sequenceDsc.m_id, GetFileInfo().c_str());
}
}
maxInputLength = max(sequence[i]->m_numberOfSamples, maxInputLength);
}
if (hasEmptyInputs)
@ -501,6 +470,17 @@ typename TextParser<ElemType>::SequenceBuffer TextParser<ElemType>::LoadSequence
{
IncrementNumberOfErrorsOrDie();
}
else if (maxInputLength < expectedRowCount)
{
if (ShouldWarn())
{
fprintf(stderr,
"WARNING: Maximum per-input number of samples for sequence (id = %" PRIu64 ") %ls"
" is less than expected (%" PRIu64 " vs. %" PRIu64 ").\n",
sequenceDsc.m_id, GetFileInfo().c_str(), maxInputLength, expectedRowCount);
}
IncrementNumberOfErrorsOrDie();
}
if (m_traceLevel >= Info)
{
@ -528,14 +508,6 @@ bool TextParser<ElemType>::TryReadRow(SequenceBuffer& sequence, size_t& bytesToR
{
char c = *m_pos;
if (c == COLUMN_DELIMITER || c == VALUE_DELIMITER || c == CARRIAGE_RETURN)
{
// skip column and value separators, as well as carriage returns.
++m_pos;
--bytesToRead;
continue;
}
if (c == ROW_DELIMITER)
{
// found the end of row, skip the delimiter, return.
@ -551,13 +523,21 @@ bool TextParser<ElemType>::TryReadRow(SequenceBuffer& sequence, size_t& bytesToR
{
fprintf(stderr,
"WARNING: Input row %ls contains more"
" samples than expected (%" PRId64 " vs. %" PRId64 ").\n",
" samples than expected (%" PRIu64 " vs. %" PRIu64 ").\n",
GetFileInfo().c_str(), numSampleRead, m_streams.size());
}
return numSampleRead > 0;
}
if (isColumnDelimiter(c))
{
// skip column (input) delimiters.
++m_pos;
--bytesToRead;
continue;
}
if (TryReadSample(sequence, bytesToRead))
{
numSampleRead++;
@ -685,9 +665,9 @@ bool TextParser<ElemType>::TryGetInputId(size_t& id, size_t& bytesToRead)
{
char c = *m_pos;
// an input id can be followed by a value marker, end of line (also, carriage return),
// column separator or the name prefix of the following input.
if (c <= VALUE_DELIMITER || c == NAME_PREFIX)
// stop as soon as there's a value delimiter, an input prefix
// or a non-printable character (e.g., newline, carriage return).
if (isValueDelimiter(c) || c == NAME_PREFIX || isNonPrintable(c))
{
size_t size = scratchIndex - m_scratch.get();
if (size)
@ -758,16 +738,24 @@ bool TextParser<ElemType>::TryReadDenseSample(vector<ElemType>& values, size_t s
{
char c = *m_pos;
if (isValueDelimiter(c))
{
// skip value delimiters
++m_pos;
--bytesToRead;
continue;
}
// return as soon as we hit a non-printable or a name prefix
if (c < VALUE_DELIMITER || c == NAME_PREFIX)
if (isNonPrintable(c) || c == NAME_PREFIX)
{
if (counter > sampleSize)
{
if (ShouldWarn())
{
fprintf(stderr,
"WARNING: Dense sample (size = %" PRId64 ") %ls"
" exceeds the expected size (%" PRId64 ").\n",
"WARNING: Dense sample (size = %" PRIu64 ") %ls"
" exceeds the expected size (%" PRIu64 ").\n",
counter, GetFileInfo().c_str(), sampleSize);
}
return false;
@ -781,7 +769,7 @@ bool TextParser<ElemType>::TryReadDenseSample(vector<ElemType>& values, size_t s
{
fprintf(stderr,
"WARNING: A dense sample %ls has a sparse suffix "
"(expected size = %" PRId64 ", actual size = %" PRId64 ").\n",
"(expected size = %" PRIu64 ", actual size = %" PRIu64 ").\n",
GetFileInfo().c_str(), sampleSize, counter);
}
for (; counter < sampleSize; ++counter)
@ -793,14 +781,6 @@ bool TextParser<ElemType>::TryReadDenseSample(vector<ElemType>& values, size_t s
return true;
}
if (c == VALUE_DELIMITER)
{
// skip value delimiters
++m_pos;
--bytesToRead;
continue;
}
if (!TryReadRealNumber(value, bytesToRead))
{
// bail out.
@ -832,14 +812,7 @@ bool TextParser<ElemType>::TryReadSparseSample(std::vector<ElemType>& values, st
{
char c = *m_pos;
// return as soon as we hit a non-printable or a name prefix
if (c < VALUE_DELIMITER || c == NAME_PREFIX)
{
// empty sparse samples are allowed ("|InputeName_1|InputName2...")
return true;
}
if (c == VALUE_DELIMITER)
if (isValueDelimiter(c))
{
// skip value delimiters
++m_pos;
@ -847,6 +820,13 @@ bool TextParser<ElemType>::TryReadSparseSample(std::vector<ElemType>& values, st
continue;
}
// return as soon as we hit a non-printable or a name prefix
if (isNonPrintable(c) || c == NAME_PREFIX)
{
// empty sparse samples are allowed ("|InputeName_1|InputName2...")
return true;
}
// read next sparse index
if (!TryReadUint64(index, bytesToRead))
{
@ -876,7 +856,7 @@ bool TextParser<ElemType>::TryReadSparseSample(std::vector<ElemType>& values, st
fprintf(stderr,
"WARNING: Unexpected character('%c')"
" in place of the index delimiter ('%c')"
" after a sparse value index (%" PRId64 ") %ls.\n",
" after a sparse value index (%" PRIu64 ") %ls.\n",
c, INDEX_DELIMITER, index, GetFileInfo().c_str());
}
return false;
@ -913,8 +893,8 @@ void TextParser<ElemType>::SkipToNextValue(size_t& bytesToRead)
while (bytesToRead && CanRead())
{
char c = *m_pos;
// skip everything until we hit either a value marker, an input marker or the end of row.
if (c == VALUE_DELIMITER || c == ROW_DELIMITER || c == NAME_PREFIX)
// skip everything until we hit either a value delimiter, an input marker or the end of row.
if (isValueDelimiter(c) || c == NAME_PREFIX || c == ROW_DELIMITER)
{
return;
}
@ -950,19 +930,7 @@ bool TextParser<ElemType>::TryReadUint64(size_t& value, size_t& bytesToRead)
if (!isdigit(c))
{
if (isDelimiter(c))
{
return found;
}
if (ShouldWarn())
{
fprintf(stderr,
"WARNING: Unexpected character('%c') in a uint64 value %ls.\n",
c, GetFileInfo().c_str());
}
return false;
return found;
}
found |= true;
@ -1213,12 +1181,6 @@ void TextParser<ElemType>::SetSkipSequenceIds(bool skip)
m_skipSequenceIds = skip;
}
template <class ElemType>
void TextParser<ElemType>::SetChunkCacheSize(unsigned int size)
{
m_chunkCacheSize = size;
}
template <class ElemType>
void TextParser<ElemType>::SetChunkSize(size_t size)
{

16
Source/Readers/CNTKTextFormatReader/TextParser.h
Просмотреть файл

@ -9,6 +9,7 @@
#include "Descriptors.h"
#include "TextConfigHelper.h"
#include "Indexer.h"
#include "CorpusDescriptor.h"
namespace Microsoft { namespace MSR { namespace CNTK {
@ -22,10 +23,9 @@ class TextParser : public DataDeserializerBase {
public:
explicit TextParser(const TextConfigHelper& helper);
~TextParser();
TextParser(CorpusDescriptorPtr corpus, const TextConfigHelper& helper);
// Builds an index of the input data.
void Initialize();
~TextParser();
// Retrieves a chunk of data.
ChunkPtr GetChunk(size_t chunkId) override;
@ -37,6 +37,9 @@ public:
void GetSequencesForChunk(size_t chunkId, std::vector<SequenceDescription>& result) override;
private:
// Builds an index of the input data.
void Initialize(CorpusDescriptorPtr corpus);
// A buffer to keep data for all samples in a (variable length) sequence
// from a single input stream.
struct InputStreamBuffer
@ -106,7 +109,6 @@ private:
unique_ptr<char[]> m_scratch; // local buffer for string parsing
size_t m_chunkSizeBytes;
unsigned int m_chunkCacheSize; // number of chunks to keep in the memory
unsigned int m_traceLevel;
bool m_hadWarnings;
unsigned int m_numAllowedErrors;
@ -114,10 +116,6 @@ private:
unsigned int m_numRetries; // specifies the number of times an unsuccessful
// file operation should be repeated (default value is 5).
// A map of currently loaded chunks
// TODO: remove caching once partial randomization is in master.
std::map<size_t, TextChunkPtr> m_chunkCache;
// throws runtime exception when number of parsing errors is
// greater than the specified threshold
void IncrementNumberOfErrorsOrDie();
@ -181,8 +179,6 @@ private:
void SetChunkSize(size_t size);
void SetChunkCacheSize(unsigned int size);
void SetNumRetries(unsigned int numRetries);
friend class CNTKTextFormatReaderTestRunner<ElemType>;

30
Source/Readers/CNTKTextFormatReader/TextReaderConstants.h
Просмотреть файл

@ -7,13 +7,37 @@
namespace Microsoft { namespace MSR { namespace CNTK {
const char COLUMN_DELIMITER = '\t';
const char SPACE_CHAR = ' ';
const char TAB_CHAR = '\t';
const char NAME_PREFIX = '|';
const char VALUE_DELIMITER = ' ';
const char INDEX_DELIMITER = ':';
const char ROW_DELIMITER = '\n';
const char CARRIAGE_RETURN = '\r';
const char ESCAPE_SYMBOL = '#';
const auto BUFFER_SIZE = 256 * 1024;
inline bool isPrintable(char c)
{
return c >= SPACE_CHAR;
}
inline bool isNonPrintable(char c)
{
return !isPrintable(c);
}
inline bool isValueDelimiter(char c)
{
return c == SPACE_CHAR || c == TAB_CHAR;
}
inline bool isColumnDelimiter(char c)
{
return isValueDelimiter(c) || (isNonPrintable(c) && c != ROW_DELIMITER);
}
}}}

132
Source/Readers/CompositeDataReader/CompositeDataReader.cpp
Просмотреть файл

@ -2,7 +2,7 @@
// Copyright (c) Microsoft. All rights reserved.
// Licensed under the MIT license. See LICENSE.md file in the project root for full license information.
//
// CompositeDataReader.cpp : Defines a reader that allows composing different deserializers.
// CompositeReader.cpp : Defines a reader that allows composing different deserializers.
// With this reader in place the users should only extend deserializers.
//
@ -11,8 +11,6 @@
#define _CRT_SECURE_NO_WARNINGS // "secure" CRT not available on all platforms --add this at the top of all CPP files that give "function or variable may be unsafe" warnings
#endif
#define DATAREADER_EXPORTS // creating the exports here
#include "CompositeDataReader.h"
#include "Bundler.h"
#include "BlockRandomizer.h"
@ -20,11 +18,15 @@
#include "FramePacker.h"
#include "SequencePacker.h"
#include "TruncatedBpttPacker.h"
#include "HeapMemoryProvider.h"
#include "CorpusDescriptor.h"
#include "ConfigUtil.h"
namespace Microsoft { namespace MSR { namespace CNTK {
// The whole CompositeDataReader is meant as a stopgap to allow deserializers/transformers composition until SGD talkes
// directly to the new Reader API.
// For more information please see its header file.
// This method composes together packers + randomizer + a set of transformers and deserializers.
CompositeDataReader::CompositeDataReader(const ConfigParameters& config, MemoryProviderPtr provider) : m_layout(make_shared<MBLayout>()),
m_corpus(std::make_shared<CorpusDescriptor>()),
m_provider(provider)
@ -55,37 +57,56 @@ CompositeDataReader::CompositeDataReader(const ConfigParameters& config, MemoryP
m_packingMode = PackingMode::sequence;
}
// Whether we need to check data between different deserializers.
bool cleanse = config(L"checkData", false);
m_precision = config("precision", "float");
// Creating deserializers.
// TODO: Currently the primary deserializer defines the corpus. The logic will be moved to CorpusDescriptor class.
CreateDeserializers(config);
// Bundling deserializers together.
// TODO: Add transformers in between.
auto bundler = std::make_shared<Bundler>(config, m_deserializers[0], m_deserializers, cleanse);
if (m_deserializers.empty())
{
InvalidArgument("Could not find deserializers in the reader config.");
}
IDataDeserializerPtr deserializer = m_deserializers.front();
if (m_deserializers.size() > 1)
{
// Bundling deserializers together.
// Option whether we need to check data between different deserializers.
bool cleanse = config(L"checkData", false);
deserializer = std::make_shared<Bundler>(config, deserializer, m_deserializers, cleanse);
}
int verbosity = config(L"verbosity", 2);
// Pick up the randomizer.
bool randomize = config(L"randomize", false);
// By default do not use omp threads for deserialization of sequences.
// It makes sense to put it to true for cases when deserialization is CPU intensive,
// i.e. decompression of images.
bool multiThreadedDeserialization = config(L"multiThreadedDeserialization", false);
if (randomize)
{
// By default randomizing the whole data set.
size_t randomizationWindow = config(L"randomizationWindow", requestDataSize);
m_randomizer = std::make_shared<BlockRandomizer>(verbosity, randomizationWindow, bundler, BlockRandomizer::DecimationMode::chunk, true);
// By default using STL random number generator.
bool useLegacyRandomization = config(L"useLegacyRandomization", false);
m_sequenceEnumerator = std::make_shared<BlockRandomizer>(verbosity, randomizationWindow, deserializer, BlockRandomizer::DecimationMode::chunk, useLegacyRandomization, multiThreadedDeserialization);
}
else
{
m_randomizer = std::make_shared<NoRandomizer>(bundler);
m_sequenceEnumerator = std::make_shared<NoRandomizer>(deserializer, multiThreadedDeserialization);
}
m_randomizer->Initialize(nullptr, config);
// In case when there are transforms, applying them to the data.
m_sequenceEnumerator = m_transforms.empty()
? m_sequenceEnumerator
: std::make_shared<TransformController>(m_transforms, m_sequenceEnumerator);
// Create output stream descriptions - where to get those? from config? what if it is not the same as network expects?
// TODO: Currently only sparse streams.
for (const auto& streamDescription : bundler->GetStreamDescriptions())
// TODO: Currently only dense output streams.
// TODO: Check here. We should already support repacking sparse into dense in the shim/matrix.
for (const auto& streamDescription : m_sequenceEnumerator->GetStreamDescriptions())
{
StreamDescriptionPtr stream = std::make_shared<StreamDescription>(*streamDescription);
stream->m_storageType = StorageType::dense;
@ -104,13 +125,17 @@ Minibatch CompositeDataReader::ReadMinibatch()
return m_packer->ReadMinibatch();
}
// Create deserializers based on the specified configuration.
// deserializers = [
// [ type = "ImageDataDeserializer" module = "ImageReader" ...]
// [ type = "CNTKTextFormatDeserializer" module = "CNTKTextFormatReader" ...]
void CompositeDataReader::CreateDeserializers(const ConfigParameters& readerConfig)
{
argvector<ConfigValue> deserializerConfigs =
readerConfig(L"deserializers", ConfigParameters::Array(argvector<ConfigValue>(vector<ConfigValue> {})));
assert(m_deserializers.empty());
bool primary = true; // CUrrently, the first deserializer becomes primary - it drives chunking.
bool primary = true; // Currently, the first deserializer becomes primary - it drives chunking.
for (size_t i = 0; i < deserializerConfigs.size(); ++i)
{
// TODO: Should go away in the future. Framing can be done on top of deserializers.
@ -124,6 +149,8 @@ void CompositeDataReader::CreateDeserializers(const ConfigParameters& readerConf
}
}
// Creates a particular deserializer based on the config: its loads the external module and calls CreateDeserializer
// factory function for a particular deserializer type.
IDataDeserializerPtr CompositeDataReader::CreateDeserializer(const ConfigParameters& deserializerConfig, bool primary)
{
typedef bool(*CreateDeserializerFactory) (IDataDeserializer** d, const std::wstring& type, const ConfigParameters& cfg, CorpusDescriptorPtr corpus, bool primary);
@ -138,10 +165,77 @@ IDataDeserializerPtr CompositeDataReader::CreateDeserializer(const ConfigParamet
RuntimeError("Cannot create deserializer. Please check module and type in the configuration.");
}
// Create transformers if necessary.
CreateTransforms(deserializerConfig);
assert(d != nullptr);
return IDataDeserializerPtr(d);
}
// Create transformers based on the configuration, i.e.
// deserializers = [
// [
// type = "ImageDataDeserializer"
// module = "ImageReader"
// inputs = [
// features = [
//----> transforms = [
// [type = "Crop"]:[type = "Scale"]...
void CompositeDataReader::CreateTransforms(const ConfigParameters& deserializerConfig)
{
std::string defaultModule = deserializerConfig("module");
argvector<ConfigParameters> inputs = deserializerConfig("input");
for (size_t i = 0; i < inputs.size(); ++i)
{
// Trying to find transfomers in a stream section of the config.
auto inputSections = TryGetSectionsWithParameter(inputs[i], "transforms");
if (inputSections.size() > 1)
{
LogicError("Only a single 'transforms' config is allowed per stream.");
}
// No need to create anything for this stream, skipping.
if (inputSections.empty())
{
continue;
}
ConfigParameters input = inputs[i](inputSections.front());
std::wstring inputName = msra::strfun::utf16(input.ConfigName());
// Read tranformers in order and appending them to the transformer pipeline.
argvector<ConfigParameters> transforms = input("transforms");
for (size_t j = 0; j < transforms.size(); ++j)
{
TransformerPtr transformer = CreateTransformer(transforms[j], defaultModule);
m_transforms.push_back(Transformation{transformer, inputName});
}
}
}
// Create a transformer for a particular configuration. Loading it from the module of the deserializer if module is not specified, i.e.
// transforms = [
// [type = "Scale" width=...]:...
TransformerPtr CompositeDataReader::CreateTransformer(const ConfigParameters& config, const string& defaultModule)
{
typedef bool(*TransformerFactory) (Transformer** t, const std::wstring& type, const ConfigParameters& cfg);
std::string transformerModule = config("module", defaultModule.c_str());
TransformerFactory f = (TransformerFactory)Plugin::Load(transformerModule, "CreateTransformer");
std::wstring transformerType = config("type");
Transformer* t;
if (!f(&t, transformerType, config))
{
RuntimeError("Cannot create transformer. Please check the module and type in the configuration.");
}
assert(t != nullptr);
return TransformerPtr(t);
}
void CompositeDataReader::StartEpoch(const EpochConfiguration& cfg)
{
EpochConfiguration config = cfg;
@ -151,7 +245,7 @@ void CompositeDataReader::StartEpoch(const EpochConfiguration& cfg)
RuntimeError("Unsupported minibatch size '%d'.", (int)config.m_totalEpochSizeInSamples);
}
m_randomizer->StartEpoch(config);
m_sequenceEnumerator->StartEpoch(config);
// TODO: As the next step the packers should be moved into the network.
switch (m_packingMode)
@ -159,13 +253,13 @@ void CompositeDataReader::StartEpoch(const EpochConfiguration& cfg)
case PackingMode::sample:
m_packer = std::make_shared<FramePacker>(
m_provider,
m_randomizer,
m_sequenceEnumerator,
m_streams);
break;
case PackingMode::sequence:
m_packer = std::make_shared<SequencePacker>(
m_provider,
m_randomizer,
m_sequenceEnumerator,
m_streams);
break;
case PackingMode::truncated:
@ -173,7 +267,7 @@ void CompositeDataReader::StartEpoch(const EpochConfiguration& cfg)
config.m_truncationSize = m_truncationLength;
m_packer = std::make_shared<TruncatedBPTTPacker>(
m_provider,
m_randomizer,
m_sequenceEnumerator,
m_streams);
break;
}

19
Source/Readers/CompositeDataReader/CompositeDataReader.h
Просмотреть файл

@ -9,7 +9,9 @@
#include <string>
#include <future>
#include "DataReader.h"
#include <Reader.h>
#include "Reader.h"
#include "Transformer.h"
#include "TransformController.h"
namespace Microsoft { namespace MSR { namespace CNTK {
@ -44,13 +46,11 @@ struct Minibatch;
// - sequences can be transformed by the transformers applied on top of deserializer (TODO: not yet in place)
// - deserializers are bound together using the bundler - it bundles sequences with the same sequence id retrieved from different deserializers
// - packer is used to pack randomized sequences into the minibatch
// The composite data reader is currently also responsible for asynchronous prefetching of the minibatch data.
// The composite reader is currently also responsible for asynchronous prefetching of the minibatch data.
// In order not to break existing configs and allow deserializers composition it exposes the same interface as the old readers, but it is not exposed
// to external developers. The actual "reader developer" now has to provide deserializer(s) only.
// TODO: Implement proper corpus descriptor.
// TODO: Add transformers as the next step.
// TODO: Same code as in ReaderLib shim, the one in the ReaderLib will be deleted as the next step.
// TODO: Change this interface when SGD is changed.
class CompositeDataReader : public Reader, protected Plugin
{
@ -68,7 +68,10 @@ public:
private:
void CreateDeserializers(const ConfigParameters& readerConfig);
void CreateTransforms(const ConfigParameters& deserializerConfig);
IDataDeserializerPtr CreateDeserializer(const ConfigParameters& readerConfig, bool primary);
TransformerPtr CreateTransformer(const ConfigParameters& config, const std::string& defaultModule);
enum class PackingMode
@ -103,9 +106,11 @@ private:
// A list of deserializers.
std::vector<IDataDeserializerPtr> m_deserializers;
// Randomizer.
// TODO: remove Transformer interface from randomizer.
TransformerPtr m_randomizer;
// A list of transformers.
std::vector<Transformation> m_transforms;
// Sequence provider.
SequenceEnumeratorPtr m_sequenceEnumerator;
// TODO: Should be removed. We already have matrices on this level.
// Should just get the corresponding pinned memory.

2
Source/Readers/CompositeDataReader/CompositeDataReader.vcxproj.filters
Просмотреть файл

@ -3,8 +3,8 @@
<ItemGroup>
<ClCompile Include="dllmain.cpp" />
<ClCompile Include="stdafx.cpp" />
<ClCompile Include="CompositeDataReader.cpp" />
<ClCompile Include="Exports.cpp" />
<ClCompile Include="CompositeDataReader.cpp" />
</ItemGroup>
<ItemGroup>
<ClInclude Include="stdafx.h" />

14
Source/Readers/ExperimentalHTKMLFReader/ConfigHelper.cpp
Просмотреть файл

@ -198,8 +198,16 @@ size_t ConfigHelper::GetRandomizationWindow()
wstring ConfigHelper::GetRandomizer()
{
// get the read method, defaults to "blockRandomize"
wstring randomizer(m_config(L"readMethod", L"blockRandomize"));
// Check (on the action) if we're writing (inputs only) or training/evaluating (inputs and outputs)
bool isActionWrite = wstring(m_config(L"action", L"")) == L"write";
// Get the read method, defaults to "blockRandomize".
wstring randomizer = m_config(L"readMethod", L"blockRandomize");
if (isActionWrite && randomizer != L"none")
{
InvalidArgument("'readMethod' must be 'none' for write action.");
}
if (randomizer == L"blockRandomize" && GetRandomizationWindow() == randomizeNone)
{
@ -231,7 +239,7 @@ vector<wstring> ConfigHelper::GetSequencePaths()
// post processing file list :
// - if users specified PrefixPath, add the prefix to each of path in filelist
// - else do the dotdotdot expansion if necessary
if (!rootPath.empty()) // use has specified a path prefix for this feature
if (!rootPath.empty()) // user has specified a path prefix for this feature
{
// first make slash consistent (sorry for Linux users:this is not necessary for you)
replace(rootPath.begin(), rootPath.end(), L'\\', L'/');

18
Source/Readers/ExperimentalHTKMLFReader/HTKDataDeserializer.cpp
Просмотреть файл

@ -36,7 +36,7 @@ HTKDataDeserializer::HTKDataDeserializer(
// TODO: This should be read in one place, potentially given by SGD.
m_frameMode = (ConfigValue)cfg("frameMode", "true");
argvector<ConfigValue> inputs = cfg("inputs");
argvector<ConfigValue> inputs = cfg("input");
if (inputs.size() != 1)
{
InvalidArgument("HTKDataDeserializer supports a single input stream only.");
@ -127,22 +127,12 @@ void HTKDataDeserializer::InitializeChunkDescriptions(ConfigHelper& config)
}
wstring key = description.GetKey();
size_t id = 0;
if (m_primary)
if (!m_corpus->IsIncluded(key))
{
// TODO: Definition of the corpus should be moved to the CorpusDescriptor
// TODO: All keys should be added there. Currently, we add them in the driving deserializer.
id = stringRegistry.AddValue(key);
}
else
{
if (!stringRegistry.TryGet(key, id))
{
// Utterance is unknown, skipping it.
continue;
}
continue;
}
size_t id = stringRegistry[key];
description.SetId(id);
utterances.push_back(description);
m_totalNumberOfFrames += numberOfFrames;

9
Source/Readers/ExperimentalHTKMLFReader/HTKMLFReader.cpp
Просмотреть файл

@ -27,9 +27,9 @@ std::vector<IDataDeserializerPtr> CreateDeserializers(const ConfigParameters& re
ConfigHelper config(readerConfig);
config.GetDataNamesFromConfig(featureNames, labelNames, notused, notused);
if (featureNames.size() < 1 || labelNames.size() < 1)
if (featureNames.size() < 1)
{
InvalidArgument("Network needs at least 1 feature and 1 label specified.");
InvalidArgument("Network needs at least 1 feature specified.");
}
CorpusDescriptorPtr corpus = std::make_shared<CorpusDescriptor>();
@ -105,7 +105,8 @@ HTKMLFReader::HTKMLFReader(MemoryProviderPtr provider,
LogicError("Please specify at least a single input stream.");
}
auto bundler = std::make_shared<Bundler>(readerConfig, deserializers[0], deserializers, false);
bool cleanse = readerConfig(L"checkData", false);
auto bundler = std::make_shared<Bundler>(readerConfig, deserializers[0], deserializers, cleanse);
int verbosity = readerConfig(L"verbosity", 2);
std::wstring readMethod = config.GetRandomizer();
@ -123,8 +124,6 @@ HTKMLFReader::HTKMLFReader(MemoryProviderPtr provider,
RuntimeError("readMethod must be 'blockRandomize' or 'none'.");
}
m_randomizer->Initialize(nullptr, readerConfig);
// Create output stream descriptions (all dense)
for (auto d : deserializers)
{

4
Source/Readers/ExperimentalHTKMLFReader/HTKMLFReader.h
Просмотреть файл

@ -8,6 +8,7 @@
#include "Reader.h"
#include "Packer.h"
#include "Config.h"
#include "SequenceEnumerator.h"
namespace Microsoft { namespace MSR { namespace CNTK {
@ -54,8 +55,7 @@ private:
// Memory provider (TODO: this will possibly change in the near future.)
MemoryProviderPtr m_provider;
// TODO: Randomizer won't implement transformer interface in the near future.
TransformerPtr m_randomizer;
SequenceEnumeratorPtr m_randomizer;
// Truncation length for BPTT mode.
size_t m_truncationLength;

4
Source/Readers/ExperimentalHTKMLFReader/MLFDataDeserializer.cpp
Просмотреть файл

@ -54,7 +54,7 @@ MLFDataDeserializer::MLFDataDeserializer(CorpusDescriptorPtr corpus, const Confi
LogicError("Mlf deserializer does not support primary mode - it cannot control chunking.");
}
argvector<ConfigValue> inputs = cfg("inputs");
argvector<ConfigValue> inputs = cfg("input");
if (inputs.size() != 1)
{
LogicError("MLFDataDeserializer supports a single input stream only.");
@ -124,7 +124,7 @@ void MLFDataDeserializer::InitializeChunkDescriptions(CorpusDescriptorPtr corpus
description.m_isValid = true;
size_t totalFrames = 0;
auto& stringRegistry = corpus->GetStringRegistry();
const auto& stringRegistry = corpus->GetStringRegistry();
// TODO resize m_keyToSequence with number of IDs from string registry

2
Source/Readers/HTKMLFReader/HTKMLFReader.cpp
Просмотреть файл

@ -599,7 +599,7 @@ void HTKMLFReader<ElemType>::PrepareForTrainingOrTesting(const ConfigRecordType&
}
// Load all input and output data.
// Note that the terms features imply be real-valued quanities and
// Note that the terms features imply be real-valued quantities and
// labels imply categorical quantities, irrespective of whether they
// are inputs or targets for the network
// TODO: lots of code dup with the other Prepare function

5
Source/Readers/HTKMLFReader/htkfeatio.h
Просмотреть файл

@ -298,11 +298,9 @@ public:
#else
W.close(numframes);
#endif
#ifdef _WIN32 // BUGBUG: and on Linux??
// rename to final destination
// (This would only fail in strange circumstances such as accidental multiple processes writing to the same file.)
renameOrDie(tmppath, path);
#endif
}
};
@ -417,7 +415,8 @@ public:
if (xpath.empty())
malformed(pathParam);
e = msra::strfun::toint(consume(xpath, L"]"));
if (!xpath.empty())
// TODO \r should be handled elsewhere; refine this
if (!xpath.empty() && xpath != L"\r")
malformed(pathParam);
isarchive = true;
}

41
Source/Readers/ImageReader/Exports.cpp
Просмотреть файл

@ -11,7 +11,9 @@
#include "ReaderShim.h"
#include "ImageReader.h"
#include "HeapMemoryProvider.h"
#include "CudaMemoryProvider.h"
#include "ImageDataDeserializer.h"
#include "ImageTransformers.h"
#include "CorpusDescriptor.h"
namespace Microsoft { namespace MSR { namespace CNTK {
@ -32,4 +34,41 @@ extern "C" DATAREADER_API void GetReaderD(IDataReader** preader)
*preader = new ReaderShim<double>(factory);
}
// TODO: Not safe from the ABI perspective. Will be uglified to make the interface ABI.
// A factory method for creating image deserializers.
extern "C" DATAREADER_API bool CreateDeserializer(IDataDeserializer** deserializer, const std::wstring& type, const ConfigParameters& deserializerConfig, CorpusDescriptorPtr corpus, bool)
{
if (type == L"ImageDataDeserializer")
*deserializer = new ImageDataDeserializer(corpus, deserializerConfig);
else
// Unknown type.
return false;
// Deserializer created.
return true;
}
// A factory method for creating image transformers.
extern "C" DATAREADER_API bool CreateTransformer(Transformer** transformer, const std::wstring& type, const ConfigParameters& config)
{
if (type == L"Crop")
*transformer = new CropTransformer(config);
else if (type == L"Scale")
*transformer = new ScaleTransformer(config);
else if (type == L"Color")
*transformer = new ColorTransformer(config);
else if (type == L"Intensity")
*transformer = new IntensityTransformer(config);
else if (type == L"Mean")
*transformer = new MeanTransformer(config);
else if (type == L"Transpose")
*transformer = new TransposeTransformer(config);
else
// Unknown type.
return false;
// Transformer created.
return true;
}
}}}

24
Source/Readers/ImageReader/ImageConfigHelper.cpp
Просмотреть файл

@ -6,33 +6,15 @@
#include "stdafx.h"
#include "ImageConfigHelper.h"
#include "StringUtil.h"
#include "ConfigUtil.h"
namespace Microsoft { namespace MSR { namespace CNTK {
std::vector<std::string> GetSectionsWithParameter(const ConfigParameters& config, const std::string& parameterName)
{
std::vector<std::string> sectionNames;
for (const std::pair<std::string, ConfigParameters>& section : config)
{
if (section.second.ExistsCurrent(parameterName))
{
sectionNames.push_back(section.first);
}
}
if (sectionNames.empty())
{
RuntimeError("ImageReader requires %s parameter.", parameterName.c_str());
}
return sectionNames;
}
ImageConfigHelper::ImageConfigHelper(const ConfigParameters& config)
: m_dataFormat(CHW)
{
std::vector<std::string> featureNames = GetSectionsWithParameter(config, "width");
std::vector<std::string> labelNames = GetSectionsWithParameter(config, "labelDim");
std::vector<std::string> featureNames = GetSectionsWithParameter("ImageReader", config, "width");
std::vector<std::string> labelNames = GetSectionsWithParameter("ImageReader", config, "labelDim");
// REVIEW alexeyk: currently support only one feature and label section.
if (featureNames.size() != 1 || labelNames.size() != 1)

4
Source/Readers/ImageReader/ImageConfigHelper.h
Просмотреть файл

@ -68,12 +68,12 @@ public:
return m_cropType == CropType::MultiView10;
}
static CropType ParseCropType(const std::string &src);
private:
ImageConfigHelper(const ImageConfigHelper&) = delete;
ImageConfigHelper& operator=(const ImageConfigHelper&) = delete;
CropType ParseCropType(const std::string &src);
std::string m_mapPath;
std::vector<StreamDescriptionPtr> m_streams;
ImageLayoutKind m_dataFormat;

109
Source/Readers/ImageReader/ImageDataDeserializer.cpp
Просмотреть файл

@ -11,6 +11,8 @@
#include <limits>
#include "ImageDataDeserializer.h"
#include "ImageConfigHelper.h"
#include "StringUtil.h"
#include "ConfigUtil.h"
namespace Microsoft { namespace MSR { namespace CNTK {
@ -115,13 +117,66 @@ public:
}
};
// A new constructor to support new compositional configuration,
// that allows composition of deserializers and transforms on inputs.
ImageDataDeserializer::ImageDataDeserializer(CorpusDescriptorPtr corpus, const ConfigParameters& config)
{
ConfigParameters inputs = config("input");
std::vector<std::string> featureNames = GetSectionsWithParameter("ImageDataDeserializer", inputs, "transforms");
std::vector<std::string> labelNames = GetSectionsWithParameter("ImageDataDeserializer", inputs, "labelDim");
// TODO: currently support only one feature and label section.
if (featureNames.size() != 1 || labelNames.size() != 1)
{
RuntimeError(
"ImageReader currently supports a single feature and label stream. '%d' features , '%d' labels found.",
static_cast<int>(featureNames.size()),
static_cast<int>(labelNames.size()));
}
string precision = (ConfigValue)config("precision", "float");
// Feature stream.
ConfigParameters featureSection = inputs(featureNames[0]);
auto features = std::make_shared<StreamDescription>();
features->m_id = 0;
features->m_name = msra::strfun::utf16(featureSection.ConfigName());
features->m_storageType = StorageType::dense;
features->m_elementType = AreEqualIgnoreCase(precision, "float") ? ElementType::tfloat : ElementType::tdouble;
m_streams.push_back(features);
// Label stream.
ConfigParameters label = inputs(labelNames[0]);
size_t labelDimension = label("labelDim");
auto labels = std::make_shared<StreamDescription>();
labels->m_id = 1;
labels->m_name = msra::strfun::utf16(label.ConfigName());
labels->m_sampleLayout = std::make_shared<TensorShape>(labelDimension);
labels->m_storageType = StorageType::sparse_csc;
labels->m_elementType = AreEqualIgnoreCase(precision, "float") ? ElementType::tfloat : ElementType::tdouble;
m_streams.push_back(labels);
m_labelGenerator = labels->m_elementType == ElementType::tfloat ?
(LabelGeneratorPtr)std::make_shared<TypedLabelGenerator<float>>(labelDimension) :
std::make_shared<TypedLabelGenerator<double>>(labelDimension);
m_grayscale = config(L"grayscale", false);
// TODO: multiview should be done on the level of randomizer/transformers - it is responsiblity of the
// TODO: randomizer to collect how many copies each transform needs and request same sequence several times.
bool multiViewCrop = config(L"multiViewCrop", false);
CreateSequenceDescriptions(corpus, config(L"file"), labelDimension, multiViewCrop);
}
// TODO: Should be removed at some point.
// Supports old type of ImageReader configuration.
ImageDataDeserializer::ImageDataDeserializer(const ConfigParameters& config)
{
ImageConfigHelper configHelper(config);
m_streams = configHelper.GetStreams();
assert(m_streams.size() == 2);
m_grayscale = configHelper.UseGrayscale();
const auto& label = m_streams[configHelper.GetLabelStreamId()];
const auto& label = m_streams[configHelper.GetLabelStreamId()];
const auto& feature = m_streams[configHelper.GetFeatureStreamId()];
// Expect data in HWC.
@ -147,7 +202,7 @@ ImageDataDeserializer::ImageDataDeserializer(const ConfigParameters& config)
RuntimeError("Unsupported label element type '%d'.", (int)label->m_elementType);
}
CreateSequenceDescriptions(configHelper.GetMapPath(), labelDimension, configHelper);
CreateSequenceDescriptions(std::make_shared<CorpusDescriptor>(), configHelper.GetMapPath(), labelDimension, configHelper.IsMultiViewCrop());
}
// Descriptions of chunks exposed by the image reader.
@ -173,7 +228,7 @@ void ImageDataDeserializer::GetSequencesForChunk(size_t chunkId, std::vector<Seq
result.push_back(m_imageSequences[chunkId]);
}
void ImageDataDeserializer::CreateSequenceDescriptions(std::string mapPath, size_t labelDimension, const ImageConfigHelper& config)
void ImageDataDeserializer::CreateSequenceDescriptions(CorpusDescriptorPtr corpus, std::string mapPath, size_t labelDimension, bool isMultiCrop)
{
std::ifstream mapFile(mapPath);
if (!mapFile)
@ -181,7 +236,7 @@ void ImageDataDeserializer::CreateSequenceDescriptions(std::string mapPath, size
RuntimeError("Could not open %s for reading.", mapPath.c_str());
}
size_t itemsPerLine = config.IsMultiViewCrop() ? 10 : 1;
size_t itemsPerLine = isMultiCrop ? 10 : 1;
size_t curId = 0;
std::string line;
PathReaderMap knownReaders;
@ -189,13 +244,30 @@ void ImageDataDeserializer::CreateSequenceDescriptions(std::string mapPath, size
description.m_numberOfSamples = 1;
description.m_isValid = true;
auto& stringRegistry = corpus->GetStringRegistry();
for (size_t lineIndex = 0; std::getline(mapFile, line); ++lineIndex)
{
std::stringstream ss(line);
std::string imagePath;
std::string classId;
if (!std::getline(ss, imagePath, '\t') || !std::getline(ss, classId, '\t'))
RuntimeError("Invalid map file format, must contain 2 tab-delimited columns, line %" PRIu64 " in file %s.", lineIndex, mapPath.c_str());
std::string imagePath, classId, sequenceKey;
// Try to parse sequence id, file path and label.
if (!std::getline(ss, sequenceKey, '\t') || !std::getline(ss, imagePath, '\t') || !std::getline(ss, classId, '\t'))
{
// In case when the sequence key is not specified we set it to the line number inside the mapping file.
// Assume that only image path and class label is given (old format).
classId = imagePath;
imagePath = sequenceKey;
sequenceKey = std::to_string(lineIndex);
if (classId.empty() || imagePath.empty())
RuntimeError("Invalid map file format, must contain 2 or 3 tab-delimited columns, line %" PRIu64 " in file %s.", lineIndex, mapPath.c_str());
}
// Skipping sequences that are not included in corpus.
auto key = msra::strfun::utf16(sequenceKey);
if (!corpus->IsIncluded(key))
{
continue;
}
char* eptr;
errno = 0;
@ -216,9 +288,10 @@ void ImageDataDeserializer::CreateSequenceDescriptions(std::string mapPath, size
description.m_chunkId = curId;
description.m_path = imagePath;
description.m_classId = cid;
description.m_key.m_sequence = description.m_id;
description.m_key.m_sequence = stringRegistry[key];
description.m_key.m_sample = 0;
m_keyToSequence[description.m_key.m_sequence] = m_imageSequences.size();
m_imageSequences.push_back(description);
RegisterByteReader(description.m_id, description.m_path, knownReaders);
}
@ -280,8 +353,24 @@ cv::Mat ImageDataDeserializer::ReadImage(size_t seqId, const std::string& path,
cv::Mat FileByteReader::Read(size_t, const std::string& path, bool grayscale)
{
assert(!path.empty());
assert(!path.empty());
return cv::imread(path, grayscale ? cv::IMREAD_GRAYSCALE : cv::IMREAD_COLOR);
}
static SequenceDescription s_invalidSequence{0, 0, 0, false};
void ImageDataDeserializer::GetSequenceDescriptionByKey(const KeyType& key, SequenceDescription& result)
{
auto index = m_keyToSequence.find(key.m_sequence);
// Checks whether it is a known sequence for us.
if (key.m_sample != 0 || index == m_keyToSequence.end())
{
result = s_invalidSequence;
return;
}
result = m_imageSequences[index->second];
}
}}}

15
Source/Readers/ImageReader/ImageDataDeserializer.h
Просмотреть файл

@ -8,8 +8,8 @@
#include "DataDeserializerBase.h"
#include "Config.h"
#include "ByteReader.h"
#include "ImageConfigHelper.h"
#include <unordered_map>
#include "CorpusDescriptor.h"
namespace Microsoft { namespace MSR { namespace CNTK {
@ -21,6 +21,11 @@ namespace Microsoft { namespace MSR { namespace CNTK {
class ImageDataDeserializer : public DataDeserializerBase
{
public:
// A new constructor to support new compositional configuration,
// that allows composition of deserializers and transforms on inputs.
ImageDataDeserializer(CorpusDescriptorPtr corpus, const ConfigParameters& config);
// TODO: This constructor should be deprecated in the future. Compositional config should be used instead.
explicit ImageDataDeserializer(const ConfigParameters& config);
// Gets sequences by specified ids. Order of returned sequences corresponds to the order of provided ids.
@ -32,9 +37,12 @@ public:
// Gets sequence descriptions for the chunk.
virtual void GetSequencesForChunk(size_t, std::vector<SequenceDescription>&) override;
// Gets sequence description by key.
void GetSequenceDescriptionByKey(const KeyType&, SequenceDescription&) override;
private:
// Creates a set of sequence descriptions.
void CreateSequenceDescriptions(std::string mapPath, size_t labelDimension, const ImageConfigHelper& config);
void CreateSequenceDescriptions(CorpusDescriptorPtr corpus, std::string mapPath, size_t labelDimension, bool isMultiCrop);
// Image sequence descriptions. Currently, a sequence contains a single sample only.
struct ImageSequenceDescription : public SequenceDescription
@ -53,6 +61,9 @@ private:
// Sequence descriptions for all input data.
std::vector<ImageSequenceDescription> m_imageSequences;
// Mapping of logical sequence key into sequence description.
std::map<size_t, size_t> m_keyToSequence;
// Element type of the feature/label stream (currently float/double only).
ElementType m_featureElementType;

40
Source/Readers/ImageReader/ImageReader.cpp
Просмотреть файл

@ -13,9 +13,13 @@
#include "ImageDataDeserializer.h"
#include "FramePacker.h"
#include <omp.h>
#include "TransformController.h"
namespace Microsoft { namespace MSR { namespace CNTK {
// TODO: This class should go away eventually.
// TODO: The composition of packer + randomizer + different deserializers in a generic manner is done in the CompositeDataReader.
// TODO: Currently preserving this for backward compatibility with current configs.
ImageReader::ImageReader(MemoryProviderPtr provider,
const ConfigParameters& config)
: m_seed(0), m_provider(provider)
@ -37,7 +41,7 @@ ImageReader::ImageReader(MemoryProviderPtr provider,
auto deserializer = std::make_shared<ImageDataDeserializer>(config);
TransformerPtr randomizer;
SequenceEnumeratorPtr randomizer;
// Request multi-threaded randomizer operation to speed up CPU-intensive image-decoding and transformations.
const bool multithreadedGetNextSequences = true;
if (configHelper.ShouldRandomize())
@ -51,35 +55,27 @@ ImageReader::ImageReader(MemoryProviderPtr provider,
randomizer = std::make_shared<NoRandomizer>(deserializer, multithreadedGetNextSequences);
}
randomizer->Initialize(nullptr, config);
// Create transformations for a single feature stream.
std::wstring featureName = m_streams[configHelper.GetFeatureStreamId()]->m_name;
ConfigParameters featureStream = config(featureName);
auto cropper = std::make_shared<CropTransformer>();
cropper->Initialize(randomizer, config);
std::vector<Transformation> transformations;
transformations.push_back(Transformation{ std::make_shared<CropTransformer>(featureStream), featureName });
transformations.push_back(Transformation{ std::make_shared<ScaleTransformer>(featureStream), featureName });
transformations.push_back(Transformation{ std::make_shared<ColorTransformer>(featureStream), featureName });
transformations.push_back(Transformation{ std::make_shared<IntensityTransformer>(featureStream), featureName });
transformations.push_back(Transformation{ std::make_shared<MeanTransformer>(featureStream), featureName });
auto scaler = std::make_shared<ScaleTransformer>();
scaler->Initialize(cropper, config);
auto color = std::make_shared<ColorTransformer>();
color->Initialize(scaler, config);
auto intensity = std::make_shared<IntensityTransformer>();
intensity->Initialize(color, config);
auto mean = std::make_shared<MeanTransformer>();
mean->Initialize(intensity, config);
TransformerPtr last = mean;
if (configHelper.GetDataFormat() == CHW)
{
last = std::make_shared<TransposeTransformer>();
last->Initialize(mean, config);
transformations.push_back(Transformation{ std::make_shared<TransposeTransformer>(featureStream), featureName });
}
m_transformer = last;
m_sequenceEnumerator = std::make_shared<TransformController>(transformations, randomizer);
m_packer = std::make_shared<FramePacker>(
m_provider,
m_transformer,
m_sequenceEnumerator,
m_streams);
}
@ -96,7 +92,7 @@ void ImageReader::StartEpoch(const EpochConfiguration& config)
RuntimeError("Epoch size cannot be 0.");
}
m_transformer->StartEpoch(config);
m_sequenceEnumerator->StartEpoch(config);
m_packer->StartEpoch(config);
}

4
Source/Readers/ImageReader/ImageReader.h
Просмотреть файл

@ -6,8 +6,8 @@
#pragma once
#include "Reader.h"
#include "ImageTransformers.h"
#include "Packer.h"
#include "SequenceEnumerator.h"
namespace Microsoft { namespace MSR { namespace CNTK {
@ -34,7 +34,7 @@ private:
std::vector<StreamDescriptionPtr> m_streams;
// A head transformer in a list of transformers.
TransformerPtr m_transformer;
SequenceEnumeratorPtr m_sequenceEnumerator;
// Packer.
PackerPtr m_packer;

12
Source/Readers/ImageReader/ImageReader.vcxproj.filters
Просмотреть файл

@ -3,23 +3,11 @@
<ItemGroup>
<ClCompile Include="Exports.cpp" />
<ClCompile Include="stdafx.cpp" />
<ClCompile Include="..\..\Common\fileutil.cpp">
<Filter>Common</Filter>
</ClCompile>
<ClCompile Include="..\..\Common\File.cpp">
<Filter>Common</Filter>
</ClCompile>
<ClCompile Include="dllmain.cpp" />
<ClCompile Include="ImageTransformers.cpp" />
<ClCompile Include="ImageDataDeserializer.cpp" />
<ClCompile Include="ImageReader.cpp" />
<ClCompile Include="ImageConfigHelper.cpp" />
<ClCompile Include="..\..\Common\Config.cpp">
<Filter>Common</Filter>
</ClCompile>
<ClCompile Include="..\..\Common\ExceptionWithCallStack.cpp">
<Filter>Common</Filter>
</ClCompile>
<ClCompile Include="ZipByteReader.cpp" />
</ItemGroup>
<ItemGroup>

228
Source/Readers/ImageReader/ImageTransformers.cpp
Просмотреть файл

@ -13,7 +13,7 @@
#include "StringUtil.h"
#include "ElementTypeUtils.h"
namespace Microsoft { namespace MSR { namespace CNTK
namespace Microsoft { namespace MSR { namespace CNTK
{
struct ImageSequenceData : DenseSequenceData
@ -23,54 +23,51 @@ struct ImageSequenceData : DenseSequenceData
SequenceDataPtr m_original;
};
void ImageTransformerBase::Initialize(TransformerPtr next,
const ConfigParameters &readerConfig)
ImageTransformerBase::ImageTransformerBase(const ConfigParameters& readerConfig) : m_imageElementType(0)
{
Base::Initialize(next, readerConfig);
m_imageConfig = std::make_unique<ImageConfigHelper>(readerConfig);
m_seed = readerConfig(L"seed", (unsigned int)0);
size_t featureStreamId = m_imageConfig->GetFeatureStreamId();
m_appliedStreamIds.push_back(featureStreamId);
if (m_appliedStreamIds.size() != 1)
{
RuntimeError("Only a single feature stream is supported.");
}
const auto &inputStreams = GetInputStreams();
m_outputStreams.resize(inputStreams.size());
std::copy(inputStreams.begin(), inputStreams.end(), m_outputStreams.begin());
m_seed = readerConfig(L"seed", 0u);
}
SequenceDataPtr
ImageTransformerBase::Apply(SequenceDataPtr sequence,
const StreamDescription &inputStream,
const StreamDescription & /*outputStream*/)
// The method describes how input stream is transformed to the output stream. Called once per applied stream.
// Currently for image transformations we only support dense streams of type double or float.
StreamDescription ImageTransformerBase::Transform(const StreamDescription& inputStream)
{
assert(inputStream.m_storageType == StorageType::dense);
auto inputSequence = static_cast<const DenseSequenceData&>(*sequence.get());
ImageDimensions dimensions(*inputSequence.m_sampleLayout, HWC);
int columns = static_cast<int>(dimensions.m_width);
int rows = static_cast<int>(dimensions.m_height);
int channels = static_cast<int>(dimensions.m_numChannels);
m_inputStream = inputStream;
m_outputStream = m_inputStream;
int typeId = 0;
if (inputStream.m_elementType == ElementType::tdouble)
if (m_inputStream.m_storageType != StorageType::dense)
{
typeId = CV_64F;
LogicError("ImageTransformerBase supports only dense input streams.");
}
else if (inputStream.m_elementType == ElementType::tfloat)
if (m_inputStream.m_elementType == ElementType::tdouble)
{
typeId = CV_32F;
m_imageElementType = CV_64F;
}
else if (m_inputStream.m_elementType == ElementType::tfloat)
{
m_imageElementType = CV_32F;
}
else
{
RuntimeError("Unsupported type");
}
return m_outputStream;
}
// Transforms a single sequence as open cv dense image. Called once per sequence.
SequenceDataPtr ImageTransformerBase::Transform(SequenceDataPtr sequence)
{
auto inputSequence = static_cast<const DenseSequenceData&>(*sequence);
ImageDimensions dimensions(*inputSequence.m_sampleLayout, HWC);
int columns = static_cast<int>(dimensions.m_width);
int rows = static_cast<int>(dimensions.m_height);
int channels = static_cast<int>(dimensions.m_numChannels);
auto result = std::make_shared<ImageSequenceData>();
int type = CV_MAKETYPE(typeId, channels);
int type = CV_MAKETYPE(m_imageElementType, channels);
cv::Mat buffer = cv::Mat(rows, columns, type, inputSequence.m_data);
Apply(sequence->m_id, buffer);
if (!buffer.isContinuous())
@ -92,16 +89,7 @@ ImageTransformerBase::Apply(SequenceDataPtr sequence,
}
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
void CropTransformer::Initialize(TransformerPtr next,
const ConfigParameters &readerConfig)
{
ImageTransformerBase::Initialize(next, readerConfig);
auto featureStreamIds = GetAppliedStreamIds();
InitFromConfig(readerConfig(GetInputStreams()[featureStreamIds[0]]->m_name));
}
void CropTransformer::InitFromConfig(const ConfigParameters &config)
CropTransformer::CropTransformer(const ConfigParameters& config) : ImageTransformerBase(config)
{
floatargvector cropRatio = config(L"cropRatio", "1.0");
m_cropRatioMin = cropRatio[0];
@ -117,9 +105,11 @@ void CropTransformer::InitFromConfig(const ConfigParameters &config)
m_jitterType = ParseJitterType(config(L"jitterType", ""));
m_cropType = ImageConfigHelper::ParseCropType(config(L"cropType", ""));
if (!config.ExistsCurrent(L"hflip"))
{
m_hFlip = m_imageConfig->GetCropType() == CropType::Random;
m_hFlip = m_cropType == CropType::Random;
}
else
{
@ -134,7 +124,6 @@ void CropTransformer::StartEpoch(const EpochConfiguration &config)
m_curAspectRatioRadius = m_aspectRatioRadius[config.m_epochIndex];
if (!(0 <= m_curAspectRatioRadius && m_curAspectRatioRadius <= 1.0))
InvalidArgument("aspectRatioRadius must be >= 0.0 and <= 1.0");
ImageTransformerBase::StartEpoch(config);
}
@ -164,9 +153,9 @@ void CropTransformer::Apply(size_t id, cv::Mat &mat)
RuntimeError("Jitter type currently not implemented.");
}
int viewIndex = m_imageConfig->IsMultiViewCrop() ? (int)(id % 10) : 0;
int viewIndex = m_cropType == CropType::MultiView10 ? (int)(id % 10) : 0;
mat = mat(GetCropRect(m_imageConfig->GetCropType(), viewIndex, mat.rows, mat.cols, ratio, *rng));
mat = mat(GetCropRect(m_cropType, viewIndex, mat.rows, mat.cols, ratio, *rng));
if ((m_hFlip && std::bernoulli_distribution()(*rng)) ||
viewIndex >= 5)
{
@ -176,7 +165,8 @@ void CropTransformer::Apply(size_t id, cv::Mat &mat)
m_rngs.push(std::move(rng));
}
CropTransformer::RatioJitterType CropTransformer::ParseJitterType(const std::string &src)
CropTransformer::RatioJitterType
CropTransformer::ParseJitterType(const std::string &src)
{
if (src.empty() || AreEqualIgnoreCase(src, "none"))
{
@ -202,7 +192,7 @@ CropTransformer::RatioJitterType CropTransformer::ParseJitterType(const std::str
}
cv::Rect CropTransformer::GetCropRect(CropType type, int viewIndex, int crow, int ccol,
double cropRatio, std::mt19937 &rng)
double cropRatio, std::mt19937 &rng)
{
assert(crow > 0);
assert(ccol > 0);
@ -254,27 +244,27 @@ cv::Rect CropTransformer::GetCropRect(CropType type, int viewIndex, int crow, in
int isubView = viewIndex % 5;
switch (isubView)
{
// top-left
// top-left
case 0:
xOff = 0;
yOff = 0;
break;
// top-right
// top-right
case 1:
xOff = ccol - cropSizeX;
yOff = 0;
break;
// bottom-left
// bottom-left
case 2:
xOff = 0;
yOff = crow - cropSizeY;
break;
// bottom-right
// bottom-right
case 3:
xOff = ccol - cropSizeX;
yOff = crow - cropSizeY;
break;
// center
// center
case 4:
xOff = (ccol - cropSizeX) / 2;
yOff = (crow - cropSizeY) / 2;
@ -293,24 +283,13 @@ cv::Rect CropTransformer::GetCropRect(CropType type, int viewIndex, int crow, in
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
void ScaleTransformer::Initialize(TransformerPtr next,
const ConfigParameters &readerConfig)
ScaleTransformer::ScaleTransformer(const ConfigParameters& config) : ImageTransformerBase(config)
{
ImageTransformerBase::Initialize(next, readerConfig);
m_interpMap.emplace("nearest", cv::INTER_NEAREST);
m_interpMap.emplace("linear", cv::INTER_LINEAR);
m_interpMap.emplace("cubic", cv::INTER_CUBIC);
m_interpMap.emplace("lanczos", cv::INTER_LANCZOS4);
auto featureStreamIds = GetAppliedStreamIds();
const auto &feature = GetInputStreams()[featureStreamIds[0]];
m_dataType = feature->m_elementType == ElementType::tfloat ? CV_32F : CV_64F;
InitFromConfig(readerConfig(feature->m_name));
}
void ScaleTransformer::InitFromConfig(const ConfigParameters &config)
{
m_imgWidth = config(L"width");
m_imgHeight = config(L"height");
m_imgChannels = config(L"channels");
@ -335,15 +314,24 @@ void ScaleTransformer::InitFromConfig(const ConfigParameters &config)
m_interp.push_back(cv::INTER_LINEAR);
}
// The method describes how input stream is transformed to the output stream. Called once per applied stream.
// Scale transformer transforms the stream so that all samples are of the same size.
StreamDescription ScaleTransformer::Transform(const StreamDescription& inputStream)
{
ImageTransformerBase::Transform(inputStream);
m_outputStream.m_sampleLayout = std::make_shared<TensorShape>(ImageDimensions(m_imgWidth, m_imgHeight, m_imgChannels).AsTensorShape(HWC));
return m_outputStream;
}
void ScaleTransformer::Apply(size_t id, cv::Mat &mat)
{
UNUSED(id);
// If matrix has not been converted to the right type, do it now as rescaling
// requires floating point type.
//
if (mat.type() != CV_MAKETYPE(m_dataType, m_imgChannels))
if (mat.type() != CV_MAKETYPE(m_imageElementType, m_imgChannels))
{
mat.convertTo(mat, m_dataType);
mat.convertTo(mat, m_imageElementType);
}
auto seed = GetSeed();
@ -351,7 +339,6 @@ void ScaleTransformer::Apply(size_t id, cv::Mat &mat)
auto index = UniIntT(0, static_cast<int>(m_interp.size()) - 1)(*rng);
assert(m_interp.size() > 0);
cv::resize(mat, mat, cv::Size((int)m_imgWidth, (int)m_imgHeight), 0, 0, m_interp[index]);
m_rngs.push(std::move(rng));
@ -359,16 +346,7 @@ void ScaleTransformer::Apply(size_t id, cv::Mat &mat)
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
void MeanTransformer::Initialize(TransformerPtr next,
const ConfigParameters &readerConfig)
{
ImageTransformerBase::Initialize(next, readerConfig);
auto featureStreamIds = GetAppliedStreamIds();
InitFromConfig(readerConfig(GetInputStreams()[featureStreamIds[0]]->m_name));
}
void MeanTransformer::InitFromConfig(const ConfigParameters &config)
MeanTransformer::MeanTransformer(const ConfigParameters& config) : ImageTransformerBase(config)
{
std::wstring meanFile = config(L"meanFile", L"");
if (meanFile.empty())
@ -401,7 +379,7 @@ void MeanTransformer::Apply(size_t id, cv::Mat &mat)
UNUSED(id);
assert(m_meanImg.size() == cv::Size(0, 0) ||
(m_meanImg.size() == mat.size() &&
m_meanImg.channels() == mat.channels()));
m_meanImg.channels() == mat.channels()));
// REVIEW alexeyk: check type conversion (float/double).
if (m_meanImg.size() == mat.size())
@ -410,48 +388,34 @@ void MeanTransformer::Apply(size_t id, cv::Mat &mat)
}
}
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
void TransposeTransformer::Initialize(TransformerPtr next,
const ConfigParameters &readerConfig)
// The method describes how input stream is transformed to the output stream. Called once per applied stream.
// Transpose transformer expects the dense input stream with samples as HWC and outputs CHW.
StreamDescription TransposeTransformer::Transform(const StreamDescription& inputStream)
{
TransformerBase::Initialize(next, readerConfig);
// Currently we only support a single stream.
ImageConfigHelper config(readerConfig);
size_t featureStreamId = config.GetFeatureStreamId();
m_appliedStreamIds.push_back(featureStreamId);
const auto &inputStreams = GetInputStreams();
m_outputStreams.resize(inputStreams.size());
std::copy(inputStreams.begin(), inputStreams.end(), m_outputStreams.begin());
for (auto id : m_appliedStreamIds)
m_inputStream = inputStream;
if (m_inputStream.m_storageType != StorageType::dense)
{
auto &stream = inputStreams[id];
ImageDimensions dimensions(*stream->m_sampleLayout, HWC);
// Changing from NHWC to NCHW (note: row-major notation)
auto changedStream = std::make_shared<StreamDescription>(*stream);
changedStream->m_sampleLayout = std::make_shared<TensorShape>(dimensions.AsTensorShape(CHW));
m_outputStreams[id] = changedStream;
LogicError("Transpose transformer supports only dense streams.");
}
// Changing from NHWC to NCHW
ImageDimensions dimensions(*m_inputStream.m_sampleLayout, HWC);
m_outputStream = m_inputStream;
m_outputStream.m_sampleLayout = std::make_shared<TensorShape>(dimensions.AsTensorShape(CHW));
return m_outputStream;
}
SequenceDataPtr
TransposeTransformer::Apply(SequenceDataPtr inputSequence,
const StreamDescription &inputStream,
const StreamDescription &outputStream)
// Transformation of the sequence.
SequenceDataPtr TransposeTransformer::Transform(SequenceDataPtr sequence)
{
if (inputStream.m_elementType == ElementType::tdouble)
if (m_inputStream.m_elementType == ElementType::tdouble)
{
return TypedApply<double>(inputSequence, inputStream, outputStream);
return TypedTransform<double>(sequence);
}
if (inputStream.m_elementType == ElementType::tfloat)
if (m_inputStream.m_elementType == ElementType::tfloat)
{
return TypedApply<float>(inputSequence, inputStream, outputStream);
return TypedTransform<float>(sequence);
}
RuntimeError("Unsupported type");
@ -466,21 +430,17 @@ struct DenseSequenceWithBuffer : DenseSequenceData
};
template <class TElemType>
SequenceDataPtr TransposeTransformer::TypedApply(SequenceDataPtr sequence,
const StreamDescription &inputStream,
const StreamDescription &outputStream)
SequenceDataPtr TransposeTransformer::TypedTransform(SequenceDataPtr sequence)
{
assert(inputStream.m_storageType == StorageType::dense);
auto inputSequence = static_cast<DenseSequenceData&>(*sequence.get());
auto inputSequence = static_cast<DenseSequenceData&>(*sequence);
assert(inputSequence.m_numberOfSamples == 1);
assert(inputStream.m_sampleLayout->GetNumElements() == outputStream.m_sampleLayout->GetNumElements());
size_t count = inputStream.m_sampleLayout->GetNumElements() * GetSizeByType(inputStream.m_elementType);
size_t count = m_inputStream.m_sampleLayout->GetNumElements() * GetSizeByType(m_inputStream.m_elementType);
auto result = std::make_shared<DenseSequenceWithBuffer>();
result->m_buffer.resize(count);
ImageDimensions dimensions(*inputStream.m_sampleLayout, ImageLayoutKind::HWC);
ImageDimensions dimensions(*m_inputStream.m_sampleLayout, ImageLayoutKind::HWC);
size_t rowCount = dimensions.m_height * dimensions.m_width;
size_t channelCount = dimensions.m_numChannels;
@ -495,7 +455,7 @@ SequenceDataPtr TransposeTransformer::TypedApply(SequenceDataPtr sequence,
}
}
result->m_sampleLayout = outputStream.m_sampleLayout;
result->m_sampleLayout = m_outputStream.m_sampleLayout;
result->m_data = result->m_buffer.data();
result->m_numberOfSamples = inputSequence.m_numberOfSamples;
return result;
@ -503,16 +463,7 @@ SequenceDataPtr TransposeTransformer::TypedApply(SequenceDataPtr sequence,
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
void IntensityTransformer::Initialize(TransformerPtr next,
const ConfigParameters &readerConfig)
{
ImageTransformerBase::Initialize(next, readerConfig);
auto featureStreamIds = GetAppliedStreamIds();
InitFromConfig(readerConfig(GetInputStreams()[featureStreamIds[0]]->m_name));
}
void IntensityTransformer::InitFromConfig(const ConfigParameters &config)
IntensityTransformer::IntensityTransformer(const ConfigParameters &config) : ImageTransformerBase(config)
{
m_stdDev = config(L"intensityStdDev", ConfigParameters::Array(doubleargvector(vector<double>{0.0})));
std::wstring intFile = config(L"intensityFile", L"");
@ -540,7 +491,6 @@ void IntensityTransformer::InitFromConfig(const ConfigParameters &config)
void IntensityTransformer::StartEpoch(const EpochConfiguration &config)
{
m_curStdDev = m_stdDev[config.m_epochIndex];
ImageTransformerBase::StartEpoch(config);
}
@ -594,15 +544,7 @@ void IntensityTransformer::Apply(cv::Mat &mat)
//////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////////
void ColorTransformer::Initialize(TransformerPtr next, const ConfigParameters &readerConfig)
{
ImageTransformerBase::Initialize(next, readerConfig);
auto featureStreamIds = GetAppliedStreamIds();
InitFromConfig(readerConfig(GetInputStreams()[featureStreamIds[0]]->m_name));
}
void ColorTransformer::InitFromConfig(const ConfigParameters &config)
ColorTransformer::ColorTransformer(const ConfigParameters &config) : ImageTransformerBase(config)
{
m_brightnessRadius = config(L"brightnessRadius", ConfigParameters::Array(doubleargvector(vector<double>{0.0})));
m_contrastRadius = config(L"contrastRadius", ConfigParameters::Array(doubleargvector(vector<double>{0.0})));

93
Source/Readers/ImageReader/ImageTransformers.h
Просмотреть файл

@ -11,7 +11,6 @@
#include "Transformer.h"
#include "ConcStack.h"
#include "TransformerBase.h"
#include "Config.h"
#include "ImageConfigHelper.h"
@ -21,48 +20,39 @@ class ConfigParameters;
// Base class for image transformations based on OpenCV
// that helps to wrap the sequences into OpenCV::Mat class.
class ImageTransformerBase : public TransformerBase
class ImageTransformerBase : public Transformer
{
public:
// Initializes the transformer.
void Initialize(TransformerPtr next, const ConfigParameters &readerConfig) override;
explicit ImageTransformerBase(const ConfigParameters& config);
void StartEpoch(const EpochConfiguration&) override {}
// Transformation of the stream.
StreamDescription Transform(const StreamDescription& inputStream) override;
// Transformation of the sequence.
SequenceDataPtr Transform(SequenceDataPtr sequence) override;
protected:
const std::vector<StreamId> &GetAppliedStreamIds() const override
{
return m_appliedStreamIds;
}
const std::vector<StreamDescriptionPtr>& GetOutputStreams() const override
{
return m_outputStreams;
}
// Seed getter.
unsigned int GetSeed() const
{
return m_seed;
}
using Base = TransformerBase;
using Base = Transformer;
using UniRealT = std::uniform_real_distribution<double>;
using UniIntT = std::uniform_int_distribution<int>;
// Applies transformation to the sequence.
SequenceDataPtr Apply(SequenceDataPtr inputSequence,
const StreamDescription &inputStream,
const StreamDescription &outputStream) override;
// The only function that should be redefined by the inherited classes.
virtual void Apply(size_t id, cv::Mat &from) = 0;
protected:
std::unique_ptr<ImageConfigHelper> m_imageConfig;
private:
std::vector<StreamDescriptionPtr> m_outputStreams;
std::vector<StreamId> m_appliedStreamIds;
StreamDescription m_inputStream;
StreamDescription m_outputStream;
unsigned int m_seed;
int m_imageElementType;
conc_stack<std::unique_ptr<std::mt19937>> m_rngs;
};
// Crop transformation of the image.
@ -70,7 +60,7 @@ private:
class CropTransformer : public ImageTransformerBase
{
public:
void Initialize(TransformerPtr next, const ConfigParameters &readerConfig) override;
explicit CropTransformer(const ConfigParameters& config);
private:
void Apply(size_t id, cv::Mat &mat) override;
@ -84,14 +74,13 @@ private:
UniArea = 3
};
void InitFromConfig(const ConfigParameters &config);
void StartEpoch(const EpochConfiguration &config) override;
RatioJitterType ParseJitterType(const std::string &src);
cv::Rect GetCropRect(CropType type, int viewIndex, int crow, int ccol, double cropRatio, std::mt19937 &rng);
conc_stack<std::unique_ptr<std::mt19937>> m_rngs;
CropType m_cropType;
double m_cropRatioMin;
double m_cropRatioMax;
RatioJitterType m_jitterType;
@ -105,11 +94,11 @@ private:
class ScaleTransformer : public ImageTransformerBase
{
public:
void Initialize(TransformerPtr next,
const ConfigParameters &readerConfig) override;
explicit ScaleTransformer(const ConfigParameters& config);
StreamDescription Transform(const StreamDescription& inputStream) override;
private:
void InitFromConfig(const ConfigParameters &config);
void Apply(size_t id, cv::Mat &mat) override;
using StrToIntMapT = std::unordered_map<std::string, int>;
@ -117,7 +106,6 @@ private:
std::vector<int> m_interp;
conc_stack<std::unique_ptr<std::mt19937>> m_rngs;
int m_dataType;
size_t m_imgWidth;
size_t m_imgHeight;
size_t m_imgChannels;
@ -127,45 +115,34 @@ private:
class MeanTransformer : public ImageTransformerBase
{
public:
void Initialize(TransformerPtr next,
const ConfigParameters &readerConfig) override;
explicit MeanTransformer(const ConfigParameters& config);
private:
void Apply(size_t id, cv::Mat &mat) override;
void InitFromConfig(const ConfigParameters &config);
cv::Mat m_meanImg;
};
// Transpose transformation from HWC to CHW (note: row-major notation).
class TransposeTransformer : public TransformerBase
class TransposeTransformer : public Transformer
{
public:
void Initialize(TransformerPtr next, const ConfigParameters &readerConfig) override;
explicit TransposeTransformer(const ConfigParameters&) {}
protected:
const std::vector<StreamId>& GetAppliedStreamIds() const override
{
return m_appliedStreamIds;
}
void StartEpoch(const EpochConfiguration&) override {}
const std::vector<StreamDescriptionPtr>& GetOutputStreams() const override
{
return m_outputStreams;
}
// Transformation of the stream.
StreamDescription Transform(const StreamDescription& inputStream) override;
SequenceDataPtr Apply(SequenceDataPtr inputSequence,
const StreamDescription &inputStream,
const StreamDescription &outputStream) override;
// Transformation of the sequence.
SequenceDataPtr Transform(SequenceDataPtr sequence) override;
private:
template <class TElement>
SequenceDataPtr TypedApply(SequenceDataPtr inputSequence,
const StreamDescription &inputStream,
const StreamDescription &outputStream);
SequenceDataPtr TypedTransform(SequenceDataPtr inputSequence);
std::vector<StreamDescriptionPtr> m_outputStreams;
std::vector<StreamId> m_appliedStreamIds;
StreamDescription m_inputStream;
StreamDescription m_outputStream;
};
// Intensity jittering based on PCA transform as described in original AlexNet paper
@ -176,11 +153,9 @@ private:
class IntensityTransformer : public ImageTransformerBase
{
public:
void Initialize(TransformerPtr next, const ConfigParameters &readerConfig) override;
explicit IntensityTransformer(const ConfigParameters& config);
private:
void InitFromConfig(const ConfigParameters &config);
void StartEpoch(const EpochConfiguration &config) override;
void Apply(size_t id, cv::Mat &mat) override;
@ -201,11 +176,9 @@ private:
class ColorTransformer : public ImageTransformerBase
{
public:
void Initialize(TransformerPtr next, const ConfigParameters &readerConfig) override;
explicit ColorTransformer(const ConfigParameters& config);
private:
void InitFromConfig(const ConfigParameters &config);
void StartEpoch(const EpochConfiguration &config) override;
void Apply(size_t id, cv::Mat &mat) override;

7
Source/Readers/ReaderLib/BlockRandomizer.h
Просмотреть файл

@ -7,7 +7,7 @@
#include <vector>
#include "Transformer.h"
#include "SequenceEnumerator.h"
#include "DataDeserializer.h"
#include "ChunkRandomizer.h"
#include "SequenceRandomizer.h"
@ -31,8 +31,7 @@ namespace Microsoft { namespace MSR { namespace CNTK {
// This class is responsible for decimation and loading the data chunks in to memory.
// Actual randomization happens in ChunkRandomizer and SequenceRandomizer.
// TODO: The behavior can be simplified by only randomizing sequences forward.
// TODO: The layering will be changed, when we move transformers under the randomizer, it won't be a transformer anymore.
class BlockRandomizer : public Transformer
class BlockRandomizer : public SequenceEnumerator
{
public:
// Currently, decimation based on sequences or chunks is supported.
@ -50,8 +49,6 @@ public:
bool useLegacyRandomization = false,
bool multithreadedGetNextSequences = false);
virtual void Initialize(TransformerPtr, const ConfigParameters&) override {};
// Starts a new epoch.
virtual void StartEpoch(const EpochConfiguration& config) override;

23
Source/Readers/ReaderLib/Bundler.cpp
Просмотреть файл

@ -133,13 +133,18 @@ void Bundler::GetSequencesForChunk(size_t chunkId, std::vector<SequenceDescripti
// Can return because all sequences are clean.
if (chunk->m_invalid.empty())
{
// Reindexing, because currently m_ids provided by some deserializers (i.e. CNTKTextFormat) are not contiguous.
for (size_t i = 0; i < sequences.size(); ++i)
{
sequences[i].m_id = i;
}
return;
}
// Do cleansing.
std::vector<SequenceDescription> result;
result.reserve(sequences.size());
for (size_t sequenceIndex = 0; sequenceIndex < sequences.size(); ++sequenceIndex)
for (size_t sequenceIndex = 0, index = 0; sequenceIndex < sequences.size(); ++sequenceIndex)
{
if (chunk->m_invalid.find(sequenceIndex) != chunk->m_invalid.end())
{
@ -147,6 +152,7 @@ void Bundler::GetSequencesForChunk(size_t chunkId, std::vector<SequenceDescripti
}
result.push_back(sequences[sequenceIndex]);
result.back().m_id = index++;
}
std::swap(sequences, result);
}
@ -159,8 +165,8 @@ class Bundler::BundlingChunk : public Chunk
size_t m_chunkId;
// A mapping between exposed sequence id and inner chunk for each deserializer.
// Index i of the vector maps to the chunk of inner sequence (i / m_numberOfInputs) of
// deserializer (i % m_numberOfInputs).
// Index i of the vector maps to the chunk of inner sequence (i / number of deserializers) of
// deserializer (i % number of deserializers).
std::vector<ChunkPtr> m_innerChunks;
// A mapping between exposed sequence id and inner sequence id for each deserializer.
// Indices as above.
@ -176,15 +182,14 @@ public:
ChunkDescriptionPtr original = chunk->m_original;
auto& deserializers = m_parent->m_deserializers;
assert(numberOfInputs == deserializers.size());
std::vector<SequenceDescription> sequences;
sequences.reserve(original->m_numberOfSequences);
// Creating chunk mapping.
m_parent->m_driver->GetSequencesForChunk(original->m_id, sequences);
ChunkPtr drivingChunk = m_parent->m_driver->GetChunk(original->m_id);
m_sequenceToSequence.resize(m_numberOfInputs * sequences.size());
m_innerChunks.resize(m_numberOfInputs * sequences.size());
m_sequenceToSequence.resize(deserializers.size() * sequences.size());
m_innerChunks.resize(deserializers.size() * sequences.size());
for (size_t sequenceIndex = 0; sequenceIndex < sequences.size(); ++sequenceIndex)
{
if (chunk->m_invalid.find(sequenceIndex) != chunk->m_invalid.end())
@ -192,7 +197,7 @@ public:
continue;
}
size_t currentIndex = sequenceIndex * m_numberOfInputs;
size_t currentIndex = sequenceIndex * deserializers.size();
m_sequenceToSequence[currentIndex] = sequences[sequenceIndex].m_id;
m_innerChunks[currentIndex] = drivingChunk;
}
@ -210,7 +215,7 @@ public:
continue;
}
size_t currentIndex = sequenceIndex * m_numberOfInputs + deserializerIndex;
size_t currentIndex = sequenceIndex * deserializers.size() + deserializerIndex;
deserializers[deserializerIndex]->GetSequenceDescriptionByKey(sequences[sequenceIndex].m_key, s);
m_sequenceToSequence[currentIndex] = s.m_id;
@ -235,7 +240,7 @@ public:
virtual void GetSequence(size_t sequenceId, std::vector<SequenceDataPtr>& result) override
{
result.reserve(m_numberOfInputs);
size_t currentIndex = sequenceId * m_numberOfInputs;
size_t currentIndex = sequenceId * m_parent->m_deserializers.size();
for (int i = 0; i < m_parent->m_deserializers.size(); ++i)
{
size_t originalSequenceId = m_sequenceToSequence[currentIndex + i];

26
Source/Readers/ReaderLib/ChunkCache.cpp Normal file
Просмотреть файл

@ -0,0 +1,26 @@
//
// Copyright (c) Microsoft. All rights reserved.
// Licensed under the MIT license. See LICENSE.md file in the project root for full license information.
//
#define _CRT_SECURE_NO_WARNINGS
#include "ChunkCache.h"
namespace Microsoft { namespace MSR { namespace CNTK {
ChunkPtr ChunkCache::GetChunk(size_t chunkId)
{
auto it = m_chunkMap.find(chunkId);
if (it != m_chunkMap.end())
{
return it->second;
}
ChunkPtr chunk = m_deserializer->GetChunk(chunkId);
m_chunkMap[chunkId] = chunk;
return chunk;
}
} } }

56
Source/Readers/ReaderLib/ChunkCache.h Normal file
Просмотреть файл

@ -0,0 +1,56 @@
//
// Copyright (c) Microsoft. All rights reserved.
// Licensed under the MIT license. See LICENSE.md file in the project root for full license information.
//
#pragma once
#include <map>
#include "DataDeserializer.h"
namespace Microsoft { namespace MSR { namespace CNTK {
// A cache to store the complete dataset (all chunks) in memory. The caching can
// be switched on/off by a boolean flag in the reader config section, independent
// of the randomization and chunking parameters. The caching should only be enabled
// when the whole dataset fits in memory.
// Implemented as a wrapping proxy around a deserializer that stores pointers to
// all chunks it sees in an internal map.
class ChunkCache : public IDataDeserializer
{
public:
ChunkCache(IDataDeserializerPtr deserializer) : m_deserializer(deserializer) { }
virtual std::vector<StreamDescriptionPtr> GetStreamDescriptions() const override
{
return m_deserializer->GetStreamDescriptions();
}
virtual ChunkDescriptions GetChunkDescriptions() override
{
return m_deserializer->GetChunkDescriptions();
}
virtual void GetSequencesForChunk(size_t chunkId, std::vector<SequenceDescription>& descriptions) override
{
return m_deserializer->GetSequencesForChunk(chunkId, descriptions);
}
virtual void GetSequenceDescriptionByKey(const KeyType& key, SequenceDescription& description) override
{
return m_deserializer->GetSequenceDescriptionByKey(key, description);
}
// Gets chunk data given its id.
virtual ChunkPtr GetChunk(size_t chunkId);
private:
// A map of currently loaded chunks
std::map<size_t, ChunkPtr> m_chunkMap;
IDataDeserializerPtr m_deserializer;
DISABLE_COPY_AND_MOVE(ChunkCache);
};
} } }

40
Source/Readers/ReaderLib/ConfigUtil.h Normal file
Просмотреть файл

@ -0,0 +1,40 @@
//
// Copyright (c) Microsoft. All rights reserved.
// Licensed under the MIT license. See LICENSE.md file in the project root for full license information.
//
#pragma once
#include <string>
#include <vector>
#include "Config.h"
namespace Microsoft { namespace MSR { namespace CNTK {
// Helper function to get sections that contains specified parameter.
inline std::vector<std::string> TryGetSectionsWithParameter(const ConfigParameters& config, const std::string& parameterName)
{
std::vector<std::string> sectionNames;
for (const std::pair<std::string, ConfigParameters>& section : config)
{
if (section.second.ExistsCurrent(parameterName))
{
sectionNames.push_back(section.first);
}
}
return sectionNames;
}
// Helper function to get sections that contains specified parameter. Throws if the parameter does not exist.
inline std::vector<std::string> GetSectionsWithParameter(const std::string& reader, const ConfigParameters& config, const std::string& parameterName)
{
auto result = TryGetSectionsWithParameter(config, parameterName);
if (result.empty())
{
RuntimeError("%s requires %s parameter.", reader.c_str(), parameterName.c_str());
}
return result;
}
}}}

8
Source/Readers/ReaderLib/FramePacker.h
Просмотреть файл

@ -15,12 +15,10 @@ class FramePacker : public SequencePacker
public:
FramePacker(
MemoryProviderPtr memoryProvider,
TransformerPtr transformer,
SequenceEnumeratorPtr sequenceEnumerator,
const std::vector<StreamDescriptionPtr>& streams) :
SequencePacker(memoryProvider, transformer, streams)
{
}
SequencePacker(memoryProvider, sequenceEnumerator, streams)
{}
private:

4
Source/Readers/ReaderLib/NoRandomizer.cpp
Просмотреть файл

@ -42,10 +42,6 @@ NoRandomizer::NoRandomizer(IDataDeserializerPtr deserializer, bool multithreaded
m_totalNumberOfSamples = sampleCount;
}
void NoRandomizer::Initialize(TransformerPtr, const ConfigParameters&)
{
}
size_t NoRandomizer::GetChunkIndexOf(size_t samplePosition)
{
auto result = std::upper_bound(m_chunkSampleOffset.begin(), m_chunkSampleOffset.end(), samplePosition);

8
Source/Readers/ReaderLib/NoRandomizer.h
Просмотреть файл

@ -6,10 +6,8 @@
#pragma once
#include <vector>
#include <map>
#include "Transformer.h"
#include "SequenceEnumerator.h"
#include "DataDeserializer.h"
#include "SequenceRandomizer.h"
namespace Microsoft { namespace MSR { namespace CNTK {
@ -17,13 +15,11 @@ namespace Microsoft { namespace MSR { namespace CNTK {
// Used training where the training data has already been pre - randomized.
// TODO: currently this code moved from the old block randomizer.
// TODO: The class will be further refactored and common based will be extracted with BlockRandomizer.
// TODO: This layering will be changed, when we move transformers under the randomizer, it won't be a transformer anymore.
class NoRandomizer : public Transformer
class NoRandomizer : public SequenceEnumerator
{
public:
NoRandomizer(IDataDeserializerPtr deserializer, bool multithreadedGetNextSequences = false);
virtual void Initialize(TransformerPtr next, const ConfigParameters& readerConfig) override;
virtual void StartEpoch(const EpochConfiguration& config) override;
virtual Sequences GetNextSequences(size_t sampleCount) override;
virtual std::vector<StreamDescriptionPtr> GetStreamDescriptions() const override

6
Source/Readers/ReaderLib/PackerBase.cpp
Просмотреть файл

@ -34,13 +34,13 @@ void PackerBase::StartEpoch(const EpochConfiguration& config)
}
PackerBase::PackerBase(MemoryProviderPtr memoryProvider,
TransformerPtr transformer,
SequenceEnumeratorPtr sequenceEnumerator,
const std::vector<StreamDescriptionPtr>& streams) :
m_transformer(transformer),
m_sequenceEnumerator(sequenceEnumerator),
m_minibatchSize(0),
m_outputStreamDescriptions(streams)
{
m_inputStreamDescriptions = m_transformer->GetStreamDescriptions();
m_inputStreamDescriptions = sequenceEnumerator->GetStreamDescriptions();
assert(m_inputStreamDescriptions.size() != 0);
assert(m_inputStreamDescriptions.size() == m_outputStreamDescriptions.size());

7
Source/Readers/ReaderLib/PackerBase.h
Просмотреть файл

@ -7,9 +7,8 @@
#include "Reader.h"
#include "MemoryProvider.h"
#include "Transformer.h"
#include "SequenceEnumerator.h"
#include "Packer.h"
#include <deque>
namespace Microsoft { namespace MSR { namespace CNTK {
@ -34,7 +33,7 @@ protected:
};
PackerBase(MemoryProviderPtr memoryProvider,
TransformerPtr transformer,
SequenceEnumeratorPtr sequenceEnumerator,
const std::vector<StreamDescriptionPtr>& streams);
typedef std::vector<SequenceDataPtr> StreamBatch;
@ -57,7 +56,7 @@ protected:
// (sampleOffset is equal to the sum of sample sizes of all preceding samples).
void PackDenseSample(char* destination, SequenceDataPtr sequence, size_t sampleOffset, size_t sampleSize);
TransformerPtr m_transformer;
SequenceEnumeratorPtr m_sequenceEnumerator;
// Input stream descriptions provided by the transformer.
std::vector<StreamDescriptionPtr> m_outputStreamDescriptions;

8
Source/Readers/ReaderLib/ReaderLib.vcxproj
Просмотреть файл

@ -40,17 +40,20 @@
</ClCompile>
</ItemDefinitionGroup>
<ItemGroup>
<ClInclude Include="ConfigUtil.h" />
<ClInclude Include="CorpusDescriptor.h" />
<ClInclude Include="Bundler.h" />
<ClInclude Include="ChunkCache.h" />
<ClInclude Include="ChunkRandomizer.h" />
<ClInclude Include="TransformController.h" />
<ClInclude Include="DataDeserializerBase.h" />
<ClInclude Include="BlockRandomizer.h" />
<ClInclude Include="Packer.h" />
<ClInclude Include="PackerBase.h" />
<ClInclude Include="SequenceEnumerator.h" />
<ClInclude Include="SequencePacker.h" />
<ClInclude Include="SequenceRandomizer.h" />
<ClInclude Include="StringToIdMap.h" />
<ClInclude Include="TransformerBase.h" />
<ClInclude Include="NoRandomizer.h" />
<ClInclude Include="CudaMemoryProvider.h" />
<ClInclude Include="DataDeserializer.h" />
@ -65,6 +68,7 @@
</ItemGroup>
<ItemGroup>
<ClCompile Include="Bundler.cpp" />
<ClCompile Include="ChunkCache.cpp" />
<ClCompile Include="ChunkRandomizer.cpp" />
<ClCompile Include="NoRandomizer.cpp" />
<ClCompile Include="BlockRandomizer.cpp" />
@ -78,4 +82,4 @@
<Import Project="$(VCTargetsPath)\Microsoft.Cpp.targets" />
<ImportGroup Label="ExtensionTargets">
</ImportGroup>
</Project>
</Project>

20
Source/Readers/ReaderLib/ReaderLib.vcxproj.filters
Просмотреть файл

@ -28,9 +28,6 @@
<ClInclude Include="ElementTypeUtils.h">
<Filter>Utils</Filter>
</ClInclude>
<ClInclude Include="TransformerBase.h">
<Filter>Transformers</Filter>
</ClInclude>
<ClInclude Include="DataDeserializerBase.h">
<Filter>Deserializers</Filter>
</ClInclude>
@ -64,9 +61,21 @@
<ClInclude Include="TruncatedBpttPacker.h">
<Filter>Packers</Filter>
</ClInclude>
<ClInclude Include="..\ExperimentalHTKMLFReader\CorpusDescriptor.h">
<ClInclude Include="SequenceEnumerator.h">
<Filter>Interfaces</Filter>
</ClInclude>
<ClInclude Include="ConfigUtil.h">
<Filter>Utils</Filter>
</ClInclude>
<ClInclude Include="ChunkCache.h">
<Filter>Utils</Filter>
</ClInclude>
<ClInclude Include="CorpusDescriptor.h">
<Filter>Interfaces</Filter>
</ClInclude>
<ClInclude Include="TransformController.h">
<Filter>Transformers</Filter>
</ClInclude>
</ItemGroup>
<ItemGroup>
<ClCompile Include="NoRandomizer.cpp">
@ -99,6 +108,9 @@
<ClCompile Include="TruncatedBpttPacker.cpp">
<Filter>Packers</Filter>
</ClCompile>
<ClCompile Include="ChunkCache.cpp">
<Filter>Utils</Filter>
</ClCompile>
</ItemGroup>
<ItemGroup>
<Filter Include="Interfaces">

7
Source/Readers/ReaderLib/ReaderShim.h
Просмотреть файл

@ -33,6 +33,13 @@ public:
virtual void Destroy() override
{
// Make sure there are no outstanding reads.
if (m_prefetchTask.valid())
{
// If there are some, give them time to finish.
m_prefetchTask.wait_for(std::chrono::seconds(5));
}
delete this;
}

52
Source/Readers/ReaderLib/SequenceEnumerator.h Normal file
Просмотреть файл

@ -0,0 +1,52 @@
//
// Copyright (c) Microsoft. All rights reserved.
// Licensed under the MIT license. See LICENSE.md file in the project root for full license information.
//
#pragma once
#include <vector>
#include "DataDeserializer.h"
namespace Microsoft { namespace MSR { namespace CNTK {
class ConfigParameters;
// Defines a set of sequences for a set of streams.
// Return by the sequence enumerator.
struct Sequences
{
// Data for up to a requested number of sequences.
// Indices in the outer vector have to correspond to the stream ids returned from the GetStreamDescriptions().
std::vector<std::vector<SequenceDataPtr>> m_data;
// Indicates whether the epoch ends with the data returned.
bool m_endOfEpoch = false;
};
class SequenceEnumerator;
typedef std::shared_ptr<SequenceEnumerator> SequenceEnumeratorPtr;
// Sequence enumerator is internal interface used by the packer to get a set of new sequences.
// It is implemented either by different randomizers or by TransformController that can wrap the randomizer
// and apply different transforms on top of data.
// This interface is not exposed to the developers of deserializers/plugins, internal to CNTK.
class SequenceEnumerator
{
public:
// Describes streams the sequence enumerator produces.
virtual std::vector<StreamDescriptionPtr> GetStreamDescriptions() const = 0;
// Sets current epoch configuration.
virtual void StartEpoch(const EpochConfiguration& config) = 0;
// Gets next sequences up to a maximum count of samples.
virtual Sequences GetNextSequences(size_t sampleCount) = 0;
virtual ~SequenceEnumerator()
{
}
};
}}}

2
Source/Readers/ReaderLib/SequencePacker.cpp
Просмотреть файл

@ -38,7 +38,7 @@ MBLayoutPtr SequencePacker::CreateMBLayout(const StreamBatch& batch)
Minibatch SequencePacker::ReadMinibatch()
{
auto sequences = m_transformer->GetNextSequences(m_minibatchSize);
auto sequences = m_sequenceEnumerator->GetNextSequences(m_minibatchSize);
const auto& batch = sequences.m_data;
Minibatch minibatch(sequences.m_endOfEpoch);

4
Source/Readers/ReaderLib/SequencePacker.h
Просмотреть файл

@ -16,9 +16,9 @@ class SequencePacker : public PackerBase
public:
SequencePacker(
MemoryProviderPtr memoryProvider,
TransformerPtr transformer,
SequenceEnumeratorPtr sequenceEnumerator,
const std::vector<StreamDescriptionPtr>& streams) :
PackerBase(memoryProvider, transformer, streams)
PackerBase(memoryProvider, sequenceEnumerator, streams)
{
}

2
Source/Readers/ReaderLib/StringToIdMap.h
Просмотреть файл

@ -33,7 +33,7 @@ public:
}
// Tries to get a value by id.
bool TryGet(const TString& value, size_t& id)
bool TryGet(const TString& value, size_t& id) const
{
const auto& it = m_values.find(value);
if (it == m_values.end())

105
Source/Readers/ReaderLib/TransformController.h Normal file
Просмотреть файл

@ -0,0 +1,105 @@
//
// Copyright (c) Microsoft. All rights reserved.
// Licensed under the MIT license. See LICENSE.md file in the project root for full license information.
//
#pragma once
#include <set>
#include "Transformer.h"
#include "SequenceEnumerator.h"
namespace Microsoft { namespace MSR { namespace CNTK {
// A pair of a transformer and the stream name to which the transformer should be a applied.
struct Transformation
{
TransformerPtr m_transformer;
std::wstring m_streamName;
};
// A class responsible for applying a list of transformers to sequences and stream descriptions.
// Delegates retrieving of sequences to another sequence provider(such as randomizer) and applies transformations after retrieving.
// Usually used by the packer to get next set of sequences.
class TransformController : public SequenceEnumerator
{
public:
TransformController(const std::vector<Transformation>& transformations, SequenceEnumeratorPtr sequenceProvider)
: m_sequenceProvider(sequenceProvider)
{
// Applying transformations to stream descriptions,
// i.e. a transformation can change a stream from dense to sparse.
std::vector<StreamDescriptionPtr> transformedStreams = m_sequenceProvider->GetStreamDescriptions();
for (auto& t : transformations)
{
size_t streamId = GetStreamId(t.m_streamName, transformedStreams);
m_transformations.push_back(std::make_pair(t, streamId));
transformedStreams[streamId] = std::make_shared<StreamDescription>(t.m_transformer->Transform(*transformedStreams[streamId]));
}
m_outputStreams = transformedStreams;
}
// Sets configuration for the current epoch.
// Some transformers can change their config based on the epoch.
virtual void StartEpoch(const EpochConfiguration &config) override
{
assert(m_sequenceProvider != nullptr);
for (auto& t : m_transformations)
{
t.first.m_transformer->StartEpoch(config);
}
m_sequenceProvider->StartEpoch(config);
}
// Description of streams that the transformer provides.
virtual std::vector<StreamDescriptionPtr> GetStreamDescriptions() const override
{
return m_outputStreams;
}
// Gets next sequences up to a maximum count of samples,
// applying transformers to particular streams.
virtual Sequences GetNextSequences(size_t sampleCount) override
{
assert(m_sequenceProvider != nullptr);
Sequences sequences = m_sequenceProvider->GetNextSequences(sampleCount);
if (sequences.m_data.empty())
{
return sequences;
}
#pragma omp parallel for schedule(dynamic)
for (int j = 0; j < sequences.m_data.front().size(); ++j)
{
for (auto& t : m_transformations)
{
sequences.m_data[t.second][j] = t.first.m_transformer->Transform(sequences.m_data[t.second][j]);
}
}
return sequences;
}
private:
size_t GetStreamId(const std::wstring streamName, const std::vector<StreamDescriptionPtr>& streams) const
{
for (const auto& s : streams)
{
if (s->m_name == streamName)
{
return s->m_id;
}
}
assert(false);
LogicError("Unexpected stream specified for transformation.");
}
SequenceEnumeratorPtr m_sequenceProvider;
std::vector<StreamDescriptionPtr> m_outputStreams;
std::vector<std::pair<Transformation, size_t>> m_transformations;
};
}}}

44
Source/Readers/ReaderLib/Transformer.h
Просмотреть файл

@ -5,57 +5,33 @@
#pragma once
#include <vector>
#include "DataDeserializer.h"
namespace Microsoft { namespace MSR { namespace CNTK {
class ConfigParameters;
// Defines a set of sequences.
struct Sequences
{
Sequences()
: m_endOfEpoch(false)
{
}
// Data for up to a requested number of sequences.
// Indices in the inner vector have to correspond to the stream IDs
// given by GetStream().
std::vector<std::vector<SequenceDataPtr>> m_data;
// Indicates whether the epoch ends with the data returned.
bool m_endOfEpoch;
};
class Transformer;
typedef std::shared_ptr<Transformer> TransformerPtr;
// Defines a data transformation interface.
// Transformers are responsible for doing custom transformation of sequences.
// For example for images, there could be scale, crop, or median transformation.
// TODO: Adopt to the C#/Java iterator pattern.
class Transformer
{
public:
// Initialization.
virtual void Initialize(
TransformerPtr next,
const ConfigParameters& readerConfig) = 0;
// Starts a new epoch. Some transformers have to change their configuration
// based on the epoch.
virtual void StartEpoch(const EpochConfiguration &config) = 0;
// Describes streams the transformer produces.
virtual std::vector<StreamDescriptionPtr> GetStreamDescriptions() const = 0;
// Transformers are applied on a particular input stream - this method should describe
// how inputStream is transformed to the output stream (return value)
virtual StreamDescription Transform(const StreamDescription& inputStream) = 0;
// Sets current epoch configuration.
virtual void StartEpoch(const EpochConfiguration& config) = 0;
// Gets next sequences up to a maximum count of samples.
// The return value can be used until the next call to GetNextSequences.
virtual Sequences GetNextSequences(size_t sampleCount) = 0;
// This method should describe how input sequences is transformed to the output sequence.
virtual SequenceDataPtr Transform(SequenceDataPtr inputSequence) = 0;
virtual ~Transformer()
{
}
};
} } }
}}}

91
Source/Readers/ReaderLib/TransformerBase.h
Просмотреть файл

@ -1,91 +0,0 @@
//
// Copyright (c) Microsoft. All rights reserved.
// Licensed under the MIT license. See LICENSE.md file in the project root for full license information.
//
#pragma once
#include <set>
#include "Transformer.h"
namespace Microsoft { namespace MSR { namespace CNTK {
class TransformerBase : public Transformer
{
public:
// Initializes the transformer.
virtual void Initialize(TransformerPtr next,
const ConfigParameters &) override
{
m_next = next;
m_inputStreams = m_next->GetStreamDescriptions();
}
// Sets configuration for the current epoch.
virtual void StartEpoch(const EpochConfiguration &config) override
{
assert(m_next != nullptr);
m_next->StartEpoch(config);
}
// Description of streams that the transformer provides.
virtual std::vector<StreamDescriptionPtr> GetStreamDescriptions() const override
{
return this->GetOutputStreams();
}
// Gets next sequences up to a maximum count of samples.
// Sequences contains data for all streams.
virtual Sequences GetNextSequences(size_t sampleCount) override
{
assert(m_next != nullptr);
Sequences samples = m_next->GetNextSequences(sampleCount);
if (samples.m_data.empty())
{
return samples;
}
const auto &appliedStreamIds = GetAppliedStreamIds();
const auto &outputStreams = GetOutputStreams();
// TODO: Move parallelization on the outer loop with collapse.
for (int j = 0; j < appliedStreamIds.size(); ++j)
{
size_t streamId = appliedStreamIds[j];
auto& allSamples = samples.m_data[streamId];
#pragma omp parallel for schedule(dynamic)
for (int i = 0; i < allSamples.size(); ++i)
{
allSamples[i] = Apply(allSamples[i], *m_inputStreams[streamId], *outputStreams[streamId]);
}
}
return samples;
}
protected:
virtual const std::vector<StreamId> &GetAppliedStreamIds() const = 0;
virtual const std::vector<StreamDescriptionPtr> &GetOutputStreams() const
{
return m_inputStreams;
}
const std::vector<StreamDescriptionPtr> &GetInputStreams()
{
return m_inputStreams;
}
private:
// Applies transformation to the sequence.
virtual SequenceDataPtr Apply(SequenceDataPtr inputSequence,
const StreamDescription &inputStream,
const StreamDescription &outputStream) = 0;
TransformerPtr m_next;
std::vector<StreamId> m_featureStreamIds;
std::vector<StreamDescriptionPtr> m_inputStreams;
};
}}}

15
Source/Readers/ReaderLib/TruncatedBpttPacker.cpp
Просмотреть файл

@ -107,9 +107,9 @@ struct SequenceBuffer
TruncatedBPTTPacker::TruncatedBPTTPacker(
MemoryProviderPtr memoryProvider,
TransformerPtr transformer,
SequenceEnumeratorPtr sequenceEnumerator,
const vector<StreamDescriptionPtr>& streams)
: PackerBase(memoryProvider, transformer, streams),
: PackerBase(memoryProvider, sequenceEnumerator, streams),
m_truncationSize(0)
{
auto sparseOutput = find_if(m_outputStreamDescriptions.begin(), m_outputStreamDescriptions.end(), [](const StreamDescriptionPtr& s){ return s->m_storageType == StorageType::sparse_csc; });
@ -148,6 +148,15 @@ void TruncatedBPTTPacker::StartEpoch(const EpochConfiguration& config)
// Estimating the number of parallel sequences to pack (slots) from the minibatch size and truncation size.
m_numParallelSequences = max(1, (int)floor(m_minibatchSize / m_truncationSize));
if (config.m_numberOfWorkers > m_numParallelSequences)
{
InvalidArgument("Too many workers for minibatch size; please increase minibatch size or decrease number of workers.");
}
m_numParallelSequences =
(m_numParallelSequences / config.m_numberOfWorkers) +
(config.m_workerRank < (m_numParallelSequences % config.m_numberOfWorkers) ? 1 : 0);
m_sequenceBufferPerStream.clear();
// Preparing the buffers.
@ -303,7 +312,7 @@ void TruncatedBPTTPacker::ReadSequencesToSlot(size_t slotIndex)
{
// We need a single sequence, potentially we can request (m_truncationSize - slot.AvailableNumberOfSamples())
// to be more efficient. In reality the truncation size usually is less the sequence size.
auto s = m_transformer->GetNextSequences(1);
auto s = m_sequenceEnumerator->GetNextSequences(1);
if (s.m_endOfEpoch)
{
break;

3
Source/Readers/ReaderLib/TruncatedBpttPacker.h
Просмотреть файл

@ -7,7 +7,6 @@
#include "Reader.h"
#include "MemoryProvider.h"
#include "Transformer.h"
#include "PackerBase.h"
namespace Microsoft { namespace MSR { namespace CNTK {
@ -23,7 +22,7 @@ class TruncatedBPTTPacker : public PackerBase
public:
TruncatedBPTTPacker(
MemoryProviderPtr memoryProvider,
TransformerPtr transformer,
SequenceEnumeratorPtr sequenceEnumerator,
const std::vector<StreamDescriptionPtr>& streams);
virtual Minibatch ReadMinibatch() override;

18
Source/SGDLib/SGD.cpp
Просмотреть файл

@ -365,7 +365,7 @@ void SGD<ElemType>::TrainOrAdaptModel(int startEpoch, ComputationNetworkPtr net,
// set dropout rate for this epoch
// We use the same seed across workers until parallel training kicks in to ensure that the workers have identical models
size_t parallelWorkerIdx = ((m_mpi == nullptr) || !UseParallelTrain(i)) ? 0 : m_mpi->CurrentNodeRank();
size_t parallelWorkerIdx = ((m_mpi == nullptr) || !UsingParallelTrain(i)) ? 0 : m_mpi->CurrentNodeRank();
size_t dropoutRandSeedBase = (parallelWorkerIdx * m_maxEpochs) + i;
ComputationNetwork::SetDropoutRate<ElemType>(net, criterionNodes[0], m_dropoutRates[i], prevDropoutRate, dropoutRandSeedBase);
ComputationNetwork::SetBatchNormalizationTimeConstants<ElemType>(net, criterionNodes[0],
@ -771,9 +771,9 @@ size_t SGD<ElemType>::TrainOneEpoch(ComputationNetworkPtr net,
int numMBsRun = 0;
bool useGradientAggregation = UseGradientAggregation(epochNumber);
bool useModelAggregation = UseModelAggregation(epochNumber);
bool useParallelTrain = UseParallelTrain(epochNumber);
bool useGradientAggregation = UsingGradientAggregation(epochNumber);
bool useModelAggregation = UsingModelAggregation(epochNumber);
bool useParallelTrain = UsingParallelTrain(epochNumber);
// MA-related variables
size_t nSamplesSinceLastModelSync = 0;
@ -1024,7 +1024,7 @@ size_t SGD<ElemType>::TrainOneEpoch(ComputationNetworkPtr net,
for (size_t i = 0; i < evaluationNodes.size(); i++)
m_gradHeader->evalErrors[i] = localEpochEvalErrors.GetCriterion(i);
bool samplesProcessed = m_distGradAgg->AggregateGradients(learnParamsGradients, m_gradHeader, epochNumber);
bool samplesProcessed = m_distGradAgg->AggregateGradients(learnParamsGradients, m_gradHeader.get(), epochNumber);
noMoreSamplesToProcess = !samplesProcessed;
aggregateNumSamples = m_gradHeader->numSamples;
@ -1796,20 +1796,22 @@ void SGD<ElemType>::InitDistGradAgg(int numEvalNodes, int traceLevel)
if (m_distGradAgg == nullptr)
{
#ifdef CNTK_PARALLEL_TRAINING_SUPPORT
m_distGradAgg = new AllReduceDistGradAggregator<ElemType>(m_mpi, m_numGradientBits, m_zeroThresholdFor1Bit, true /*useQuantizationForSelfStripe*/, m_bufferedAsyncGradientAggregation, traceLevel, m_syncStatsTrace);
m_distGradAgg = std::make_shared<AllReduceDistGradAggregator<ElemType>>(m_mpi, m_numGradientBits, m_zeroThresholdFor1Bit, true /*useQuantizationForSelfStripe*/, m_bufferedAsyncGradientAggregation, traceLevel, m_syncStatsTrace);
#else
if (m_numGradientBits != (8 * sizeof(ElemType)))
{
RuntimeError("Gradient quantization is unsupported in CNTK binaries built without quantized gradient aggregation support!");
}
m_distGradAgg = new SimpleDistGradAggregator<ElemType>(m_mpi, m_bufferedAsyncGradientAggregation, m_syncStatsTrace);
m_distGradAgg = std::make_shared<SimpleDistGradAggregator<ElemType>>(m_mpi, m_bufferedAsyncGradientAggregation, m_syncStatsTrace);
#endif // !CNTK_PARALLEL_TRAINING_SUPPORT
}
if (m_gradHeader == nullptr)
{
m_gradHeader = DistGradHeader::Create(numEvalNodes);
m_gradHeader.reset(DistGradHeader::Create(numEvalNodes), [](DistGradHeader* ptr) {
DistGradHeader::Destroy(ptr);
});
}
}
}

12
Source/SGDLib/SGD.h
Просмотреть файл

@ -557,8 +557,8 @@ protected:
size_t m_prevChosenMinibatchSize;
double m_lastFinishedEpochTrainLoss;
IDistGradAggregator<ElemType>* m_distGradAgg;
struct DistGradHeader* m_gradHeader;
std::shared_ptr<IDistGradAggregator<ElemType>> m_distGradAgg;
std::shared_ptr<struct DistGradHeader> m_gradHeader;
shared_ptr<IMASGD<ElemType>> m_pMASGDHelper;
@ -566,21 +566,21 @@ private:
void InitializeAndCheckBlockMomentumSGDParameters();
void MarkDropoutNodesEvalTimeStampAsOutdated(const ComputationNetworkPtr& net, const ComputationNodeBasePtr& criterionNode);
bool UseGradientAggregation(size_t epochNumber)
bool UsingGradientAggregation(size_t epochNumber) const
{
return ((GetParallelizationMethod() == ParallelizationMethod::dataParallelSGD) && (epochNumber >= m_parallelizationStartEpochNum));
}
bool UseModelAggregation(size_t epochNumber)
bool UsingModelAggregation(size_t epochNumber) const
{
return ((GetParallelizationMethod() == ParallelizationMethod::modelAveragingSGD ||
GetParallelizationMethod() == ParallelizationMethod::blockMomentumSGD) &&
(epochNumber >= m_parallelizationStartEpochNum));
}
bool UseParallelTrain(size_t epochNumber)
bool UsingParallelTrain(size_t epochNumber) const
{
return UseGradientAggregation(epochNumber) || UseModelAggregation(epochNumber);
return UsingGradientAggregation(epochNumber) || UsingModelAggregation(epochNumber);
}
};

10
Source/SGDLib/SimpleEvaluator.h
Просмотреть файл

@ -177,7 +177,9 @@ public:
{
if (m_gradHeader == nullptr)
{
m_gradHeader = DistGradHeader::Create(evalNodes.size());
m_gradHeader.reset(DistGradHeader::Create(evalNodes.size()), [](DistGradHeader* ptr) {
DistGradHeader::Destroy(ptr);
});
m_distGradAgg = make_shared<SimpleDistGradAggregator<ElemType>>(m_mpi, false, m_traceLevel);
}
@ -199,7 +201,7 @@ public:
// Using SimpleDistAggregator for eval results only. At some point we should rename the class to be just
// IDistAggregator and SimpleDistAggregator.
bool samplesProcessed = m_distGradAgg->AggregateGradients(learnParamsGradients, m_gradHeader, 0);
bool samplesProcessed = m_distGradAgg->AggregateGradients(learnParamsGradients, m_gradHeader.get(), 0);
noMoreSamplesToProcess = !samplesProcessed;
aggregateNumSamplesWithLabel = m_gradHeader->numSamplesWithLabel;
@ -299,8 +301,8 @@ protected:
MPIWrapperPtr m_mpi;
bool m_enableDistributedMBReading;
shared_ptr<IDistGradAggregator<ElemType>> m_distGradAgg;
struct DistGradHeader* m_gradHeader;
std::shared_ptr<IDistGradAggregator<ElemType>> m_distGradAgg;
std::shared_ptr<struct DistGradHeader> m_gradHeader;
int m_traceLevel;
void operator=(const SimpleEvaluator&); // (not assignable)
};

13
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden/README.txt Normal file
Просмотреть файл

@ -0,0 +1,13 @@
Test runtimes
Windows:
Running test CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden (debug cpu) - [OK] 808.62 sec
Running test CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden (debug gpu) - [OK] 86.82 sec
Running test CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden (release cpu) - [OK] 49.23 sec
Running test CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden (release gpu) - [OK] 19.37 sec
Linux:
Running test CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden (debug cpu) - [OK] 170.07 sec
Running test CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden (debug gpu) - [OK] 24.82 sec
Running test CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden (release cpu) - [OK] 20.71 sec
Running test CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden (release gpu) - [OK] 8.17 sec

458
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden/baseline.linux.txt Normal file
Просмотреть файл

@ -0,0 +1,458 @@
=== Running /home/philly/jenkins/workspace/CNTK-Test-Linux-W1/build/gpu/release/bin/cntk configFile=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden/../../../../../../../Examples/Image/MNIST/Config/../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/Config/01_OneHidden.cntk currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/TestData RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/TestData ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden/../../../../../../../Examples/Image/MNIST/Config OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu DeviceId=0 timestamping=true train=[SGD=[maxEpochs=3]] imageLayout="cudnn"
-------------------------------------------------------------------
Build info:
Built time: May 13 2016 14:50:25
Last modified date: Thu May 12 14:00:37 2016
Build type: release
Build target: GPU
With 1bit-SGD: no
Math lib: acml
CUDA_PATH: /usr/local/cuda-7.5
CUB_PATH: /usr/local/cub-1.4.1
CUDNN_PATH: /usr/local/cudnn-4.0
Build Branch: HEAD
Build SHA1: 35fadc316f045d843bbd9b85061250a959268787
Built by philly on d8dc82703b0f
Build Path: /home/philly/jenkins/workspace/CNTK-Build-Linux
-------------------------------------------------------------------
Changed current directory to /tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/TestData
05/13/2016 15:10:02: -------------------------------------------------------------------
05/13/2016 15:10:02: Build info:
05/13/2016 15:10:02: Built time: May 13 2016 14:50:25
05/13/2016 15:10:02: Last modified date: Thu May 12 14:00:37 2016
05/13/2016 15:10:02: Build type: release
05/13/2016 15:10:02: Build target: GPU
05/13/2016 15:10:02: With 1bit-SGD: no
05/13/2016 15:10:02: Math lib: acml
05/13/2016 15:10:02: CUDA_PATH: /usr/local/cuda-7.5
05/13/2016 15:10:02: CUB_PATH: /usr/local/cub-1.4.1
05/13/2016 15:10:02: CUDNN_PATH: /usr/local/cudnn-4.0
05/13/2016 15:10:02: Build Branch: HEAD
05/13/2016 15:10:02: Build SHA1: 35fadc316f045d843bbd9b85061250a959268787
05/13/2016 15:10:02: Built by philly on d8dc82703b0f
05/13/2016 15:10:02: Build Path: /home/philly/jenkins/workspace/CNTK-Build-Linux
05/13/2016 15:10:02: -------------------------------------------------------------------
05/13/2016 15:10:02: Running on localhost at 2016/05/13 15:10:02
05/13/2016 15:10:02: Command line:
/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/build/gpu/release/bin/cntk configFile=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden/../../../../../../../Examples/Image/MNIST/Config/../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/Config/01_OneHidden.cntk currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/TestData RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/TestData ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden/../../../../../../../Examples/Image/MNIST/Config OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu DeviceId=0 timestamping=true train=[SGD=[maxEpochs=3]] imageLayout="cudnn"
05/13/2016 15:10:02: >>>>>>>>>>>>>>>>>>>> RAW CONFIG (VARIABLES NOT RESOLVED) >>>>>>>>>>>>>>>>>>>>
05/13/2016 15:10:02: RootDir = ".."
ConfigDir = "$RootDir$/Config"
DataDir = "$RootDir$/Data"
OutputDir = "$RootDir$/Output"
ModelDir = "$OutputDir$/Models"
deviceId = 0
imageLayout = "cudnn"
command = train:test
precision = "float"
modelPath = "$ModelDir$/01_OneHidden"
ndlMacros = "$ConfigDir$/Macros.ndl"
traceLevel=1
numMBsToShowResult=500
initOnCPUOnly=true
train = [
action = "train"
NDLNetworkBuilder = [
networkDescription = "$ConfigDir$/01_OneHidden.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.1
momentumPerMB = 0
maxEpochs = 30
]
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
test = [
action = "test"
minibatchSize = 16
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/TestData
RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu
DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/TestData
ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden/../../../../../../../Examples/Image/MNIST/Config
OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu
DeviceId=0
timestamping=true
train=[SGD=[maxEpochs=3]]
imageLayout="cudnn"
05/13/2016 15:10:02: <<<<<<<<<<<<<<<<<<<< RAW CONFIG (VARIABLES NOT RESOLVED) <<<<<<<<<<<<<<<<<<<<
05/13/2016 15:10:02: >>>>>>>>>>>>>>>>>>>> RAW CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>
05/13/2016 15:10:02: RootDir = ".."
ConfigDir = "../Config"
DataDir = "../Data"
OutputDir = "../Output"
ModelDir = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/Models"
deviceId = 0
imageLayout = "cudnn"
command = train:test
precision = "float"
modelPath = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/Models/01_OneHidden"
ndlMacros = "/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden/../../../../../../../Examples/Image/MNIST/Config/Macros.ndl"
traceLevel=1
numMBsToShowResult=500
initOnCPUOnly=true
train = [
action = "train"
NDLNetworkBuilder = [
networkDescription = "/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden/../../../../../../../Examples/Image/MNIST/Config/01_OneHidden.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.1
momentumPerMB = 0
maxEpochs = 30
]
reader = [
readerType = "CNTKTextFormatReader"
file = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/TestData/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
test = [
action = "test"
minibatchSize = 16
reader = [
readerType = "CNTKTextFormatReader"
file = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/TestData/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/TestData
RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu
DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/TestData
ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden/../../../../../../../Examples/Image/MNIST/Config
OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu
DeviceId=0
timestamping=true
train=[SGD=[maxEpochs=3]]
imageLayout="cudnn"
05/13/2016 15:10:02: <<<<<<<<<<<<<<<<<<<< RAW CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<
05/13/2016 15:10:02: >>>>>>>>>>>>>>>>>>>> PROCESSED CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>
configparameters: 01_OneHidden.cntk:command=train:test
configparameters: 01_OneHidden.cntk:ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden/../../../../../../../Examples/Image/MNIST/Config
configparameters: 01_OneHidden.cntk:currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/TestData
configparameters: 01_OneHidden.cntk:DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/TestData
configparameters: 01_OneHidden.cntk:deviceId=0
configparameters: 01_OneHidden.cntk:imageLayout=cudnn
configparameters: 01_OneHidden.cntk:initOnCPUOnly=true
configparameters: 01_OneHidden.cntk:ModelDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/Models
configparameters: 01_OneHidden.cntk:modelPath=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/Models/01_OneHidden
configparameters: 01_OneHidden.cntk:ndlMacros=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden/../../../../../../../Examples/Image/MNIST/Config/Macros.ndl
configparameters: 01_OneHidden.cntk:numMBsToShowResult=500
configparameters: 01_OneHidden.cntk:OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu
configparameters: 01_OneHidden.cntk:precision=float
configparameters: 01_OneHidden.cntk:RootDir=..
configparameters: 01_OneHidden.cntk:RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu
configparameters: 01_OneHidden.cntk:test=[
action = "test"
minibatchSize = 16
reader = [
readerType = "CNTKTextFormatReader"
file = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/TestData/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
configparameters: 01_OneHidden.cntk:timestamping=true
configparameters: 01_OneHidden.cntk:traceLevel=1
configparameters: 01_OneHidden.cntk:train=[
action = "train"
NDLNetworkBuilder = [
networkDescription = "/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden/../../../../../../../Examples/Image/MNIST/Config/01_OneHidden.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.1
momentumPerMB = 0
maxEpochs = 30
]
reader = [
readerType = "CNTKTextFormatReader"
file = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/TestData/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
] [SGD=[maxEpochs=3]]
05/13/2016 15:10:02: <<<<<<<<<<<<<<<<<<<< PROCESSED CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<
05/13/2016 15:10:02: Commands: train test
05/13/2016 15:10:02: Precision = "float"
05/13/2016 15:10:02: CNTKModelPath: /tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/Models/01_OneHidden
05/13/2016 15:10:02: CNTKCommandTrainInfo: train : 3
05/13/2016 15:10:02: CNTKCommandTrainInfo: CNTKNoMoreCommands_Total : 3
05/13/2016 15:10:02: ##############################################################################
05/13/2016 15:10:02: # #
05/13/2016 15:10:02: # Action "train" #
05/13/2016 15:10:02: # #
05/13/2016 15:10:02: ##############################################################################
05/13/2016 15:10:02: CNTKCommandTrainBegin: train
NDLBuilder Using GPU 0
05/13/2016 15:10:02: Creating virgin network.
Post-processing network...
4 roots:
ce = CrossEntropyWithSoftmax()
err = ErrorPrediction()
errTop1 = ErrorPrediction()
ol.z = Plus()
Validating network. 17 nodes to process in pass 1.
Validating --> labels = InputValue() : -> [10 x *]
Validating --> ol.W = LearnableParameter() : -> [10 x 200]
Validating --> h1.W = LearnableParameter() : -> [200 x 784]
Validating --> featScale = LearnableParameter() : -> [1 x 1]
Validating --> features = InputValue() : -> [784 x *]
Validating --> featScaled = ElementTimes (featScale, features) : [1 x 1], [784 x *] -> [784 x 1 x *]
Validating --> h1.t = Times (h1.W, featScaled) : [200 x 784], [784 x 1 x *] -> [200 x 1 x *]
Validating --> h1.b = LearnableParameter() : -> [200 x 1]
Validating --> h1.z = Plus (h1.t, h1.b) : [200 x 1 x *], [200 x 1] -> [200 x 1 x *]
Validating --> h1.y = Sigmoid (h1.z) : [200 x 1 x *] -> [200 x 1 x *]
Validating --> ol.t = Times (ol.W, h1.y) : [10 x 200], [200 x 1 x *] -> [10 x 1 x *]
Validating --> ol.b = LearnableParameter() : -> [10 x 1]
Validating --> ol.z = Plus (ol.t, ol.b) : [10 x 1 x *], [10 x 1] -> [10 x 1 x *]
Validating --> ce = CrossEntropyWithSoftmax (labels, ol.z) : [10 x *], [10 x 1 x *] -> [1]
Validating --> err = ErrorPrediction (labels, ol.z) : [10 x *], [10 x 1 x *] -> [1]
Validating --> unnamed81 = LearnableParameter() : -> [1 x 1]
Validating --> errTop1 = ErrorPrediction (labels, ol.z, unnamed81) : [10 x *], [10 x 1 x *], [1 x 1] -> [1]
Validating network. 9 nodes to process in pass 2.
Validating network, final pass.
9 out of 17 nodes do not share the minibatch layout with the input data.
Post-processing network complete.
05/13/2016 15:10:02: Created model with 17 nodes on GPU 0.
05/13/2016 15:10:02: Training criterion node(s):
05/13/2016 15:10:02: ce = CrossEntropyWithSoftmax
05/13/2016 15:10:02: Evaluation criterion node(s):
05/13/2016 15:10:02: errTop1 = ErrorPrediction
05/13/2016 15:10:02: err = ErrorPrediction
Allocating matrices for forward and/or backward propagation.
Memory Sharing Structure:
(nil): {[err Gradient[1]] [errTop1 Gradient[1]] [featScale Gradient[1 x 1]] [featScaled Gradient[784 x 1 x *]] [features Gradient[784 x *]] [labels Gradient[10 x *]] [unnamed81 Gradient[1 x 1]] }
0x1bde9d8: {[errTop1 Value[1]] }
0x1bdeb98: {[err Value[1]] }
0x1be1e38: {[features Value[784 x *]] }
0x2447ab8: {[featScale Value[1 x 1]] }
0x2448c28: {[labels Value[10 x *]] }
0x2449368: {[h1.W Value[200 x 784]] }
0x29577e8: {[h1.b Value[200 x 1]] }
0x2958938: {[ol.W Value[10 x 200]] }
0x2959808: {[ol.b Value[10 x 1]] }
0x295b198: {[unnamed81 Value[1 x 1]] }
0x295ece8: {[featScaled Value[784 x 1 x *]] }
0x295ef48: {[ol.z Value[10 x 1 x *]] }
0x295f108: {[ce Value[1]] }
0x29609d8: {[h1.t Value[200 x 1 x *]] }
0x2960d88: {[h1.W Gradient[200 x 784]] [h1.z Value[200 x 1 x *]] }
0x2960ee8: {[h1.t Gradient[200 x 1 x *]] [h1.y Value[200 x 1 x *]] }
0x2961048: {[h1.z Gradient[200 x 1 x *]] [ol.t Value[10 x 1 x *]] }
0x2961fa8: {[ce Gradient[1]] }
0x2962168: {[ol.W Gradient[10 x 200]] [ol.z Gradient[10 x 1 x *]] }
0x2962328: {[ol.t Gradient[10 x 1 x *]] }
0x29624e8: {[ol.b Gradient[10 x 1]] }
0x29626a8: {[h1.b Gradient[200 x 1]] [h1.y Gradient[200 x 1 x *]] }
05/13/2016 15:10:02: No PreCompute nodes found, skipping PreCompute step.
05/13/2016 15:10:02: Starting Epoch 1: learning rate per sample = 0.003125 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 15:10:02: Starting minibatch loop.
05/13/2016 15:10:03: Epoch[ 1 of 3]-Minibatch[ 1- 500, 26.67%]: ce = 1.30072449 * 16000; errTop1 = 0.38468750 * 16000; err = 0.38468750 * 16000; time = 1.2825s; samplesPerSecond = 12475.2
05/13/2016 15:10:04: Epoch[ 1 of 3]-Minibatch[ 501-1000, 53.33%]: ce = 0.49017273 * 16000; errTop1 = 0.13037500 * 16000; err = 0.13037500 * 16000; time = 0.2861s; samplesPerSecond = 55923.1
05/13/2016 15:10:04: Epoch[ 1 of 3]-Minibatch[1001-1500, 80.00%]: ce = 0.39744922 * 16000; errTop1 = 0.11168750 * 16000; err = 0.11168750 * 16000; time = 0.2889s; samplesPerSecond = 55389.2
05/13/2016 15:10:04: Finished Epoch[ 1 of 3]: [Training] ce = 0.65501042 * 60000; errTop1 = 0.18685000 * 60000; err = 0.18685000 * 60000; totalSamplesSeen = 60000; learningRatePerSample = 0.003125; epochTime=2.09125s
05/13/2016 15:10:04: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/Models/01_OneHidden.1'
05/13/2016 15:10:04: Starting Epoch 2: learning rate per sample = 0.003125 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 15:10:04: Starting minibatch loop.
05/13/2016 15:10:04: Epoch[ 2 of 3]-Minibatch[ 1- 500, 26.67%]: ce = 0.32870679 * 16000; errTop1 = 0.09531250 * 16000; err = 0.09531250 * 16000; time = 0.2809s; samplesPerSecond = 56955.1
05/13/2016 15:10:05: Epoch[ 2 of 3]-Minibatch[ 501-1000, 53.33%]: ce = 0.31809930 * 16000; errTop1 = 0.09206250 * 16000; err = 0.09206250 * 16000; time = 0.2862s; samplesPerSecond = 55905.9
05/13/2016 15:10:05: Epoch[ 2 of 3]-Minibatch[1001-1500, 80.00%]: ce = 0.31002502 * 16000; errTop1 = 0.08762500 * 16000; err = 0.08762500 * 16000; time = 0.2946s; samplesPerSecond = 54305.4
05/13/2016 15:10:05: Finished Epoch[ 2 of 3]: [Training] ce = 0.31494245 * 60000; errTop1 = 0.09090000 * 60000; err = 0.09090000 * 60000; totalSamplesSeen = 120000; learningRatePerSample = 0.003125; epochTime=1.08973s
05/13/2016 15:10:05: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/Models/01_OneHidden.2'
05/13/2016 15:10:06: Starting Epoch 3: learning rate per sample = 0.003125 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 15:10:06: Starting minibatch loop.
05/13/2016 15:10:06: Epoch[ 3 of 3]-Minibatch[ 1- 500, 26.67%]: ce = 0.28016867 * 16000; errTop1 = 0.08187500 * 16000; err = 0.08187500 * 16000; time = 0.2894s; samplesPerSecond = 55283.2
05/13/2016 15:10:06: Epoch[ 3 of 3]-Minibatch[ 501-1000, 53.33%]: ce = 0.28037985 * 16000; errTop1 = 0.08093750 * 16000; err = 0.08093750 * 16000; time = 0.2860s; samplesPerSecond = 55935.8
05/13/2016 15:10:06: Epoch[ 3 of 3]-Minibatch[1001-1500, 80.00%]: ce = 0.27621069 * 16000; errTop1 = 0.08237500 * 16000; err = 0.08237500 * 16000; time = 0.2791s; samplesPerSecond = 57323.8
05/13/2016 15:10:07: Finished Epoch[ 3 of 3]: [Training] ce = 0.27476087 * 60000; errTop1 = 0.08011667 * 60000; err = 0.08011667 * 60000; totalSamplesSeen = 180000; learningRatePerSample = 0.003125; epochTime=1.07334s
05/13/2016 15:10:07: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_01_OneHidden@release_gpu/Models/01_OneHidden'
05/13/2016 15:10:07: CNTKCommandTrainEnd: train
05/13/2016 15:10:07: Action "train" complete.
05/13/2016 15:10:07: ##############################################################################
05/13/2016 15:10:07: # #
05/13/2016 15:10:07: # Action "test" #
05/13/2016 15:10:07: # #
05/13/2016 15:10:07: ##############################################################################
Post-processing network...
4 roots:
ce = CrossEntropyWithSoftmax()
err = ErrorPrediction()
errTop1 = ErrorPrediction()
ol.z = Plus()
Validating network. 17 nodes to process in pass 1.
Validating --> labels = InputValue() : -> [10 x *1]
Validating --> ol.W = LearnableParameter() : -> [10 x 200]
Validating --> h1.W = LearnableParameter() : -> [200 x 784]
Validating --> featScale = LearnableParameter() : -> [1 x 1]
Validating --> features = InputValue() : -> [784 x *1]
Validating --> featScaled = ElementTimes (featScale, features) : [1 x 1], [784 x *1] -> [784 x 1 x *1]
Validating --> h1.t = Times (h1.W, featScaled) : [200 x 784], [784 x 1 x *1] -> [200 x 1 x *1]
Validating --> h1.b = LearnableParameter() : -> [200 x 1]
Validating --> h1.z = Plus (h1.t, h1.b) : [200 x 1 x *1], [200 x 1] -> [200 x 1 x *1]
Validating --> h1.y = Sigmoid (h1.z) : [200 x 1 x *1] -> [200 x 1 x *1]
Validating --> ol.t = Times (ol.W, h1.y) : [10 x 200], [200 x 1 x *1] -> [10 x 1 x *1]
Validating --> ol.b = LearnableParameter() : -> [10 x 1]
Validating --> ol.z = Plus (ol.t, ol.b) : [10 x 1 x *1], [10 x 1] -> [10 x 1 x *1]
Validating --> ce = CrossEntropyWithSoftmax (labels, ol.z) : [10 x *1], [10 x 1 x *1] -> [1]
Validating --> err = ErrorPrediction (labels, ol.z) : [10 x *1], [10 x 1 x *1] -> [1]
Validating --> unnamed81 = LearnableParameter() : -> [1 x 1]
Validating --> errTop1 = ErrorPrediction (labels, ol.z, unnamed81) : [10 x *1], [10 x 1 x *1], [1 x 1] -> [1]
Validating network. 9 nodes to process in pass 2.
Validating network, final pass.
9 out of 17 nodes do not share the minibatch layout with the input data.
Post-processing network complete.
evalNodeNames are not specified, using all the default evalnodes and training criterion nodes.
Allocating matrices for forward and/or backward propagation.
Memory Sharing Structure:
(nil): {[ce Gradient[1]] [err Gradient[1]] [errTop1 Gradient[1]] [featScale Gradient[1 x 1]] [featScaled Gradient[784 x 1 x *1]] [features Gradient[784 x *1]] [h1.W Gradient[200 x 784]] [h1.b Gradient[200 x 1]] [h1.t Gradient[200 x 1 x *1]] [h1.y Gradient[200 x 1 x *1]] [h1.z Gradient[200 x 1 x *1]] [labels Gradient[10 x *1]] [ol.W Gradient[10 x 200]] [ol.b Gradient[10 x 1]] [ol.t Gradient[10 x 1 x *1]] [ol.z Gradient[10 x 1 x *1]] [unnamed81 Gradient[1 x 1]] }
0x7f0d2f269e18: {[labels Value[10 x *1]] }
0x7f0d2f26a4c8: {[ol.b Value[10 x 1]] }
0x7f0d2f26b578: {[ol.W Value[10 x 200]] }
0x7f0d2f26bd18: {[unnamed81 Value[1 x 1]] }
0x7f0d2f270658: {[errTop1 Value[1]] }
0x7f0d2f270818: {[err Value[1]] }
0x7f0d2f2709d8: {[ce Value[1]] }
0x7f0d2f270f28: {[h1.t Value[200 x 1 x *1]] }
0x7f0d2f2720b8: {[featScaled Value[784 x 1 x *1]] }
0x7f0d2f272588: {[h1.z Value[200 x 1 x *1]] }
0x7f0d2f272748: {[h1.y Value[200 x 1 x *1]] }
0x7f0d2f272908: {[ol.t Value[10 x 1 x *1]] }
0x7f0d2f272ac8: {[ol.z Value[10 x 1 x *1]] }
0x7f0d35693b68: {[featScale Value[1 x 1]] }
0x7f0d4bd02258: {[h1.b Value[200 x 1]] }
0x7f0d4bd02f98: {[features Value[784 x *1]] }
0x7f0d4bd03c78: {[h1.W Value[200 x 784]] }
05/13/2016 15:10:10: Final Results: Minibatch[1-625]: errTop1 = 0.07140000 * 10000; err = 0.07140000 * 10000; ce = 0.25287636 * 10000; perplexity = 1.28772405
05/13/2016 15:10:10: Action "test" complete.
05/13/2016 15:10:10: __COMPLETED__

456
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden/baseline.windows.txt Normal file
Просмотреть файл

@ -0,0 +1,456 @@
=== Running /cygdrive/c/jenkins/workspace/CNTK-Test-Windows-W1/x64/release/cntk.exe configFile=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/Config/01_OneHidden.cntk currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu\TestData RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu\TestData ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu DeviceId=0 timestamping=true train=[SGD=[maxEpochs=3]] imageLayout="cudnn"
-------------------------------------------------------------------
Build info:
Built time: May 13 2016 08:06:01
Last modified date: Thu May 12 07:31:50 2016
Build type: Release
Build target: GPU
With 1bit-SGD: no
CUDA_PATH: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5
CUB_PATH: c:\src\cub-1.4.1
CUDNN_PATH: c:\NVIDIA\cudnn-4.0\cuda
Build Branch: HEAD
Build SHA1: 35fadc316f045d843bbd9b85061250a959268787
Built by svcphil on Philly-Pool3
Build Path: c:\Jenkins\workspace\CNTK-Build-Windows\Source\CNTK\
-------------------------------------------------------------------
Changed current directory to C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu\TestData
05/13/2016 08:15:51: -------------------------------------------------------------------
05/13/2016 08:15:51: Build info:
05/13/2016 08:15:51: Built time: May 13 2016 08:06:01
05/13/2016 08:15:51: Last modified date: Thu May 12 07:31:50 2016
05/13/2016 08:15:51: Build type: Release
05/13/2016 08:15:51: Build target: GPU
05/13/2016 08:15:51: With 1bit-SGD: no
05/13/2016 08:15:51: CUDA_PATH: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5
05/13/2016 08:15:51: CUB_PATH: c:\src\cub-1.4.1
05/13/2016 08:15:51: CUDNN_PATH: c:\NVIDIA\cudnn-4.0\cuda
05/13/2016 08:15:51: Build Branch: HEAD
05/13/2016 08:15:51: Build SHA1: 35fadc316f045d843bbd9b85061250a959268787
05/13/2016 08:15:51: Built by svcphil on Philly-Pool3
05/13/2016 08:15:51: Build Path: c:\Jenkins\workspace\CNTK-Build-Windows\Source\CNTK\
05/13/2016 08:15:51: -------------------------------------------------------------------
05/13/2016 08:15:51: Running on Philly-Pool2 at 2016/05/13 08:15:51
05/13/2016 08:15:51: Command line:
C:\jenkins\workspace\CNTK-Test-Windows-W1\x64\release\cntk.exe configFile=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/Config/01_OneHidden.cntk currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu\TestData RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu\TestData ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu DeviceId=0 timestamping=true train=[SGD=[maxEpochs=3]] imageLayout="cudnn"
05/13/2016 08:15:51: >>>>>>>>>>>>>>>>>>>> RAW CONFIG (VARIABLES NOT RESOLVED) >>>>>>>>>>>>>>>>>>>>
05/13/2016 08:15:51: RootDir = ".."
ConfigDir = "$RootDir$/Config"
DataDir = "$RootDir$/Data"
OutputDir = "$RootDir$/Output"
ModelDir = "$OutputDir$/Models"
deviceId = 0
imageLayout = "cudnn"
command = train:test
precision = "float"
modelPath = "$ModelDir$/01_OneHidden"
ndlMacros = "$ConfigDir$/Macros.ndl"
traceLevel=1
numMBsToShowResult=500
initOnCPUOnly=true
train = [
action = "train"
NDLNetworkBuilder = [
networkDescription = "$ConfigDir$/01_OneHidden.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.1
momentumPerMB = 0
maxEpochs = 30
]
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
test = [
action = "test"
minibatchSize = 16
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu\TestData
RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu
DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu\TestData
ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config
OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu
DeviceId=0
timestamping=true
train=[SGD=[maxEpochs=3]]
imageLayout="cudnn"
05/13/2016 08:15:51: <<<<<<<<<<<<<<<<<<<< RAW CONFIG (VARIABLES NOT RESOLVED) <<<<<<<<<<<<<<<<<<<<
05/13/2016 08:15:51: >>>>>>>>>>>>>>>>>>>> RAW CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>
05/13/2016 08:15:51: RootDir = ".."
ConfigDir = "../Config"
DataDir = "../Data"
OutputDir = "../Output"
ModelDir = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu/Models"
deviceId = 0
imageLayout = "cudnn"
command = train:test
precision = "float"
modelPath = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu/Models/01_OneHidden"
ndlMacros = "C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/Macros.ndl"
traceLevel=1
numMBsToShowResult=500
initOnCPUOnly=true
train = [
action = "train"
NDLNetworkBuilder = [
networkDescription = "C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/01_OneHidden.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.1
momentumPerMB = 0
maxEpochs = 30
]
reader = [
readerType = "CNTKTextFormatReader"
file = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu\TestData/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
test = [
action = "test"
minibatchSize = 16
reader = [
readerType = "CNTKTextFormatReader"
file = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu\TestData/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu\TestData
RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu
DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu\TestData
ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config
OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu
DeviceId=0
timestamping=true
train=[SGD=[maxEpochs=3]]
imageLayout="cudnn"
05/13/2016 08:15:51: <<<<<<<<<<<<<<<<<<<< RAW CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<
05/13/2016 08:15:51: >>>>>>>>>>>>>>>>>>>> PROCESSED CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>
configparameters: 01_OneHidden.cntk:command=train:test
configparameters: 01_OneHidden.cntk:ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config
configparameters: 01_OneHidden.cntk:currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu\TestData
configparameters: 01_OneHidden.cntk:DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu\TestData
configparameters: 01_OneHidden.cntk:deviceId=0
configparameters: 01_OneHidden.cntk:imageLayout=cudnn
configparameters: 01_OneHidden.cntk:initOnCPUOnly=true
configparameters: 01_OneHidden.cntk:ModelDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu/Models
configparameters: 01_OneHidden.cntk:modelPath=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu/Models/01_OneHidden
configparameters: 01_OneHidden.cntk:ndlMacros=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/Macros.ndl
configparameters: 01_OneHidden.cntk:numMBsToShowResult=500
configparameters: 01_OneHidden.cntk:OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu
configparameters: 01_OneHidden.cntk:precision=float
configparameters: 01_OneHidden.cntk:RootDir=..
configparameters: 01_OneHidden.cntk:RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu
configparameters: 01_OneHidden.cntk:test=[
action = "test"
minibatchSize = 16
reader = [
readerType = "CNTKTextFormatReader"
file = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu\TestData/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
configparameters: 01_OneHidden.cntk:timestamping=true
configparameters: 01_OneHidden.cntk:traceLevel=1
configparameters: 01_OneHidden.cntk:train=[
action = "train"
NDLNetworkBuilder = [
networkDescription = "C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/01_OneHidden.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.1
momentumPerMB = 0
maxEpochs = 30
]
reader = [
readerType = "CNTKTextFormatReader"
file = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu\TestData/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
] [SGD=[maxEpochs=3]]
05/13/2016 08:15:51: <<<<<<<<<<<<<<<<<<<< PROCESSED CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<
05/13/2016 08:15:51: Commands: train test
05/13/2016 08:15:51: Precision = "float"
05/13/2016 08:15:51: CNTKModelPath: C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu/Models/01_OneHidden
05/13/2016 08:15:51: CNTKCommandTrainInfo: train : 3
05/13/2016 08:15:51: CNTKCommandTrainInfo: CNTKNoMoreCommands_Total : 3
05/13/2016 08:15:51: ##############################################################################
05/13/2016 08:15:51: # #
05/13/2016 08:15:51: # Action "train" #
05/13/2016 08:15:51: # #
05/13/2016 08:15:51: ##############################################################################
05/13/2016 08:15:51: CNTKCommandTrainBegin: train
NDLBuilder Using GPU 0
05/13/2016 08:15:52: Creating virgin network.
Post-processing network...
4 roots:
ce = CrossEntropyWithSoftmax()
err = ErrorPrediction()
errTop1 = ErrorPrediction()
ol.z = Plus()
Validating network. 17 nodes to process in pass 1.
Validating --> labels = InputValue() : -> [10 x *]
Validating --> ol.W = LearnableParameter() : -> [10 x 200]
Validating --> h1.W = LearnableParameter() : -> [200 x 784]
Validating --> featScale = LearnableParameter() : -> [1 x 1]
Validating --> features = InputValue() : -> [784 x *]
Validating --> featScaled = ElementTimes (featScale, features) : [1 x 1], [784 x *] -> [784 x 1 x *]
Validating --> h1.t = Times (h1.W, featScaled) : [200 x 784], [784 x 1 x *] -> [200 x 1 x *]
Validating --> h1.b = LearnableParameter() : -> [200 x 1]
Validating --> h1.z = Plus (h1.t, h1.b) : [200 x 1 x *], [200 x 1] -> [200 x 1 x *]
Validating --> h1.y = Sigmoid (h1.z) : [200 x 1 x *] -> [200 x 1 x *]
Validating --> ol.t = Times (ol.W, h1.y) : [10 x 200], [200 x 1 x *] -> [10 x 1 x *]
Validating --> ol.b = LearnableParameter() : -> [10 x 1]
Validating --> ol.z = Plus (ol.t, ol.b) : [10 x 1 x *], [10 x 1] -> [10 x 1 x *]
Validating --> ce = CrossEntropyWithSoftmax (labels, ol.z) : [10 x *], [10 x 1 x *] -> [1]
Validating --> err = ErrorPrediction (labels, ol.z) : [10 x *], [10 x 1 x *] -> [1]
Validating --> unnamed81 = LearnableParameter() : -> [1 x 1]
Validating --> errTop1 = ErrorPrediction (labels, ol.z, unnamed81) : [10 x *], [10 x 1 x *], [1 x 1] -> [1]
Validating network. 9 nodes to process in pass 2.
Validating network, final pass.
9 out of 17 nodes do not share the minibatch layout with the input data.
Post-processing network complete.
05/13/2016 08:15:53: Created model with 17 nodes on GPU 0.
05/13/2016 08:15:53: Training criterion node(s):
05/13/2016 08:15:53: ce = CrossEntropyWithSoftmax
05/13/2016 08:15:53: Evaluation criterion node(s):
05/13/2016 08:15:53: errTop1 = ErrorPrediction
05/13/2016 08:15:53: err = ErrorPrediction
Allocating matrices for forward and/or backward propagation.
Memory Sharing Structure:
0000000000000000: {[err Gradient[1]] [errTop1 Gradient[1]] [featScale Gradient[1 x 1]] [featScaled Gradient[784 x 1 x *]] [features Gradient[784 x *]] [labels Gradient[10 x *]] [unnamed81 Gradient[1 x 1]] }
000000780D2D93A0: {[unnamed81 Value[1 x 1]] }
000000780D2D9440: {[featScaled Value[784 x 1 x *]] }
000000780D2D94E0: {[h1.W Gradient[200 x 784]] [h1.z Value[200 x 1 x *]] }
000000780D2D9580: {[h1.t Gradient[200 x 1 x *]] [h1.y Value[200 x 1 x *]] }
000000780D2D9620: {[h1.z Gradient[200 x 1 x *]] [ol.t Value[10 x 1 x *]] }
000000780D2D96C0: {[ol.W Value[10 x 200]] }
000000780D2D9760: {[ol.b Value[10 x 1]] }
000000780D2D99E0: {[errTop1 Value[1]] }
000000780D2D9EE0: {[err Value[1]] }
000000780D2DA0C0: {[ol.z Value[10 x 1 x *]] }
000000780D2DA160: {[ce Value[1]] }
000000780D2DA2A0: {[h1.t Value[200 x 1 x *]] }
000000780D33AB50: {[ce Gradient[1]] }
000000780D33ABF0: {[ol.t Gradient[10 x 1 x *]] }
000000780D33AFB0: {[ol.b Gradient[10 x 1]] }
000000780D33C270: {[h1.b Gradient[200 x 1]] [h1.y Gradient[200 x 1 x *]] }
000000780D33C9F0: {[ol.W Gradient[10 x 200]] [ol.z Gradient[10 x 1 x *]] }
00000078767789E0: {[featScale Value[1 x 1]] }
0000007876778A80: {[labels Value[10 x *]] }
0000007876778B20: {[h1.W Value[200 x 784]] }
0000007876778BC0: {[h1.b Value[200 x 1]] }
000000787677A1A0: {[features Value[784 x *]] }
05/13/2016 08:15:53: No PreCompute nodes found, skipping PreCompute step.
05/13/2016 08:15:53: Starting Epoch 1: learning rate per sample = 0.003125 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 08:15:53: Starting minibatch loop.
05/13/2016 08:15:56: Epoch[ 1 of 3]-Minibatch[ 1- 500, 26.67%]: ce = 1.29023352 * 16000; errTop1 = 0.37981250 * 16000; err = 0.37981250 * 16000; time = 3.1210s; samplesPerSecond = 5126.5
05/13/2016 08:15:57: Epoch[ 1 of 3]-Minibatch[ 501-1000, 53.33%]: ce = 0.50742346 * 16000; errTop1 = 0.13900000 * 16000; err = 0.13900000 * 16000; time = 0.6202s; samplesPerSecond = 25796.5
05/13/2016 08:15:57: Epoch[ 1 of 3]-Minibatch[1001-1500, 80.00%]: ce = 0.39415479 * 16000; errTop1 = 0.11081250 * 16000; err = 0.11081250 * 16000; time = 0.6195s; samplesPerSecond = 25828.0
05/13/2016 08:15:58: Finished Epoch[ 1 of 3]: [Training] ce = 0.65521146 * 60000; errTop1 = 0.18846667 * 60000; err = 0.18846667 * 60000; totalSamplesSeen = 60000; learningRatePerSample = 0.003125; epochTime=4.86409s
05/13/2016 08:15:58: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu/Models/01_OneHidden.1'
05/13/2016 08:15:58: Starting Epoch 2: learning rate per sample = 0.003125 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 08:15:58: Starting minibatch loop.
05/13/2016 08:15:59: Epoch[ 2 of 3]-Minibatch[ 1- 500, 26.67%]: ce = 0.33321408 * 16000; errTop1 = 0.09581250 * 16000; err = 0.09581250 * 16000; time = 0.6590s; samplesPerSecond = 24277.8
05/13/2016 08:15:59: Epoch[ 2 of 3]-Minibatch[ 501-1000, 53.33%]: ce = 0.31547781 * 16000; errTop1 = 0.09287500 * 16000; err = 0.09287500 * 16000; time = 0.6704s; samplesPerSecond = 23866.0
05/13/2016 08:16:00: Epoch[ 2 of 3]-Minibatch[1001-1500, 80.00%]: ce = 0.31882251 * 16000; errTop1 = 0.09218750 * 16000; err = 0.09218750 * 16000; time = 0.6720s; samplesPerSecond = 23808.7
05/13/2016 08:16:00: Finished Epoch[ 2 of 3]: [Training] ce = 0.31533239 * 60000; errTop1 = 0.09158333 * 60000; err = 0.09158333 * 60000; totalSamplesSeen = 120000; learningRatePerSample = 0.003125; epochTime=2.52448s
05/13/2016 08:16:00: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu/Models/01_OneHidden.2'
05/13/2016 08:16:00: Starting Epoch 3: learning rate per sample = 0.003125 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 08:16:00: Starting minibatch loop.
05/13/2016 08:16:01: Epoch[ 3 of 3]-Minibatch[ 1- 500, 26.67%]: ce = 0.28751190 * 16000; errTop1 = 0.08393750 * 16000; err = 0.08393750 * 16000; time = 0.6195s; samplesPerSecond = 25825.2
05/13/2016 08:16:02: Epoch[ 3 of 3]-Minibatch[ 501-1000, 53.33%]: ce = 0.27455951 * 16000; errTop1 = 0.07950000 * 16000; err = 0.07950000 * 16000; time = 0.6193s; samplesPerSecond = 25834.3
05/13/2016 08:16:02: Epoch[ 3 of 3]-Minibatch[1001-1500, 80.00%]: ce = 0.27693610 * 16000; errTop1 = 0.07987500 * 16000; err = 0.07987500 * 16000; time = 0.6192s; samplesPerSecond = 25839.8
05/13/2016 08:16:03: Finished Epoch[ 3 of 3]: [Training] ce = 0.27493141 * 60000; errTop1 = 0.07983333 * 60000; err = 0.07983333 * 60000; totalSamplesSeen = 180000; learningRatePerSample = 0.003125; epochTime=2.34147s
05/13/2016 08:16:03: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_01_OneHidden@release_gpu/Models/01_OneHidden'
05/13/2016 08:16:03: CNTKCommandTrainEnd: train
05/13/2016 08:16:03: Action "train" complete.
05/13/2016 08:16:03: ##############################################################################
05/13/2016 08:16:03: # #
05/13/2016 08:16:03: # Action "test" #
05/13/2016 08:16:03: # #
05/13/2016 08:16:03: ##############################################################################
Post-processing network...
4 roots:
ce = CrossEntropyWithSoftmax()
err = ErrorPrediction()
errTop1 = ErrorPrediction()
ol.z = Plus()
Validating network. 17 nodes to process in pass 1.
Validating --> labels = InputValue() : -> [10 x *1]
Validating --> ol.W = LearnableParameter() : -> [10 x 200]
Validating --> h1.W = LearnableParameter() : -> [200 x 784]
Validating --> featScale = LearnableParameter() : -> [1 x 1]
Validating --> features = InputValue() : -> [784 x *1]
Validating --> featScaled = ElementTimes (featScale, features) : [1 x 1], [784 x *1] -> [784 x 1 x *1]
Validating --> h1.t = Times (h1.W, featScaled) : [200 x 784], [784 x 1 x *1] -> [200 x 1 x *1]
Validating --> h1.b = LearnableParameter() : -> [200 x 1]
Validating --> h1.z = Plus (h1.t, h1.b) : [200 x 1 x *1], [200 x 1] -> [200 x 1 x *1]
Validating --> h1.y = Sigmoid (h1.z) : [200 x 1 x *1] -> [200 x 1 x *1]
Validating --> ol.t = Times (ol.W, h1.y) : [10 x 200], [200 x 1 x *1] -> [10 x 1 x *1]
Validating --> ol.b = LearnableParameter() : -> [10 x 1]
Validating --> ol.z = Plus (ol.t, ol.b) : [10 x 1 x *1], [10 x 1] -> [10 x 1 x *1]
Validating --> ce = CrossEntropyWithSoftmax (labels, ol.z) : [10 x *1], [10 x 1 x *1] -> [1]
Validating --> err = ErrorPrediction (labels, ol.z) : [10 x *1], [10 x 1 x *1] -> [1]
Validating --> unnamed81 = LearnableParameter() : -> [1 x 1]
Validating --> errTop1 = ErrorPrediction (labels, ol.z, unnamed81) : [10 x *1], [10 x 1 x *1], [1 x 1] -> [1]
Validating network. 9 nodes to process in pass 2.
Validating network, final pass.
9 out of 17 nodes do not share the minibatch layout with the input data.
Post-processing network complete.
evalNodeNames are not specified, using all the default evalnodes and training criterion nodes.
Allocating matrices for forward and/or backward propagation.
Memory Sharing Structure:
0000000000000000: {[ce Gradient[1]] [err Gradient[1]] [errTop1 Gradient[1]] [featScale Gradient[1 x 1]] [featScaled Gradient[784 x 1 x *1]] [features Gradient[784 x *1]] [h1.W Gradient[200 x 784]] [h1.b Gradient[200 x 1]] [h1.t Gradient[200 x 1 x *1]] [h1.y Gradient[200 x 1 x *1]] [h1.z Gradient[200 x 1 x *1]] [labels Gradient[10 x *1]] [ol.W Gradient[10 x 200]] [ol.b Gradient[10 x 1]] [ol.t Gradient[10 x 1 x *1]] [ol.z Gradient[10 x 1 x *1]] [unnamed81 Gradient[1 x 1]] }
000000780D33B230: {[labels Value[10 x *1]] }
000000780D33BA50: {[ol.b Value[10 x 1]] }
000000780D33BD70: {[featScale Value[1 x 1]] }
000000780D33BF50: {[h1.b Value[200 x 1]] }
000000780D33C6D0: {[features Value[784 x *1]] }
000000780D33C770: {[h1.W Value[200 x 784]] }
000000787673E350: {[ol.z Value[10 x 1 x *1]] }
000000787673E850: {[ol.t Value[10 x 1 x *1]] }
00000078767789E0: {[ol.W Value[10 x 200]] }
0000007876778A80: {[unnamed81 Value[1 x 1]] }
0000007876779020: {[errTop1 Value[1]] }
00000078767790C0: {[err Value[1]] }
0000007876779160: {[ce Value[1]] }
00000078767792A0: {[h1.t Value[200 x 1 x *1]] }
00000078767793E0: {[h1.z Value[200 x 1 x *1]] }
00000078767795C0: {[h1.y Value[200 x 1 x *1]] }
00000078767797A0: {[featScaled Value[784 x 1 x *1]] }
05/13/2016 08:16:11: Final Results: Minibatch[1-625]: errTop1 = 0.07460000 * 10000; err = 0.07460000 * 10000; ce = 0.26425332 * 10000; perplexity = 1.30245809
05/13/2016 08:16:11: Action "test" complete.
05/13/2016 08:16:11: __COMPLETED__

11
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden/run-test Executable file
Просмотреть файл

@ -0,0 +1,11 @@
#!/bin/bash
. $TEST_DIR/../run-test-common
cntkrun $CNTKTextConfigDir/01_OneHidden.cntk "train=[SGD=[maxEpochs=3]] imageLayout=\"$imageLayout\""
ExitCode=$?
# Delete the test data if copied
[[ "$Copied" -eq "1" ]] && rm -rf "$DataDir"
exit $ExitCode

41
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/01_OneHidden/testcases.yml Normal file
Просмотреть файл

@ -0,0 +1,41 @@
dataDir: ../../../../../../../Examples/Image/MNIST/Data
tags:
# In BVT, run Release GPU
- bvt-e (build_sku=='gpu') and (device=='gpu') and (flavor=='release')
# In Nightly on Linux, additionally run Debug GPU and Release CPU
- nightly-e (build_sku=='gpu') and (((device=='gpu') and (flavor=='release')) or (os=='linux' and ((flavor=='debug') ^ (device=='cpu'))))
testCases:
CNTK Run must be completed:
patterns:
- __COMPLETED__
Must train epochs in exactly same order and parameters:
patterns:
- Starting Epoch {{integer}}
- learning rate per sample = {{float}}
- momentum = {{float}}
Epochs must be finished with expected results:
patterns:
- Finished Epoch[{{integer}} of {{integer}}]
- ce = {{float,tolerance=.1%}} * {{integer}}
- errTop1 = {{float,tolerance=.1%}} * {{integer}}
- err = {{float,tolerance=.1%}} * {{integer}}
- totalSamplesSeen = {{integer}}
- learningRatePerSample = {{float,tolerance=0.001%}}
Per-minibatch training results must match:
patterns:
- Epoch[{{integer}} of {{integer}}]-Minibatch[{{integer}}-{{integer}}
- ce = {{float,tolerance=.1%}} * {{integer}}
- errTop1 = {{float,tolerance=.1%}} * {{integer}}
- err = {{float,tolerance=.1%}} * {{integer}}
Final test results must match:
patterns:
- "Final Results: Minibatch[{{integer}}-{{integer}}]"
- errTop1 = {{float,tolerance=.1%}} * {{integer}}
- err = {{float,tolerance=.1%}} * {{integer}}
- ce = {{float,tolerance=.1%}} * {{integer}}

9
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution/README.txt Normal file
Просмотреть файл

@ -0,0 +1,9 @@
Test runtimes
Windows:
Running test CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution (debug gpu) - [OK] 157.53 sec
Running test CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution (release gpu) - [OK] 64.93 sec
Linux:
Running test CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution (debug gpu) - [OK] 40.43 sec
Running test CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution (release gpu) - [OK] 18.92 sec

524
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution/baseline.linux.txt Normal file
Просмотреть файл

@ -0,0 +1,524 @@
=== Running /home/philly/jenkins/workspace/CNTK-Test-Linux-W1/build/gpu/release/bin/cntk configFile=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution/../../../../../../../Examples/Image/MNIST/Config/../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/Config/02_Convolution.cntk currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/TestData RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/TestData ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution/../../../../../../../Examples/Image/MNIST/Config OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu DeviceId=0 timestamping=true train=[SGD=[maxEpochs=3]] imageLayout="cudnn"
-------------------------------------------------------------------
Build info:
Built time: May 13 2016 14:50:25
Last modified date: Thu May 12 14:00:37 2016
Build type: release
Build target: GPU
With 1bit-SGD: no
Math lib: acml
CUDA_PATH: /usr/local/cuda-7.5
CUB_PATH: /usr/local/cub-1.4.1
CUDNN_PATH: /usr/local/cudnn-4.0
Build Branch: HEAD
Build SHA1: 35fadc316f045d843bbd9b85061250a959268787
Built by philly on d8dc82703b0f
Build Path: /home/philly/jenkins/workspace/CNTK-Build-Linux
-------------------------------------------------------------------
Changed current directory to /tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/TestData
05/13/2016 15:10:11: -------------------------------------------------------------------
05/13/2016 15:10:11: Build info:
05/13/2016 15:10:11: Built time: May 13 2016 14:50:25
05/13/2016 15:10:11: Last modified date: Thu May 12 14:00:37 2016
05/13/2016 15:10:11: Build type: release
05/13/2016 15:10:11: Build target: GPU
05/13/2016 15:10:11: With 1bit-SGD: no
05/13/2016 15:10:11: Math lib: acml
05/13/2016 15:10:11: CUDA_PATH: /usr/local/cuda-7.5
05/13/2016 15:10:11: CUB_PATH: /usr/local/cub-1.4.1
05/13/2016 15:10:11: CUDNN_PATH: /usr/local/cudnn-4.0
05/13/2016 15:10:11: Build Branch: HEAD
05/13/2016 15:10:11: Build SHA1: 35fadc316f045d843bbd9b85061250a959268787
05/13/2016 15:10:11: Built by philly on d8dc82703b0f
05/13/2016 15:10:11: Build Path: /home/philly/jenkins/workspace/CNTK-Build-Linux
05/13/2016 15:10:11: -------------------------------------------------------------------
05/13/2016 15:10:11: Running on localhost at 2016/05/13 15:10:11
05/13/2016 15:10:11: Command line:
/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/build/gpu/release/bin/cntk configFile=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution/../../../../../../../Examples/Image/MNIST/Config/../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/Config/02_Convolution.cntk currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/TestData RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/TestData ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution/../../../../../../../Examples/Image/MNIST/Config OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu DeviceId=0 timestamping=true train=[SGD=[maxEpochs=3]] imageLayout="cudnn"
05/13/2016 15:10:11: >>>>>>>>>>>>>>>>>>>> RAW CONFIG (VARIABLES NOT RESOLVED) >>>>>>>>>>>>>>>>>>>>
05/13/2016 15:10:11: RootDir = ".."
ConfigDir = "$RootDir$/Config"
DataDir = "$RootDir$/Data"
OutputDir = "$RootDir$/Output"
ModelDir = "$OutputDir$/Models"
deviceId = 0
imageLayout = "cudnn"
command = train:test
precision = "float"
modelPath = "$ModelDir$/02_Convolution"
ndlMacros = "$ConfigDir$/Macros.ndl"
traceLevel=1
numMBsToShowResult=500
prefetch=true
initOnCPUOnly=true
train = [
action = "train"
NDLNetworkBuilder = [
networkDescription = "$ConfigDir$/02_Convolution.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.1*5:0.3
momentumPerMB = 0*10:0.7
maxEpochs = 15
]
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
test = [
action = test
minibatchSize = 16
NDLNetworkBuilder = [
networkDescription = "$ConfigDir$/02_Convolution.ndl"
]
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/TestData
RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu
DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/TestData
ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution/../../../../../../../Examples/Image/MNIST/Config
OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu
DeviceId=0
timestamping=true
train=[SGD=[maxEpochs=3]]
imageLayout="cudnn"
05/13/2016 15:10:11: <<<<<<<<<<<<<<<<<<<< RAW CONFIG (VARIABLES NOT RESOLVED) <<<<<<<<<<<<<<<<<<<<
05/13/2016 15:10:11: >>>>>>>>>>>>>>>>>>>> RAW CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>
05/13/2016 15:10:11: RootDir = ".."
ConfigDir = "../Config"
DataDir = "../Data"
OutputDir = "../Output"
ModelDir = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/Models"
deviceId = 0
imageLayout = "cudnn"
command = train:test
precision = "float"
modelPath = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/Models/02_Convolution"
ndlMacros = "/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution/../../../../../../../Examples/Image/MNIST/Config/Macros.ndl"
traceLevel=1
numMBsToShowResult=500
prefetch=true
initOnCPUOnly=true
train = [
action = "train"
NDLNetworkBuilder = [
networkDescription = "/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution/../../../../../../../Examples/Image/MNIST/Config/02_Convolution.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.1*5:0.3
momentumPerMB = 0*10:0.7
maxEpochs = 15
]
reader = [
readerType = "CNTKTextFormatReader"
file = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/TestData/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
test = [
action = test
minibatchSize = 16
NDLNetworkBuilder = [
networkDescription = "/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution/../../../../../../../Examples/Image/MNIST/Config/02_Convolution.ndl"
]
reader = [
readerType = "CNTKTextFormatReader"
file = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/TestData/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/TestData
RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu
DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/TestData
ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution/../../../../../../../Examples/Image/MNIST/Config
OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu
DeviceId=0
timestamping=true
train=[SGD=[maxEpochs=3]]
imageLayout="cudnn"
05/13/2016 15:10:11: <<<<<<<<<<<<<<<<<<<< RAW CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<
05/13/2016 15:10:11: >>>>>>>>>>>>>>>>>>>> PROCESSED CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>
configparameters: 02_Convolution.cntk:command=train:test
configparameters: 02_Convolution.cntk:ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution/../../../../../../../Examples/Image/MNIST/Config
configparameters: 02_Convolution.cntk:currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/TestData
configparameters: 02_Convolution.cntk:DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/TestData
configparameters: 02_Convolution.cntk:deviceId=0
configparameters: 02_Convolution.cntk:imageLayout=cudnn
configparameters: 02_Convolution.cntk:initOnCPUOnly=true
configparameters: 02_Convolution.cntk:ModelDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/Models
configparameters: 02_Convolution.cntk:modelPath=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/Models/02_Convolution
configparameters: 02_Convolution.cntk:ndlMacros=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution/../../../../../../../Examples/Image/MNIST/Config/Macros.ndl
configparameters: 02_Convolution.cntk:numMBsToShowResult=500
configparameters: 02_Convolution.cntk:OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu
configparameters: 02_Convolution.cntk:precision=float
configparameters: 02_Convolution.cntk:prefetch=true
configparameters: 02_Convolution.cntk:RootDir=..
configparameters: 02_Convolution.cntk:RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu
configparameters: 02_Convolution.cntk:test=[
action = test
minibatchSize = 16
NDLNetworkBuilder = [
networkDescription = "/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution/../../../../../../../Examples/Image/MNIST/Config/02_Convolution.ndl"
]
reader = [
readerType = "CNTKTextFormatReader"
file = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/TestData/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
configparameters: 02_Convolution.cntk:timestamping=true
configparameters: 02_Convolution.cntk:traceLevel=1
configparameters: 02_Convolution.cntk:train=[
action = "train"
NDLNetworkBuilder = [
networkDescription = "/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution/../../../../../../../Examples/Image/MNIST/Config/02_Convolution.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.1*5:0.3
momentumPerMB = 0*10:0.7
maxEpochs = 15
]
reader = [
readerType = "CNTKTextFormatReader"
file = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/TestData/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
] [SGD=[maxEpochs=3]]
05/13/2016 15:10:11: <<<<<<<<<<<<<<<<<<<< PROCESSED CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<
05/13/2016 15:10:11: Commands: train test
05/13/2016 15:10:11: Precision = "float"
05/13/2016 15:10:11: CNTKModelPath: /tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/Models/02_Convolution
05/13/2016 15:10:11: CNTKCommandTrainInfo: train : 3
05/13/2016 15:10:11: CNTKCommandTrainInfo: CNTKNoMoreCommands_Total : 3
05/13/2016 15:10:11: ##############################################################################
05/13/2016 15:10:11: # #
05/13/2016 15:10:11: # Action "train" #
05/13/2016 15:10:11: # #
05/13/2016 15:10:11: ##############################################################################
05/13/2016 15:10:11: CNTKCommandTrainBegin: train
NDLBuilder Using GPU 0
05/13/2016 15:10:11: Creating virgin network.
Post-processing network...
3 roots:
ce = CrossEntropyWithSoftmax()
err = ErrorPrediction()
ol.z = Plus()
Validating network. 27 nodes to process in pass 1.
Validating --> labels = InputValue() : -> [10 x *]
Validating --> ol.W = LearnableParameter() : -> [10 x 128]
Validating --> h1.W = LearnableParameter() : -> [128 x 7 x 7 x 32]
Validating --> conv2.w.W = LearnableParameter() : -> [32 x 400]
Validating --> conv1.w.W = LearnableParameter() : -> [16 x 25]
Validating --> featScale = LearnableParameter() : -> [1 x 1]
Validating --> features = InputValue() : -> [28 x 28 x 1 x *]
Validating --> featScaled = ElementTimes (featScale, features) : [1 x 1], [28 x 28 x 1 x *] -> [28 x 28 x 1 x *]
Validating --> conv1.c.c = Convolution (conv1.w.W, featScaled) : [16 x 25], [28 x 28 x 1 x *] -> [28 x 28 x 16 x *]
Validating --> conv1.b.b = LearnableParameter() : -> [1 x 1 x 16]
Validating --> conv1.cpb = Plus (conv1.c.c, conv1.b.b) : [28 x 28 x 16 x *], [1 x 1 x 16] -> [28 x 28 x 16 x *]
Validating --> conv1.out = RectifiedLinear (conv1.cpb) : [28 x 28 x 16 x *] -> [28 x 28 x 16 x *]
Validating --> pool1 = MaxPooling (conv1.out) : [28 x 28 x 16 x *] -> [14 x 14 x 16 x *]
Validating --> conv2.c.c = Convolution (conv2.w.W, pool1) : [32 x 400], [14 x 14 x 16 x *] -> [14 x 14 x 32 x *]
Validating --> conv2.b.b = LearnableParameter() : -> [1 x 1 x 32]
Validating --> conv2.cpb = Plus (conv2.c.c, conv2.b.b) : [14 x 14 x 32 x *], [1 x 1 x 32] -> [14 x 14 x 32 x *]
Validating --> conv2.out = RectifiedLinear (conv2.cpb) : [14 x 14 x 32 x *] -> [14 x 14 x 32 x *]
Validating --> pool2.p = Pooling (conv2.out) : [14 x 14 x 32 x *] -> [7 x 7 x 32 x *]
Validating --> h1.t = Times (h1.W, pool2.p) : [128 x 7 x 7 x 32], [7 x 7 x 32 x *] -> [128 x *]
Validating --> h1.b = LearnableParameter() : -> [128 x 1]
Validating --> h1.z = Plus (h1.t, h1.b) : [128 x *], [128 x 1] -> [128 x 1 x *]
Validating --> h1.y = Sigmoid (h1.z) : [128 x 1 x *] -> [128 x 1 x *]
Validating --> ol.t = Times (ol.W, h1.y) : [10 x 128], [128 x 1 x *] -> [10 x 1 x *]
Validating --> ol.b = LearnableParameter() : -> [10 x 1]
Validating --> ol.z = Plus (ol.t, ol.b) : [10 x 1 x *], [10 x 1] -> [10 x 1 x *]
Validating --> ce = CrossEntropyWithSoftmax (labels, ol.z) : [10 x *], [10 x 1 x *] -> [1]
Validating --> err = ErrorPrediction (labels, ol.z) : [10 x *], [10 x 1 x *] -> [1]
Validating network. 16 nodes to process in pass 2.
Validating network, final pass.
Using cuDNN convolution engine for geometry: Input: 28 x 28 x 1, Output: 28 x 28 x 16, Kernel: 5 x 5 x 1, Map: 1 x 1 x 16, Stride: 1 x 1 x 1, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 28 x 28 x 16, Output: 14 x 14 x 16, Kernel: 2 x 2 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 14 x 14 x 16, Output: 14 x 14 x 32, Kernel: 5 x 5 x 16, Map: 32, Stride: 1 x 1 x 16, Sharing: (1, 1, 1), AutoPad: (1, 1, 0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 14 x 14 x 32, Output: 7 x 7 x 32, Kernel: 2 x 2 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (1, 1, 0), LowerPad: 0, UpperPad: 0.
11 out of 27 nodes do not share the minibatch layout with the input data.
Post-processing network complete.
05/13/2016 15:10:11: Created model with 27 nodes on GPU 0.
05/13/2016 15:10:11: Training criterion node(s):
05/13/2016 15:10:11: ce = CrossEntropyWithSoftmax
05/13/2016 15:10:11: Evaluation criterion node(s):
05/13/2016 15:10:11: err = ErrorPrediction
Allocating matrices for forward and/or backward propagation.
Memory Sharing Structure:
(nil): {[err Gradient[1]] [featScale Gradient[1 x 1]] [featScaled Gradient[28 x 28 x 1 x *]] [features Gradient[28 x 28 x 1 x *]] [labels Gradient[10 x *]] }
0x132d628: {[features Value[28 x 28 x 1 x *]] }
0x1854138: {[featScale Value[1 x 1]] }
0x1ca8388: {[labels Value[10 x *]] }
0x1ca8b18: {[conv1.w.W Value[16 x 25]] }
0x1ca9778: {[conv1.b.b Value[1 x 1 x 16]] }
0x1caaa88: {[conv2.w.W Value[32 x 400]] }
0x1cac278: {[conv2.b.b Value[1 x 1 x 32]] }
0x1cb04f8: {[h1.W Value[128 x 7 x 7 x 32]] }
0x1cb1728: {[h1.b Value[128 x 1]] }
0x1cb2318: {[ol.W Value[10 x 128]] }
0x1cb3468: {[ol.b Value[10 x 1]] }
0x7f427f204c08: {[conv1.c.c Value[28 x 28 x 16 x *]] }
0x7f427f20bd48: {[h1.b Gradient[128 x 1]] [h1.y Gradient[128 x 1 x *]] }
0x7f427f4d3118: {[err Value[1]] }
0x7f427f4e3b08: {[featScaled Value[28 x 28 x 1 x *]] }
0x7f427f4e3db8: {[conv1.cpb Value[28 x 28 x 16 x *]] [conv1.w.W Gradient[16 x 25]] }
0x7f427f4e42d8: {[conv1.c.c Gradient[28 x 28 x 16 x *]] [conv1.out Value[28 x 28 x 16 x *]] }
0x7f427f4e4498: {[conv1.cpb Gradient[28 x 28 x 16 x *]] [pool1 Value[14 x 14 x 16 x *]] }
0x7f427f4e4658: {[conv2.c.c Value[14 x 14 x 32 x *]] }
0x7f427f4e4818: {[conv1.b.b Gradient[1 x 1 x 16]] [conv1.out Gradient[28 x 28 x 16 x *]] }
0x7f427f4e49d8: {[conv2.cpb Value[14 x 14 x 32 x *]] [conv2.w.W Gradient[32 x 400]] }
0x7f427f4e4b98: {[conv2.c.c Gradient[14 x 14 x 32 x *]] [conv2.out Value[14 x 14 x 32 x *]] }
0x7f427f4e4d58: {[conv2.cpb Gradient[14 x 14 x 32 x *]] [pool1 Gradient[14 x 14 x 16 x *]] [pool2.p Value[7 x 7 x 32 x *]] }
0x7f427f4e4f18: {[conv2.b.b Gradient[1 x 1 x 32]] [conv2.out Gradient[14 x 14 x 32 x *]] [h1.t Value[128 x *]] }
0x7f427f4e50d8: {[h1.W Gradient[128 x 7 x 7 x 32]] [h1.z Value[128 x 1 x *]] }
0x7f427f4e5298: {[h1.t Gradient[128 x *]] [h1.y Value[128 x 1 x *]] }
0x7f427f4e5458: {[h1.z Gradient[128 x 1 x *]] [ol.t Value[10 x 1 x *]] [pool2.p Gradient[7 x 7 x 32 x *]] }
0x7f427f4e5f38: {[ce Gradient[1]] }
0x7f427f4e60f8: {[ol.W Gradient[10 x 128]] [ol.z Gradient[10 x 1 x *]] }
0x7f427f4e62b8: {[ol.t Gradient[10 x 1 x *]] }
0x7f427f4e6478: {[ol.b Gradient[10 x 1]] }
0x7f427f4ff658: {[ce Value[1]] }
0x7f427f4ffea8: {[ol.z Value[10 x 1 x *]] }
05/13/2016 15:10:11: No PreCompute nodes found, skipping PreCompute step.
05/13/2016 15:10:11: Starting Epoch 1: learning rate per sample = 0.003125 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 15:10:11: Starting minibatch loop.
05/13/2016 15:10:13: Epoch[ 1 of 3]-Minibatch[ 1- 500, 26.67%]: ce = 1.05460791 * 16000; err = 0.35256250 * 16000; time = 2.0377s; samplesPerSecond = 7852.2
05/13/2016 15:10:14: Epoch[ 1 of 3]-Minibatch[ 501-1000, 53.33%]: ce = 0.16176135 * 16000; err = 0.04425000 * 16000; time = 0.9884s; samplesPerSecond = 16187.9
05/13/2016 15:10:15: Epoch[ 1 of 3]-Minibatch[1001-1500, 80.00%]: ce = 0.10889783 * 16000; err = 0.03043750 * 16000; time = 0.9868s; samplesPerSecond = 16214.2
05/13/2016 15:10:16: Finished Epoch[ 1 of 3]: [Training] ce = 0.37214827 * 60000; err = 0.11981667 * 60000; totalSamplesSeen = 60000; learningRatePerSample = 0.003125; epochTime=4.77593s
05/13/2016 15:10:16: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/Models/02_Convolution.1'
05/13/2016 15:10:16: Starting Epoch 2: learning rate per sample = 0.003125 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 15:10:16: Starting minibatch loop.
05/13/2016 15:10:17: Epoch[ 2 of 3]-Minibatch[ 1- 500, 26.67%]: ce = 0.07178102 * 16000; err = 0.02206250 * 16000; time = 0.9982s; samplesPerSecond = 16029.6
05/13/2016 15:10:18: Epoch[ 2 of 3]-Minibatch[ 501-1000, 53.33%]: ce = 0.06225394 * 16000; err = 0.01800000 * 16000; time = 0.9949s; samplesPerSecond = 16082.6
05/13/2016 15:10:19: Epoch[ 2 of 3]-Minibatch[1001-1500, 80.00%]: ce = 0.06624252 * 16000; err = 0.02025000 * 16000; time = 0.9961s; samplesPerSecond = 16062.5
05/13/2016 15:10:19: Finished Epoch[ 2 of 3]: [Training] ce = 0.06652122 * 60000; err = 0.01995000 * 60000; totalSamplesSeen = 120000; learningRatePerSample = 0.003125; epochTime=3.74643s
05/13/2016 15:10:20: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/Models/02_Convolution.2'
05/13/2016 15:10:20: Starting Epoch 3: learning rate per sample = 0.003125 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 15:10:20: Starting minibatch loop.
05/13/2016 15:10:21: Epoch[ 3 of 3]-Minibatch[ 1- 500, 26.67%]: ce = 0.04257084 * 16000; err = 0.01256250 * 16000; time = 0.9942s; samplesPerSecond = 16093.1
05/13/2016 15:10:21: Epoch[ 3 of 3]-Minibatch[ 501-1000, 53.33%]: ce = 0.04675156 * 16000; err = 0.01418750 * 16000; time = 0.9927s; samplesPerSecond = 16118.2
05/13/2016 15:10:22: Epoch[ 3 of 3]-Minibatch[1001-1500, 80.00%]: ce = 0.04904524 * 16000; err = 0.01475000 * 16000; time = 0.9901s; samplesPerSecond = 16160.8
05/13/2016 15:10:23: Finished Epoch[ 3 of 3]: [Training] ce = 0.04529028 * 60000; err = 0.01366667 * 60000; totalSamplesSeen = 180000; learningRatePerSample = 0.003125; epochTime=3.73418s
05/13/2016 15:10:23: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_02_Convolution@release_gpu/Models/02_Convolution'
05/13/2016 15:10:23: CNTKCommandTrainEnd: train
05/13/2016 15:10:23: Action "train" complete.
05/13/2016 15:10:23: ##############################################################################
05/13/2016 15:10:23: # #
05/13/2016 15:10:23: # Action "test" #
05/13/2016 15:10:23: # #
05/13/2016 15:10:23: ##############################################################################
Post-processing network...
3 roots:
ce = CrossEntropyWithSoftmax()
err = ErrorPrediction()
ol.z = Plus()
Validating network. 27 nodes to process in pass 1.
Validating --> labels = InputValue() : -> [10 x *1]
Validating --> ol.W = LearnableParameter() : -> [10 x 128]
Validating --> h1.W = LearnableParameter() : -> [128 x 7 x 7 x 32]
Validating --> conv2.w.W = LearnableParameter() : -> [32 x 400]
Validating --> conv1.w.W = LearnableParameter() : -> [16 x 25]
Validating --> featScale = LearnableParameter() : -> [1 x 1]
Validating --> features = InputValue() : -> [28 x 28 x 1 x *1]
Validating --> featScaled = ElementTimes (featScale, features) : [1 x 1], [28 x 28 x 1 x *1] -> [28 x 28 x 1 x *1]
Validating --> conv1.c.c = Convolution (conv1.w.W, featScaled) : [16 x 25], [28 x 28 x 1 x *1] -> [28 x 28 x 16 x *1]
Validating --> conv1.b.b = LearnableParameter() : -> [1 x 1 x 16]
Validating --> conv1.cpb = Plus (conv1.c.c, conv1.b.b) : [28 x 28 x 16 x *1], [1 x 1 x 16] -> [28 x 28 x 16 x *1]
Validating --> conv1.out = RectifiedLinear (conv1.cpb) : [28 x 28 x 16 x *1] -> [28 x 28 x 16 x *1]
Validating --> pool1 = MaxPooling (conv1.out) : [28 x 28 x 16 x *1] -> [14 x 14 x 16 x *1]
Validating --> conv2.c.c = Convolution (conv2.w.W, pool1) : [32 x 400], [14 x 14 x 16 x *1] -> [14 x 14 x 32 x *1]
Validating --> conv2.b.b = LearnableParameter() : -> [1 x 1 x 32]
Validating --> conv2.cpb = Plus (conv2.c.c, conv2.b.b) : [14 x 14 x 32 x *1], [1 x 1 x 32] -> [14 x 14 x 32 x *1]
Validating --> conv2.out = RectifiedLinear (conv2.cpb) : [14 x 14 x 32 x *1] -> [14 x 14 x 32 x *1]
Validating --> pool2.p = Pooling (conv2.out) : [14 x 14 x 32 x *1] -> [7 x 7 x 32 x *1]
Validating --> h1.t = Times (h1.W, pool2.p) : [128 x 7 x 7 x 32], [7 x 7 x 32 x *1] -> [128 x *1]
Validating --> h1.b = LearnableParameter() : -> [128 x 1]
Validating --> h1.z = Plus (h1.t, h1.b) : [128 x *1], [128 x 1] -> [128 x 1 x *1]
Validating --> h1.y = Sigmoid (h1.z) : [128 x 1 x *1] -> [128 x 1 x *1]
Validating --> ol.t = Times (ol.W, h1.y) : [10 x 128], [128 x 1 x *1] -> [10 x 1 x *1]
Validating --> ol.b = LearnableParameter() : -> [10 x 1]
Validating --> ol.z = Plus (ol.t, ol.b) : [10 x 1 x *1], [10 x 1] -> [10 x 1 x *1]
Validating --> ce = CrossEntropyWithSoftmax (labels, ol.z) : [10 x *1], [10 x 1 x *1] -> [1]
Validating --> err = ErrorPrediction (labels, ol.z) : [10 x *1], [10 x 1 x *1] -> [1]
Validating network. 16 nodes to process in pass 2.
Validating network, final pass.
Using cuDNN convolution engine for geometry: Input: 28 x 28 x 1, Output: 28 x 28 x 16, Kernel: 5 x 5 x 1, Map: 1 x 1 x 16, Stride: 1 x 1 x 1, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 28 x 28 x 16, Output: 14 x 14 x 16, Kernel: 2 x 2 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 14 x 14 x 16, Output: 14 x 14 x 32, Kernel: 5 x 5 x 16, Map: 32, Stride: 1 x 1 x 16, Sharing: (1, 1, 1), AutoPad: (1, 1, 0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 14 x 14 x 32, Output: 7 x 7 x 32, Kernel: 2 x 2 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (1, 1, 0), LowerPad: 0, UpperPad: 0.
11 out of 27 nodes do not share the minibatch layout with the input data.
Post-processing network complete.
evalNodeNames are not specified, using all the default evalnodes and training criterion nodes.
Allocating matrices for forward and/or backward propagation.
Memory Sharing Structure:
(nil): {[ce Gradient[1]] [conv1.b.b Gradient[1 x 1 x 16]] [conv1.c.c Gradient[28 x 28 x 16 x *1]] [conv1.cpb Gradient[28 x 28 x 16 x *1]] [conv1.out Gradient[28 x 28 x 16 x *1]] [conv1.w.W Gradient[16 x 25]] [conv2.b.b Gradient[1 x 1 x 32]] [conv2.c.c Gradient[14 x 14 x 32 x *1]] [conv2.cpb Gradient[14 x 14 x 32 x *1]] [conv2.out Gradient[14 x 14 x 32 x *1]] [conv2.w.W Gradient[32 x 400]] [err Gradient[1]] [featScale Gradient[1 x 1]] [featScaled Gradient[28 x 28 x 1 x *1]] [features Gradient[28 x 28 x 1 x *1]] [h1.W Gradient[128 x 7 x 7 x 32]] [h1.b Gradient[128 x 1]] [h1.t Gradient[128 x *1]] [h1.y Gradient[128 x 1 x *1]] [h1.z Gradient[128 x 1 x *1]] [labels Gradient[10 x *1]] [ol.W Gradient[10 x 128]] [ol.b Gradient[10 x 1]] [ol.t Gradient[10 x 1 x *1]] [ol.z Gradient[10 x 1 x *1]] [pool1 Gradient[14 x 14 x 16 x *1]] [pool2.p Gradient[7 x 7 x 32 x *1]] }
0x7f4274adf028: {[conv1.b.b Value[1 x 1 x 16]] }
0x7f4274adfe98: {[h1.b Value[128 x 1]] }
0x7f427ae42308: {[err Value[1]] }
0x7f427ae62498: {[featScaled Value[28 x 28 x 1 x *1]] }
0x7f427ae62748: {[conv1.c.c Value[28 x 28 x 16 x *1]] }
0x7f427ae62c08: {[conv1.cpb Value[28 x 28 x 16 x *1]] }
0x7f427ae62dc8: {[conv1.out Value[28 x 28 x 16 x *1]] }
0x7f427ae62f88: {[pool1 Value[14 x 14 x 16 x *1]] }
0x7f427ae63148: {[conv2.c.c Value[14 x 14 x 32 x *1]] }
0x7f427ae634c8: {[conv2.cpb Value[14 x 14 x 32 x *1]] }
0x7f427ae63688: {[conv2.out Value[14 x 14 x 32 x *1]] }
0x7f427ae63848: {[pool2.p Value[7 x 7 x 32 x *1]] }
0x7f427ae646a8: {[labels Value[10 x *1]] }
0x7f427ae64b18: {[ol.b Value[10 x 1]] }
0x7f427ae668a8: {[conv2.w.W Value[32 x 400]] }
0x7f427ae72368: {[h1.W Value[128 x 7 x 7 x 32]] }
0x7f427f20cb08: {[ol.W Value[10 x 128]] }
0x7f427f20e888: {[featScale Value[1 x 1]] }
0x7f427f20ea48: {[features Value[28 x 28 x 1 x *1]] }
0x7f427f4d37a8: {[conv1.w.W Value[16 x 25]] }
0x7f427f4d3968: {[conv2.b.b Value[1 x 1 x 32]] }
0x7f427f4e2108: {[ce Value[1]] }
0x7f427f4fcea8: {[h1.t Value[128 x *1]] }
0x7f427f4fd068: {[h1.z Value[128 x 1 x *1]] }
0x7f427f4fd228: {[h1.y Value[128 x 1 x *1]] }
0x7f427f4fd3e8: {[ol.t Value[10 x 1 x *1]] }
0x7f427f4fd5a8: {[ol.z Value[10 x 1 x *1]] }
05/13/2016 15:10:28: Final Results: Minibatch[1-625]: err = 0.01460000 * 10000; ce = 0.04549626 * 10000; perplexity = 1.04654709
05/13/2016 15:10:28: Action "test" complete.
05/13/2016 15:10:28: __COMPLETED__

522
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution/baseline.windows.txt Normal file
Просмотреть файл

@ -0,0 +1,522 @@
=== Running /cygdrive/c/jenkins/workspace/CNTK-Test-Windows-W1/x64/release/cntk.exe configFile=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/Config/02_Convolution.cntk currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu\TestData RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu\TestData ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu DeviceId=0 timestamping=true train=[SGD=[maxEpochs=3]] imageLayout="cudnn"
-------------------------------------------------------------------
Build info:
Built time: May 13 2016 08:06:01
Last modified date: Thu May 12 07:31:50 2016
Build type: Release
Build target: GPU
With 1bit-SGD: no
CUDA_PATH: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5
CUB_PATH: c:\src\cub-1.4.1
CUDNN_PATH: c:\NVIDIA\cudnn-4.0\cuda
Build Branch: HEAD
Build SHA1: 35fadc316f045d843bbd9b85061250a959268787
Built by svcphil on Philly-Pool3
Build Path: c:\Jenkins\workspace\CNTK-Build-Windows\Source\CNTK\
-------------------------------------------------------------------
Changed current directory to C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu\TestData
05/13/2016 08:16:16: -------------------------------------------------------------------
05/13/2016 08:16:16: Build info:
05/13/2016 08:16:16: Built time: May 13 2016 08:06:01
05/13/2016 08:16:16: Last modified date: Thu May 12 07:31:50 2016
05/13/2016 08:16:16: Build type: Release
05/13/2016 08:16:16: Build target: GPU
05/13/2016 08:16:16: With 1bit-SGD: no
05/13/2016 08:16:16: CUDA_PATH: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5
05/13/2016 08:16:16: CUB_PATH: c:\src\cub-1.4.1
05/13/2016 08:16:16: CUDNN_PATH: c:\NVIDIA\cudnn-4.0\cuda
05/13/2016 08:16:16: Build Branch: HEAD
05/13/2016 08:16:16: Build SHA1: 35fadc316f045d843bbd9b85061250a959268787
05/13/2016 08:16:16: Built by svcphil on Philly-Pool3
05/13/2016 08:16:16: Build Path: c:\Jenkins\workspace\CNTK-Build-Windows\Source\CNTK\
05/13/2016 08:16:16: -------------------------------------------------------------------
05/13/2016 08:16:16: Running on Philly-Pool2 at 2016/05/13 08:16:16
05/13/2016 08:16:16: Command line:
C:\jenkins\workspace\CNTK-Test-Windows-W1\x64\release\cntk.exe configFile=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/Config/02_Convolution.cntk currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu\TestData RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu\TestData ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu DeviceId=0 timestamping=true train=[SGD=[maxEpochs=3]] imageLayout="cudnn"
05/13/2016 08:16:16: >>>>>>>>>>>>>>>>>>>> RAW CONFIG (VARIABLES NOT RESOLVED) >>>>>>>>>>>>>>>>>>>>
05/13/2016 08:16:16: RootDir = ".."
ConfigDir = "$RootDir$/Config"
DataDir = "$RootDir$/Data"
OutputDir = "$RootDir$/Output"
ModelDir = "$OutputDir$/Models"
deviceId = 0
imageLayout = "cudnn"
command = train:test
precision = "float"
modelPath = "$ModelDir$/02_Convolution"
ndlMacros = "$ConfigDir$/Macros.ndl"
traceLevel=1
numMBsToShowResult=500
prefetch=true
initOnCPUOnly=true
train = [
action = "train"
NDLNetworkBuilder = [
networkDescription = "$ConfigDir$/02_Convolution.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.1*5:0.3
momentumPerMB = 0*10:0.7
maxEpochs = 15
]
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
test = [
action = test
minibatchSize = 16
NDLNetworkBuilder = [
networkDescription = "$ConfigDir$/02_Convolution.ndl"
]
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu\TestData
RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu
DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu\TestData
ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config
OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu
DeviceId=0
timestamping=true
train=[SGD=[maxEpochs=3]]
imageLayout="cudnn"
05/13/2016 08:16:16: <<<<<<<<<<<<<<<<<<<< RAW CONFIG (VARIABLES NOT RESOLVED) <<<<<<<<<<<<<<<<<<<<
05/13/2016 08:16:16: >>>>>>>>>>>>>>>>>>>> RAW CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>
05/13/2016 08:16:16: RootDir = ".."
ConfigDir = "../Config"
DataDir = "../Data"
OutputDir = "../Output"
ModelDir = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu/Models"
deviceId = 0
imageLayout = "cudnn"
command = train:test
precision = "float"
modelPath = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu/Models/02_Convolution"
ndlMacros = "C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/Macros.ndl"
traceLevel=1
numMBsToShowResult=500
prefetch=true
initOnCPUOnly=true
train = [
action = "train"
NDLNetworkBuilder = [
networkDescription = "C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/02_Convolution.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.1*5:0.3
momentumPerMB = 0*10:0.7
maxEpochs = 15
]
reader = [
readerType = "CNTKTextFormatReader"
file = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu\TestData/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
test = [
action = test
minibatchSize = 16
NDLNetworkBuilder = [
networkDescription = "C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/02_Convolution.ndl"
]
reader = [
readerType = "CNTKTextFormatReader"
file = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu\TestData/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu\TestData
RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu
DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu\TestData
ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config
OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu
DeviceId=0
timestamping=true
train=[SGD=[maxEpochs=3]]
imageLayout="cudnn"
05/13/2016 08:16:16: <<<<<<<<<<<<<<<<<<<< RAW CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<
05/13/2016 08:16:16: >>>>>>>>>>>>>>>>>>>> PROCESSED CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>
configparameters: 02_Convolution.cntk:command=train:test
configparameters: 02_Convolution.cntk:ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config
configparameters: 02_Convolution.cntk:currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu\TestData
configparameters: 02_Convolution.cntk:DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu\TestData
configparameters: 02_Convolution.cntk:deviceId=0
configparameters: 02_Convolution.cntk:imageLayout=cudnn
configparameters: 02_Convolution.cntk:initOnCPUOnly=true
configparameters: 02_Convolution.cntk:ModelDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu/Models
configparameters: 02_Convolution.cntk:modelPath=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu/Models/02_Convolution
configparameters: 02_Convolution.cntk:ndlMacros=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/Macros.ndl
configparameters: 02_Convolution.cntk:numMBsToShowResult=500
configparameters: 02_Convolution.cntk:OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu
configparameters: 02_Convolution.cntk:precision=float
configparameters: 02_Convolution.cntk:prefetch=true
configparameters: 02_Convolution.cntk:RootDir=..
configparameters: 02_Convolution.cntk:RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu
configparameters: 02_Convolution.cntk:test=[
action = test
minibatchSize = 16
NDLNetworkBuilder = [
networkDescription = "C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/02_Convolution.ndl"
]
reader = [
readerType = "CNTKTextFormatReader"
file = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu\TestData/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
configparameters: 02_Convolution.cntk:timestamping=true
configparameters: 02_Convolution.cntk:traceLevel=1
configparameters: 02_Convolution.cntk:train=[
action = "train"
NDLNetworkBuilder = [
networkDescription = "C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/02_Convolution.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.1*5:0.3
momentumPerMB = 0*10:0.7
maxEpochs = 15
]
reader = [
readerType = "CNTKTextFormatReader"
file = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu\TestData/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
] [SGD=[maxEpochs=3]]
05/13/2016 08:16:16: <<<<<<<<<<<<<<<<<<<< PROCESSED CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<
05/13/2016 08:16:16: Commands: train test
05/13/2016 08:16:16: Precision = "float"
05/13/2016 08:16:16: CNTKModelPath: C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu/Models/02_Convolution
05/13/2016 08:16:16: CNTKCommandTrainInfo: train : 3
05/13/2016 08:16:16: CNTKCommandTrainInfo: CNTKNoMoreCommands_Total : 3
05/13/2016 08:16:16: ##############################################################################
05/13/2016 08:16:16: # #
05/13/2016 08:16:16: # Action "train" #
05/13/2016 08:16:16: # #
05/13/2016 08:16:16: ##############################################################################
05/13/2016 08:16:16: CNTKCommandTrainBegin: train
NDLBuilder Using GPU 0
05/13/2016 08:16:16: Creating virgin network.
Post-processing network...
3 roots:
ce = CrossEntropyWithSoftmax()
err = ErrorPrediction()
ol.z = Plus()
Validating network. 27 nodes to process in pass 1.
Validating --> labels = InputValue() : -> [10 x *]
Validating --> ol.W = LearnableParameter() : -> [10 x 128]
Validating --> h1.W = LearnableParameter() : -> [128 x 7 x 7 x 32]
Validating --> conv2.w.W = LearnableParameter() : -> [32 x 400]
Validating --> conv1.w.W = LearnableParameter() : -> [16 x 25]
Validating --> featScale = LearnableParameter() : -> [1 x 1]
Validating --> features = InputValue() : -> [28 x 28 x 1 x *]
Validating --> featScaled = ElementTimes (featScale, features) : [1 x 1], [28 x 28 x 1 x *] -> [28 x 28 x 1 x *]
Validating --> conv1.c.c = Convolution (conv1.w.W, featScaled) : [16 x 25], [28 x 28 x 1 x *] -> [28 x 28 x 16 x *]
Validating --> conv1.b.b = LearnableParameter() : -> [1 x 1 x 16]
Validating --> conv1.cpb = Plus (conv1.c.c, conv1.b.b) : [28 x 28 x 16 x *], [1 x 1 x 16] -> [28 x 28 x 16 x *]
Validating --> conv1.out = RectifiedLinear (conv1.cpb) : [28 x 28 x 16 x *] -> [28 x 28 x 16 x *]
Validating --> pool1 = MaxPooling (conv1.out) : [28 x 28 x 16 x *] -> [14 x 14 x 16 x *]
Validating --> conv2.c.c = Convolution (conv2.w.W, pool1) : [32 x 400], [14 x 14 x 16 x *] -> [14 x 14 x 32 x *]
Validating --> conv2.b.b = LearnableParameter() : -> [1 x 1 x 32]
Validating --> conv2.cpb = Plus (conv2.c.c, conv2.b.b) : [14 x 14 x 32 x *], [1 x 1 x 32] -> [14 x 14 x 32 x *]
Validating --> conv2.out = RectifiedLinear (conv2.cpb) : [14 x 14 x 32 x *] -> [14 x 14 x 32 x *]
Validating --> pool2.p = Pooling (conv2.out) : [14 x 14 x 32 x *] -> [7 x 7 x 32 x *]
Validating --> h1.t = Times (h1.W, pool2.p) : [128 x 7 x 7 x 32], [7 x 7 x 32 x *] -> [128 x *]
Validating --> h1.b = LearnableParameter() : -> [128 x 1]
Validating --> h1.z = Plus (h1.t, h1.b) : [128 x *], [128 x 1] -> [128 x 1 x *]
Validating --> h1.y = Sigmoid (h1.z) : [128 x 1 x *] -> [128 x 1 x *]
Validating --> ol.t = Times (ol.W, h1.y) : [10 x 128], [128 x 1 x *] -> [10 x 1 x *]
Validating --> ol.b = LearnableParameter() : -> [10 x 1]
Validating --> ol.z = Plus (ol.t, ol.b) : [10 x 1 x *], [10 x 1] -> [10 x 1 x *]
Validating --> ce = CrossEntropyWithSoftmax (labels, ol.z) : [10 x *], [10 x 1 x *] -> [1]
Validating --> err = ErrorPrediction (labels, ol.z) : [10 x *], [10 x 1 x *] -> [1]
Validating network. 16 nodes to process in pass 2.
Validating network, final pass.
Using cuDNN convolution engine for geometry: Input: 28 x 28 x 1, Output: 28 x 28 x 16, Kernel: 5 x 5 x 1, Map: 1 x 1 x 16, Stride: 1 x 1 x 1, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 28 x 28 x 16, Output: 14 x 14 x 16, Kernel: 2 x 2 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 14 x 14 x 16, Output: 14 x 14 x 32, Kernel: 5 x 5 x 16, Map: 32, Stride: 1 x 1 x 16, Sharing: (1, 1, 1), AutoPad: (1, 1, 0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 14 x 14 x 32, Output: 7 x 7 x 32, Kernel: 2 x 2 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (1, 1, 0), LowerPad: 0, UpperPad: 0.
11 out of 27 nodes do not share the minibatch layout with the input data.
Post-processing network complete.
05/13/2016 08:16:18: Created model with 27 nodes on GPU 0.
05/13/2016 08:16:18: Training criterion node(s):
05/13/2016 08:16:18: ce = CrossEntropyWithSoftmax
05/13/2016 08:16:18: Evaluation criterion node(s):
05/13/2016 08:16:18: err = ErrorPrediction
Allocating matrices for forward and/or backward propagation.
Memory Sharing Structure:
0000000000000000: {[err Gradient[1]] [featScale Gradient[1 x 1]] [featScaled Gradient[28 x 28 x 1 x *]] [features Gradient[28 x 28 x 1 x *]] [labels Gradient[10 x *]] }
000000CB919F83E0: {[features Value[28 x 28 x 1 x *]] }
000000CBAA188380: {[h1.W Value[128 x 7 x 7 x 32]] }
000000CBAA188560: {[ol.W Value[10 x 128]] }
000000CBAA1887E0: {[ol.b Value[10 x 1]] }
000000CBAA188A60: {[featScale Value[1 x 1]] }
000000CBAA188CE0: {[conv1.w.W Value[16 x 25]] }
000000CBAA1890A0: {[labels Value[10 x *]] }
000000CBAA189320: {[conv2.b.b Value[1 x 1 x 32]] }
000000CBAA1893C0: {[conv2.w.W Value[32 x 400]] }
000000CBAA189C80: {[conv1.b.b Value[1 x 1 x 16]] }
000000CBAA189DC0: {[h1.b Value[128 x 1]] }
000000CBB0834910: {[ol.z Value[10 x 1 x *]] }
000000CBB0834AF0: {[err Value[1]] }
000000CBB0834B90: {[ol.t Gradient[10 x 1 x *]] }
000000CBB0834F50: {[conv1.c.c Gradient[28 x 28 x 16 x *]] [conv1.out Value[28 x 28 x 16 x *]] }
000000CBB0834FF0: {[conv1.c.c Value[28 x 28 x 16 x *]] }
000000CBB08353B0: {[featScaled Value[28 x 28 x 1 x *]] }
000000CBB0835770: {[ce Value[1]] }
000000CBB0835950: {[conv2.c.c Value[14 x 14 x 32 x *]] }
000000CBB0835B30: {[conv2.b.b Gradient[1 x 1 x 32]] [conv2.out Gradient[14 x 14 x 32 x *]] [h1.t Value[128 x *]] }
000000CBB0835BD0: {[h1.W Gradient[128 x 7 x 7 x 32]] [h1.z Value[128 x 1 x *]] }
000000CBB0835C70: {[h1.t Gradient[128 x *]] [h1.y Value[128 x 1 x *]] }
000000CBB0835DB0: {[conv2.cpb Gradient[14 x 14 x 32 x *]] [pool1 Gradient[14 x 14 x 16 x *]] [pool2.p Value[7 x 7 x 32 x *]] }
000000CBB0835F90: {[ce Gradient[1]] }
000000CBB0836350: {[conv1.cpb Value[28 x 28 x 16 x *]] [conv1.w.W Gradient[16 x 25]] }
000000CBB08363F0: {[conv1.b.b Gradient[1 x 1 x 16]] [conv1.out Gradient[28 x 28 x 16 x *]] }
000000CBB0836490: {[h1.z Gradient[128 x 1 x *]] [ol.t Value[10 x 1 x *]] [pool2.p Gradient[7 x 7 x 32 x *]] }
000000CBB0836670: {[ol.b Gradient[10 x 1]] }
000000CBB0836990: {[conv2.c.c Gradient[14 x 14 x 32 x *]] [conv2.out Value[14 x 14 x 32 x *]] }
000000CBB0836A30: {[ol.W Gradient[10 x 128]] [ol.z Gradient[10 x 1 x *]] }
000000CBB0836B70: {[conv2.cpb Value[14 x 14 x 32 x *]] [conv2.w.W Gradient[32 x 400]] }
000000CBB0836CB0: {[h1.b Gradient[128 x 1]] [h1.y Gradient[128 x 1 x *]] }
000000CBB0836E90: {[conv1.cpb Gradient[28 x 28 x 16 x *]] [pool1 Value[14 x 14 x 16 x *]] }
05/13/2016 08:16:18: No PreCompute nodes found, skipping PreCompute step.
05/13/2016 08:16:18: Starting Epoch 1: learning rate per sample = 0.003125 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 08:16:18: Starting minibatch loop.
05/13/2016 08:16:22: Epoch[ 1 of 3]-Minibatch[ 1- 500, 26.67%]: ce = 1.52245886 * 16000; err = 0.53531250 * 16000; time = 4.2213s; samplesPerSecond = 3790.3
05/13/2016 08:16:24: Epoch[ 1 of 3]-Minibatch[ 501-1000, 53.33%]: ce = 0.20213049 * 16000; err = 0.05737500 * 16000; time = 1.6650s; samplesPerSecond = 9609.8
05/13/2016 08:16:26: Epoch[ 1 of 3]-Minibatch[1001-1500, 80.00%]: ce = 0.11822998 * 16000; err = 0.03400000 * 16000; time = 1.6662s; samplesPerSecond = 9602.5
05/13/2016 08:16:27: Finished Epoch[ 1 of 3]: [Training] ce = 0.51029333 * 60000; err = 0.17250000 * 60000; totalSamplesSeen = 60000; learningRatePerSample = 0.003125; epochTime=8.83729s
05/13/2016 08:16:27: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu/Models/02_Convolution.1'
05/13/2016 08:16:27: Starting Epoch 2: learning rate per sample = 0.003125 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 08:16:27: Starting minibatch loop.
05/13/2016 08:16:29: Epoch[ 2 of 3]-Minibatch[ 1- 500, 26.67%]: ce = 0.07765988 * 16000; err = 0.02281250 * 16000; time = 1.6655s; samplesPerSecond = 9606.6
05/13/2016 08:16:30: Epoch[ 2 of 3]-Minibatch[ 501-1000, 53.33%]: ce = 0.06650398 * 16000; err = 0.01943750 * 16000; time = 1.6661s; samplesPerSecond = 9603.4
05/13/2016 08:16:32: Epoch[ 2 of 3]-Minibatch[1001-1500, 80.00%]: ce = 0.06597272 * 16000; err = 0.02025000 * 16000; time = 1.6655s; samplesPerSecond = 9607.0
05/13/2016 08:16:33: Finished Epoch[ 2 of 3]: [Training] ce = 0.06707618 * 60000; err = 0.01993333 * 60000; totalSamplesSeen = 120000; learningRatePerSample = 0.003125; epochTime=6.26303s
05/13/2016 08:16:33: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu/Models/02_Convolution.2'
05/13/2016 08:16:33: Starting Epoch 3: learning rate per sample = 0.003125 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 08:16:33: Starting minibatch loop.
05/13/2016 08:16:35: Epoch[ 3 of 3]-Minibatch[ 1- 500, 26.67%]: ce = 0.04900096 * 16000; err = 0.01531250 * 16000; time = 1.6660s; samplesPerSecond = 9603.7
05/13/2016 08:16:37: Epoch[ 3 of 3]-Minibatch[ 501-1000, 53.33%]: ce = 0.04317124 * 16000; err = 0.01300000 * 16000; time = 1.6655s; samplesPerSecond = 9606.5
05/13/2016 08:16:38: Epoch[ 3 of 3]-Minibatch[1001-1500, 80.00%]: ce = 0.04517576 * 16000; err = 0.01293750 * 16000; time = 1.6628s; samplesPerSecond = 9622.2
05/13/2016 08:16:40: Finished Epoch[ 3 of 3]: [Training] ce = 0.04463579 * 60000; err = 0.01335000 * 60000; totalSamplesSeen = 180000; learningRatePerSample = 0.003125; epochTime=6.25721s
05/13/2016 08:16:40: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_02_Convolution@release_gpu/Models/02_Convolution'
05/13/2016 08:16:40: CNTKCommandTrainEnd: train
05/13/2016 08:16:40: Action "train" complete.
05/13/2016 08:16:40: ##############################################################################
05/13/2016 08:16:40: # #
05/13/2016 08:16:40: # Action "test" #
05/13/2016 08:16:40: # #
05/13/2016 08:16:40: ##############################################################################
Post-processing network...
3 roots:
ce = CrossEntropyWithSoftmax()
err = ErrorPrediction()
ol.z = Plus()
Validating network. 27 nodes to process in pass 1.
Validating --> labels = InputValue() : -> [10 x *1]
Validating --> ol.W = LearnableParameter() : -> [10 x 128]
Validating --> h1.W = LearnableParameter() : -> [128 x 7 x 7 x 32]
Validating --> conv2.w.W = LearnableParameter() : -> [32 x 400]
Validating --> conv1.w.W = LearnableParameter() : -> [16 x 25]
Validating --> featScale = LearnableParameter() : -> [1 x 1]
Validating --> features = InputValue() : -> [28 x 28 x 1 x *1]
Validating --> featScaled = ElementTimes (featScale, features) : [1 x 1], [28 x 28 x 1 x *1] -> [28 x 28 x 1 x *1]
Validating --> conv1.c.c = Convolution (conv1.w.W, featScaled) : [16 x 25], [28 x 28 x 1 x *1] -> [28 x 28 x 16 x *1]
Validating --> conv1.b.b = LearnableParameter() : -> [1 x 1 x 16]
Validating --> conv1.cpb = Plus (conv1.c.c, conv1.b.b) : [28 x 28 x 16 x *1], [1 x 1 x 16] -> [28 x 28 x 16 x *1]
Validating --> conv1.out = RectifiedLinear (conv1.cpb) : [28 x 28 x 16 x *1] -> [28 x 28 x 16 x *1]
Validating --> pool1 = MaxPooling (conv1.out) : [28 x 28 x 16 x *1] -> [14 x 14 x 16 x *1]
Validating --> conv2.c.c = Convolution (conv2.w.W, pool1) : [32 x 400], [14 x 14 x 16 x *1] -> [14 x 14 x 32 x *1]
Validating --> conv2.b.b = LearnableParameter() : -> [1 x 1 x 32]
Validating --> conv2.cpb = Plus (conv2.c.c, conv2.b.b) : [14 x 14 x 32 x *1], [1 x 1 x 32] -> [14 x 14 x 32 x *1]
Validating --> conv2.out = RectifiedLinear (conv2.cpb) : [14 x 14 x 32 x *1] -> [14 x 14 x 32 x *1]
Validating --> pool2.p = Pooling (conv2.out) : [14 x 14 x 32 x *1] -> [7 x 7 x 32 x *1]
Validating --> h1.t = Times (h1.W, pool2.p) : [128 x 7 x 7 x 32], [7 x 7 x 32 x *1] -> [128 x *1]
Validating --> h1.b = LearnableParameter() : -> [128 x 1]
Validating --> h1.z = Plus (h1.t, h1.b) : [128 x *1], [128 x 1] -> [128 x 1 x *1]
Validating --> h1.y = Sigmoid (h1.z) : [128 x 1 x *1] -> [128 x 1 x *1]
Validating --> ol.t = Times (ol.W, h1.y) : [10 x 128], [128 x 1 x *1] -> [10 x 1 x *1]
Validating --> ol.b = LearnableParameter() : -> [10 x 1]
Validating --> ol.z = Plus (ol.t, ol.b) : [10 x 1 x *1], [10 x 1] -> [10 x 1 x *1]
Validating --> ce = CrossEntropyWithSoftmax (labels, ol.z) : [10 x *1], [10 x 1 x *1] -> [1]
Validating --> err = ErrorPrediction (labels, ol.z) : [10 x *1], [10 x 1 x *1] -> [1]
Validating network. 16 nodes to process in pass 2.
Validating network, final pass.
Using cuDNN convolution engine for geometry: Input: 28 x 28 x 1, Output: 28 x 28 x 16, Kernel: 5 x 5 x 1, Map: 1 x 1 x 16, Stride: 1 x 1 x 1, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 28 x 28 x 16, Output: 14 x 14 x 16, Kernel: 2 x 2 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 14 x 14 x 16, Output: 14 x 14 x 32, Kernel: 5 x 5 x 16, Map: 32, Stride: 1 x 1 x 16, Sharing: (1, 1, 1), AutoPad: (1, 1, 0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 14 x 14 x 32, Output: 7 x 7 x 32, Kernel: 2 x 2 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (1, 1, 0), LowerPad: 0, UpperPad: 0.
11 out of 27 nodes do not share the minibatch layout with the input data.
Post-processing network complete.
evalNodeNames are not specified, using all the default evalnodes and training criterion nodes.
Allocating matrices for forward and/or backward propagation.
Memory Sharing Structure:
0000000000000000: {[ce Gradient[1]] [conv1.b.b Gradient[1 x 1 x 16]] [conv1.c.c Gradient[28 x 28 x 16 x *1]] [conv1.cpb Gradient[28 x 28 x 16 x *1]] [conv1.out Gradient[28 x 28 x 16 x *1]] [conv1.w.W Gradient[16 x 25]] [conv2.b.b Gradient[1 x 1 x 32]] [conv2.c.c Gradient[14 x 14 x 32 x *1]] [conv2.cpb Gradient[14 x 14 x 32 x *1]] [conv2.out Gradient[14 x 14 x 32 x *1]] [conv2.w.W Gradient[32 x 400]] [err Gradient[1]] [featScale Gradient[1 x 1]] [featScaled Gradient[28 x 28 x 1 x *1]] [features Gradient[28 x 28 x 1 x *1]] [h1.W Gradient[128 x 7 x 7 x 32]] [h1.b Gradient[128 x 1]] [h1.t Gradient[128 x *1]] [h1.y Gradient[128 x 1 x *1]] [h1.z Gradient[128 x 1 x *1]] [labels Gradient[10 x *1]] [ol.W Gradient[10 x 128]] [ol.b Gradient[10 x 1]] [ol.t Gradient[10 x 1 x *1]] [ol.z Gradient[10 x 1 x *1]] [pool1 Gradient[14 x 14 x 16 x *1]] [pool2.p Gradient[7 x 7 x 32 x *1]] }
000000CBAA188420: {[conv2.cpb Value[14 x 14 x 32 x *1]] }
000000CBAA188BA0: {[pool2.p Value[7 x 7 x 32 x *1]] }
000000CBAA1890A0: {[conv1.out Value[28 x 28 x 16 x *1]] }
000000CBAA189140: {[conv2.c.c Value[14 x 14 x 32 x *1]] }
000000CBAA1891E0: {[h1.t Value[128 x *1]] }
000000CBAA189320: {[h1.z Value[128 x 1 x *1]] }
000000CBAA1895A0: {[ol.t Value[10 x 1 x *1]] }
000000CBAA189780: {[pool1 Value[14 x 14 x 16 x *1]] }
000000CBAA189820: {[ol.z Value[10 x 1 x *1]] }
000000CBAA189DC0: {[h1.y Value[128 x 1 x *1]] }
000000CBAA18A0E0: {[conv2.out Value[14 x 14 x 32 x *1]] }
000000CBB0834AF0: {[ol.W Value[10 x 128]] }
000000CBB0834C30: {[h1.b Value[128 x 1]] }
000000CBB0834FF0: {[features Value[28 x 28 x 1 x *1]] }
000000CBB0835770: {[h1.W Value[128 x 7 x 7 x 32]] }
000000CBB08358B0: {[featScale Value[1 x 1]] }
000000CBB0835BD0: {[conv1.w.W Value[16 x 25]] }
000000CBB08360D0: {[labels Value[10 x *1]] }
000000CBB0836350: {[ol.b Value[10 x 1]] }
000000CBB0836490: {[conv2.b.b Value[1 x 1 x 32]] }
000000CBB0836A30: {[conv2.w.W Value[32 x 400]] }
000000CBB0836CB0: {[conv1.b.b Value[1 x 1 x 16]] }
000000CBB08371B0: {[err Value[1]] }
000000CBB08372F0: {[conv1.cpb Value[28 x 28 x 16 x *1]] }
000000CBB0837F70: {[featScaled Value[28 x 28 x 1 x *1]] }
000000CBB08381F0: {[ce Value[1]] }
000000CBB08383D0: {[conv1.c.c Value[28 x 28 x 16 x *1]] }
05/13/2016 08:16:51: Final Results: Minibatch[1-625]: err = 0.01520000 * 10000; ce = 0.04488435 * 10000; perplexity = 1.04590689
05/13/2016 08:16:51: Action "test" complete.
05/13/2016 08:16:51: __COMPLETED__

11
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution/run-test Executable file
Просмотреть файл

@ -0,0 +1,11 @@
#!/bin/bash
. $TEST_DIR/../run-test-common
cntkrun $CNTKTextConfigDir/02_Convolution.cntk "train=[SGD=[maxEpochs=3]] imageLayout=\"$imageLayout\""
ExitCode=$?
# Delete the test data if copied
[[ "$Copied" -eq "1" ]] && rm -rf "$DataDir"
exit $ExitCode

41
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/02_Convolution/testcases.yml Normal file
Просмотреть файл

@ -0,0 +1,41 @@
dataDir: ../../../../../../../Examples/Image/MNIST/Data
tags:
# In BVT, run Release GPU
- bvt-e (build_sku=='gpu') and (device=='gpu') and (flavor=='release')
# In Nightly on Linux, run Debug GPU in addition.
- nightly-e (build_sku=='gpu') and (device=='gpu')
testCases:
CNTK Run must be completed:
patterns:
- __COMPLETED__
Must train epochs in exactly same order and parameters:
patterns:
- Starting Epoch {{integer}}
- learning rate per sample = {{float}}
- momentum = {{float}}
# TODO destabilized by integration of 16b606c
# Epochs (with low train loss) must be finished with expected results:
# patterns:
# - Finished Epoch[{{integer}} of {{integer}}]
# - ce = {{float,tolerance=0.05}} * {{integer}}
# - err = {{float,tolerance=0.005}} * {{integer}}
# - totalSamplesSeen = {{integer}}
# - learningRatePerSample = {{float,tolerance=0.1%}}
Per-minibatch (with low train loss) training results must match:
patterns:
# Ignores first set of minibatches at start of epoch, which sometimes has a larger deviation:
- 01-
- Epoch[{{integer}} of {{integer}}]-Minibatch[{{integer}}-{{integer}}
- ce = {{float,tolerance=0.05}} * {{integer}}
- err = {{float,tolerance=0.005}} * {{integer}}
Final test results must match:
patterns:
- "Final Results: Minibatch[{{integer}}-{{integer}}]"
- err = {{float,tolerance=0.005}} * {{integer}}
- ce = {{float,tolerance=0.05}} * {{integer}}

9
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm/README.txt Normal file
Просмотреть файл

@ -0,0 +1,9 @@
Test runtimes
Windows:
Running test CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm (debug gpu) - [OK] 214.39 sec
Running test CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm (release gpu) - [OK] 64.81 sec
Linux:
Running test CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm (debug gpu) - [OK] 44.82 sec
Running test CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm (release gpu) - [OK] 19.90 sec

581
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm/baseline.linux.txt Normal file
Просмотреть файл

@ -0,0 +1,581 @@
=== Running /home/philly/jenkins/workspace/CNTK-Test-Linux-W1/build/gpu/release/bin/cntk configFile=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm/../../../../../../../Examples/Image/MNIST/Config/../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/Config/03_ConvBatchNorm.cntk currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/TestData RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/TestData ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm/../../../../../../../Examples/Image/MNIST/Config OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu DeviceId=0 timestamping=true train=[SGD=[maxEpochs=3]] imageLayout="cudnn"
-------------------------------------------------------------------
Build info:
Built time: May 13 2016 14:50:25
Last modified date: Thu May 12 14:00:37 2016
Build type: release
Build target: GPU
With 1bit-SGD: no
Math lib: acml
CUDA_PATH: /usr/local/cuda-7.5
CUB_PATH: /usr/local/cub-1.4.1
CUDNN_PATH: /usr/local/cudnn-4.0
Build Branch: HEAD
Build SHA1: 35fadc316f045d843bbd9b85061250a959268787
Built by philly on d8dc82703b0f
Build Path: /home/philly/jenkins/workspace/CNTK-Build-Linux
-------------------------------------------------------------------
Changed current directory to /tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/TestData
05/13/2016 15:10:29: -------------------------------------------------------------------
05/13/2016 15:10:29: Build info:
05/13/2016 15:10:29: Built time: May 13 2016 14:50:25
05/13/2016 15:10:29: Last modified date: Thu May 12 14:00:37 2016
05/13/2016 15:10:29: Build type: release
05/13/2016 15:10:29: Build target: GPU
05/13/2016 15:10:29: With 1bit-SGD: no
05/13/2016 15:10:29: Math lib: acml
05/13/2016 15:10:29: CUDA_PATH: /usr/local/cuda-7.5
05/13/2016 15:10:29: CUB_PATH: /usr/local/cub-1.4.1
05/13/2016 15:10:29: CUDNN_PATH: /usr/local/cudnn-4.0
05/13/2016 15:10:29: Build Branch: HEAD
05/13/2016 15:10:29: Build SHA1: 35fadc316f045d843bbd9b85061250a959268787
05/13/2016 15:10:29: Built by philly on d8dc82703b0f
05/13/2016 15:10:29: Build Path: /home/philly/jenkins/workspace/CNTK-Build-Linux
05/13/2016 15:10:29: -------------------------------------------------------------------
05/13/2016 15:10:29: Running on localhost at 2016/05/13 15:10:29
05/13/2016 15:10:29: Command line:
/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/build/gpu/release/bin/cntk configFile=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm/../../../../../../../Examples/Image/MNIST/Config/../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/Config/03_ConvBatchNorm.cntk currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/TestData RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/TestData ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm/../../../../../../../Examples/Image/MNIST/Config OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu DeviceId=0 timestamping=true train=[SGD=[maxEpochs=3]] imageLayout="cudnn"
05/13/2016 15:10:29: >>>>>>>>>>>>>>>>>>>> RAW CONFIG (VARIABLES NOT RESOLVED) >>>>>>>>>>>>>>>>>>>>
05/13/2016 15:10:29: RootDir = ".."
ConfigDir = "$RootDir$/Config"
DataDir = "$RootDir$/Data"
OutputDir = "$RootDir$/Output"
ModelDir = "$OutputDir$/Models"
deviceId = 0
imageLayout = "cudnn"
command = train:test
precision = "float"
modelPath = "$ModelDir$/03_ConvBatchNorm"
ndlMacros = "$ConfigDir$/Macros.ndl"
traceLevel=1
numMBsToShowResult=500
initOnCPUOnly=true
train = [
action = "train"
NDLNetworkBuilder = [
networkDescription = "$ConfigDir$/03_ConvBatchNorm.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.5:0.1
momentumPerMB = 0.9
maxEpochs = 2
batchNormalizationBlendTimeConstant=0:1#INF
]
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
test = [
action = "test"
minibatchSize = 32
modelPath=$ModelDir$/03_ConvBatchNorm
NDLNetworkBuilder = [
networkDescription = "$ConfigDir$/03_ConvBatchNorm.ndl"
]
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/TestData
RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu
DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/TestData
ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm/../../../../../../../Examples/Image/MNIST/Config
OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu
DeviceId=0
timestamping=true
train=[SGD=[maxEpochs=3]]
imageLayout="cudnn"
05/13/2016 15:10:29: <<<<<<<<<<<<<<<<<<<< RAW CONFIG (VARIABLES NOT RESOLVED) <<<<<<<<<<<<<<<<<<<<
05/13/2016 15:10:29: >>>>>>>>>>>>>>>>>>>> RAW CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>
05/13/2016 15:10:29: RootDir = ".."
ConfigDir = "../Config"
DataDir = "../Data"
OutputDir = "../Output"
ModelDir = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/Models"
deviceId = 0
imageLayout = "cudnn"
command = train:test
precision = "float"
modelPath = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/Models/03_ConvBatchNorm"
ndlMacros = "/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm/../../../../../../../Examples/Image/MNIST/Config/Macros.ndl"
traceLevel=1
numMBsToShowResult=500
initOnCPUOnly=true
train = [
action = "train"
NDLNetworkBuilder = [
networkDescription = "/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm/../../../../../../../Examples/Image/MNIST/Config/03_ConvBatchNorm.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.5:0.1
momentumPerMB = 0.9
maxEpochs = 2
batchNormalizationBlendTimeConstant=0:1#INF
]
reader = [
readerType = "CNTKTextFormatReader"
file = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/TestData/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
test = [
action = "test"
minibatchSize = 32
modelPath=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/Models/03_ConvBatchNorm
NDLNetworkBuilder = [
networkDescription = "/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm/../../../../../../../Examples/Image/MNIST/Config/03_ConvBatchNorm.ndl"
]
reader = [
readerType = "CNTKTextFormatReader"
file = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/TestData/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/TestData
RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu
DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/TestData
ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm/../../../../../../../Examples/Image/MNIST/Config
OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu
DeviceId=0
timestamping=true
train=[SGD=[maxEpochs=3]]
imageLayout="cudnn"
05/13/2016 15:10:29: <<<<<<<<<<<<<<<<<<<< RAW CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<
05/13/2016 15:10:29: >>>>>>>>>>>>>>>>>>>> PROCESSED CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>
configparameters: 03_ConvBatchNorm.cntk:command=train:test
configparameters: 03_ConvBatchNorm.cntk:ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm/../../../../../../../Examples/Image/MNIST/Config
configparameters: 03_ConvBatchNorm.cntk:currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/TestData
configparameters: 03_ConvBatchNorm.cntk:DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/TestData
configparameters: 03_ConvBatchNorm.cntk:deviceId=0
configparameters: 03_ConvBatchNorm.cntk:imageLayout=cudnn
configparameters: 03_ConvBatchNorm.cntk:initOnCPUOnly=true
configparameters: 03_ConvBatchNorm.cntk:ModelDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/Models
configparameters: 03_ConvBatchNorm.cntk:modelPath=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/Models/03_ConvBatchNorm
configparameters: 03_ConvBatchNorm.cntk:ndlMacros=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm/../../../../../../../Examples/Image/MNIST/Config/Macros.ndl
configparameters: 03_ConvBatchNorm.cntk:numMBsToShowResult=500
configparameters: 03_ConvBatchNorm.cntk:OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu
configparameters: 03_ConvBatchNorm.cntk:precision=float
configparameters: 03_ConvBatchNorm.cntk:RootDir=..
configparameters: 03_ConvBatchNorm.cntk:RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu
configparameters: 03_ConvBatchNorm.cntk:test=[
action = "test"
minibatchSize = 32
modelPath=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/Models/03_ConvBatchNorm
NDLNetworkBuilder = [
networkDescription = "/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm/../../../../../../../Examples/Image/MNIST/Config/03_ConvBatchNorm.ndl"
]
reader = [
readerType = "CNTKTextFormatReader"
file = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/TestData/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
configparameters: 03_ConvBatchNorm.cntk:timestamping=true
configparameters: 03_ConvBatchNorm.cntk:traceLevel=1
configparameters: 03_ConvBatchNorm.cntk:train=[
action = "train"
NDLNetworkBuilder = [
networkDescription = "/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm/../../../../../../../Examples/Image/MNIST/Config/03_ConvBatchNorm.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.5:0.1
momentumPerMB = 0.9
maxEpochs = 2
batchNormalizationBlendTimeConstant=0:1#INF
]
reader = [
readerType = "CNTKTextFormatReader"
file = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/TestData/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
] [SGD=[maxEpochs=3]]
05/13/2016 15:10:29: <<<<<<<<<<<<<<<<<<<< PROCESSED CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<
05/13/2016 15:10:29: Commands: train test
05/13/2016 15:10:29: Precision = "float"
05/13/2016 15:10:29: CNTKModelPath: /tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/Models/03_ConvBatchNorm
05/13/2016 15:10:29: CNTKCommandTrainInfo: train : 3
05/13/2016 15:10:29: CNTKCommandTrainInfo: CNTKNoMoreCommands_Total : 3
05/13/2016 15:10:29: ##############################################################################
05/13/2016 15:10:29: # #
05/13/2016 15:10:29: # Action "train" #
05/13/2016 15:10:29: # #
05/13/2016 15:10:29: ##############################################################################
05/13/2016 15:10:29: CNTKCommandTrainBegin: train
NDLBuilder Using GPU 0
05/13/2016 15:10:29: Creating virgin network.
Post-processing network...
3 roots:
ce = CrossEntropyWithSoftmax()
err = ErrorPrediction()
ol.z = Plus()
Validating network. 36 nodes to process in pass 1.
Validating --> labels = InputValue() : -> [10 x *]
Validating --> ol.W = LearnableParameter() : -> [10 x 128]
Validating --> h1.W = LearnableParameter() : -> [128 x 1568]
Validating --> conv2.c.W = LearnableParameter() : -> [32 x 400]
Validating --> conv1.c.W = LearnableParameter() : -> [16 x 25]
Validating --> featScale = LearnableParameter() : -> [1 x 1]
Validating --> features = InputValue() : -> [28 x 28 x 1 x *]
Validating --> featScaled = ElementTimes (featScale, features) : [1 x 1], [28 x 28 x 1 x *] -> [28 x 28 x 1 x *]
Validating --> conv1.c.c.c = Convolution (conv1.c.W, featScaled) : [16 x 25], [28 x 28 x 1 x *] -> [28 x 28 x 16 x *]
Validating --> conv1.c.c.sc = LearnableParameter() : -> [16 x 1]
Validating --> conv1.c.c.b = LearnableParameter() : -> [16 x 1]
Validating --> conv1.c.c.m = LearnableParameter() : -> [16 x 1]
Validating --> conv1.c.c.isd = LearnableParameter() : -> [16 x 1]
Validating --> conv1.c.c.y = BatchNormalization (conv1.c.c.c, conv1.c.c.sc, conv1.c.c.b, conv1.c.c.m, conv1.c.c.isd) : [28 x 28 x 16 x *], [16 x 1], [16 x 1], [16 x 1], [16 x 1] -> [28 x 28 x 16 x *]
Validating --> conv1.y = RectifiedLinear (conv1.c.c.y) : [28 x 28 x 16 x *] -> [28 x 28 x 16 x *]
Validating --> pool1 = MaxPooling (conv1.y) : [28 x 28 x 16 x *] -> [14 x 14 x 16 x *]
Validating --> conv2.c.c.c = Convolution (conv2.c.W, pool1) : [32 x 400], [14 x 14 x 16 x *] -> [14 x 14 x 32 x *]
Validating --> conv2.c.c.sc = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.b = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.m = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.isd = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.y = BatchNormalization (conv2.c.c.c, conv2.c.c.sc, conv2.c.c.b, conv2.c.c.m, conv2.c.c.isd) : [14 x 14 x 32 x *], [32 x 1], [32 x 1], [32 x 1], [32 x 1] -> [14 x 14 x 32 x *]
Validating --> conv2.y = RectifiedLinear (conv2.c.c.y) : [14 x 14 x 32 x *] -> [14 x 14 x 32 x *]
Validating --> pool2 = MaxPooling (conv2.y) : [14 x 14 x 32 x *] -> [7 x 7 x 32 x *]
h1.t Times operation: For legacy compatibility, the sample layout of left input (h1.W LearnableParameter operation) was patched to [128 x 7 x 7 x 32] (from [128 x 1568])
Validating --> h1.t = Times (h1.W, pool2) : [128 x 7 x 7 x 32], [7 x 7 x 32 x *] -> [128 x *]
Validating --> h1.sc = LearnableParameter() : -> [128 x 1]
Validating --> h1.b = LearnableParameter() : -> [128 x 1]
Validating --> h1.m = LearnableParameter() : -> [128 x 1]
Validating --> h1.isd = LearnableParameter() : -> [128 x 1]
Validating --> h1.bn = BatchNormalization (h1.t, h1.sc, h1.b, h1.m, h1.isd) : [128 x *], [128 x 1], [128 x 1], [128 x 1], [128 x 1] -> [128 x *]
Validating --> h1.y = RectifiedLinear (h1.bn) : [128 x *] -> [128 x *]
Validating --> ol.t = Times (ol.W, h1.y) : [10 x 128], [128 x *] -> [10 x *]
Validating --> ol.b = LearnableParameter() : -> [10 x 1]
Validating --> ol.z = Plus (ol.t, ol.b) : [10 x *], [10 x 1] -> [10 x 1 x *]
Validating --> ce = CrossEntropyWithSoftmax (labels, ol.z) : [10 x *], [10 x 1 x *] -> [1]
Validating --> err = ErrorPrediction (labels, ol.z) : [10 x *], [10 x 1 x *] -> [1]
Validating network. 16 nodes to process in pass 2.
Validating network, final pass.
Using cuDNN convolution engine for geometry: Input: 28 x 28 x 1, Output: 28 x 28 x 16, Kernel: 5 x 5 x 1, Map: 1 x 1 x 16, Stride: 1 x 1 x 1, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
Using cuDNN convolution engine for geometry: Input: 28 x 28 x 16, Output: 14 x 14 x 16, Kernel: 2 x 2 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 14 x 14 x 16, Output: 14 x 14 x 32, Kernel: 5 x 5 x 16, Map: 1 x 1 x 32, Stride: 1 x 1 x 16, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
Using cuDNN convolution engine for geometry: Input: 14 x 14 x 32, Output: 7 x 7 x 32, Kernel: 2 x 2 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
20 out of 36 nodes do not share the minibatch layout with the input data.
Post-processing network complete.
05/13/2016 15:10:29: Created model with 36 nodes on GPU 0.
05/13/2016 15:10:29: Training criterion node(s):
05/13/2016 15:10:29: ce = CrossEntropyWithSoftmax
05/13/2016 15:10:29: Evaluation criterion node(s):
05/13/2016 15:10:29: err = ErrorPrediction
Allocating matrices for forward and/or backward propagation.
Memory Sharing Structure:
(nil): {[conv1.c.c.isd Gradient[16 x 1]] [conv1.c.c.m Gradient[16 x 1]] [conv2.c.c.isd Gradient[32 x 1]] [conv2.c.c.m Gradient[32 x 1]] [err Gradient[1]] [featScale Gradient[1 x 1]] [featScaled Gradient[28 x 28 x 1 x *]] [features Gradient[28 x 28 x 1 x *]] [h1.isd Gradient[128 x 1]] [h1.m Gradient[128 x 1]] [labels Gradient[10 x *]] }
0x2643328: {[features Value[28 x 28 x 1 x *]] }
0x33a9468: {[featScale Value[1 x 1]] }
0x33aa5e8: {[labels Value[10 x *]] }
0x33ab128: {[conv1.c.W Value[16 x 25]] }
0x33ab818: {[conv1.c.c.b Value[16 x 1]] }
0x33ac238: {[conv1.c.c.sc Value[16 x 1]] }
0x33ad108: {[conv1.c.c.m Value[16 x 1]] }
0x33adbf8: {[conv1.c.c.isd Value[16 x 1]] }
0x33af968: {[conv2.c.W Value[32 x 400]] }
0x33b0878: {[conv2.c.c.b Value[32 x 1]] }
0x33b1258: {[conv2.c.c.sc Value[32 x 1]] }
0x33b1908: {[ol.b Value[10 x 1]] }
0x33b1e78: {[conv2.c.c.m Value[32 x 1]] }
0x33b29c8: {[conv2.c.c.isd Value[32 x 1]] }
0x33b3968: {[h1.W Value[128 x 7 x 7 x 32]] }
0x33b5408: {[h1.b Value[128 x 1]] }
0x33b5e38: {[h1.sc Value[128 x 1]] }
0x33b6738: {[h1.m Value[128 x 1]] }
0x33b70b8: {[h1.isd Value[128 x 1]] }
0x33b7618: {[ol.W Value[10 x 128]] }
0x33be778: {[ce Value[1]] }
0x33bfba8: {[ol.z Value[10 x 1 x *]] }
0x33ff558: {[err Value[1]] }
0x788fe48: {[conv1.c.c.c Value[28 x 28 x 16 x *]] }
0x7890188: {[featScaled Value[28 x 28 x 1 x *]] }
0x7890438: {[conv1.c.c.y Value[28 x 28 x 16 x *]] }
0x7891238: {[conv1.c.c.c Gradient[28 x 28 x 16 x *]] [conv1.y Value[28 x 28 x 16 x *]] }
0x78913f8: {[conv1.c.c.y Gradient[28 x 28 x 16 x *]] [pool1 Value[14 x 14 x 16 x *]] }
0x78915b8: {[conv1.c.W Gradient[16 x 25]] [conv2.c.c.c Value[14 x 14 x 32 x *]] }
0x7891778: {[conv1.c.c.sc Gradient[16 x 1]] [conv1.y Gradient[28 x 28 x 16 x *]] }
0x7891938: {[conv2.c.c.y Value[14 x 14 x 32 x *]] }
0x7891e78: {[conv1.c.c.b Gradient[16 x 1]] [conv2.c.c.c Gradient[14 x 14 x 32 x *]] [conv2.y Value[14 x 14 x 32 x *]] }
0x7892038: {[conv2.c.c.y Gradient[14 x 14 x 32 x *]] [pool2 Value[7 x 7 x 32 x *]] }
0x78921f8: {[conv2.c.c.sc Gradient[32 x 1]] [conv2.y Gradient[14 x 14 x 32 x *]] [h1.t Value[128 x *]] }
0x78923b8: {[h1.bn Value[128 x *]] }
0x7892738: {[conv2.c.c.b Gradient[32 x 1]] }
0x78928f8: {[conv2.c.W Gradient[32 x 400]] [h1.t Gradient[128 x *]] [h1.y Value[128 x *]] }
0x7892ab8: {[h1.bn Gradient[128 x *]] [ol.t Value[10 x *]] }
0x78999e8: {[ce Gradient[1]] }
0x7899ba8: {[ol.W Gradient[10 x 128]] [ol.z Gradient[10 x 1 x *]] }
0x7899d68: {[ol.t Gradient[10 x *]] [pool1 Gradient[14 x 14 x 16 x *]] [pool2 Gradient[7 x 7 x 32 x *]] }
0x7899f28: {[ol.b Gradient[10 x 1]] }
0x789a0e8: {[h1.sc Gradient[128 x 1]] [h1.y Gradient[128 x *]] }
0x789a2d8: {[h1.W Gradient[128 x 7 x 7 x 32]] }
0x789a498: {[h1.b Gradient[128 x 1]] }
05/13/2016 15:10:29: No PreCompute nodes found, skipping PreCompute step.
05/13/2016 15:10:29: Starting Epoch 1: learning rate per sample = 0.015625 effective momentum = 0.900000 momentum as time constant = 303.7 samples
05/13/2016 15:10:29: Starting minibatch loop.
05/13/2016 15:10:31: Epoch[ 1 of 3]-Minibatch[ 1- 500, 26.67%]: ce = 0.18369328 * 16000; err = 0.05750000 * 16000; time = 2.0641s; samplesPerSecond = 7751.5
05/13/2016 15:10:32: Epoch[ 1 of 3]-Minibatch[ 501-1000, 53.33%]: ce = 0.08101009 * 16000; err = 0.02425000 * 16000; time = 1.0283s; samplesPerSecond = 15560.4
05/13/2016 15:10:33: Epoch[ 1 of 3]-Minibatch[1001-1500, 80.00%]: ce = 0.06876743 * 16000; err = 0.02125000 * 16000; time = 1.0403s; samplesPerSecond = 15380.9
05/13/2016 15:10:34: Finished Epoch[ 1 of 3]: [Training] ce = 0.09983698 * 60000; err = 0.03098333 * 60000; totalSamplesSeen = 60000; learningRatePerSample = 0.015625; epochTime=4.9337s
05/13/2016 15:10:34: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/Models/03_ConvBatchNorm.1'
Setting batch normalization blend time constant to inf.
05/13/2016 15:10:34: Starting Epoch 2: learning rate per sample = 0.003125 effective momentum = 0.900000 momentum as time constant = 303.7 samples
05/13/2016 15:10:34: Starting minibatch loop.
05/13/2016 15:10:35: Epoch[ 2 of 3]-Minibatch[ 1- 500, 26.67%]: ce = 0.02224222 * 16000; err = 0.00756250 * 16000; time = 1.0463s; samplesPerSecond = 15292.5
05/13/2016 15:10:36: Epoch[ 2 of 3]-Minibatch[ 501-1000, 53.33%]: ce = 0.01788928 * 16000; err = 0.00568750 * 16000; time = 1.0489s; samplesPerSecond = 15254.3
05/13/2016 15:10:37: Epoch[ 2 of 3]-Minibatch[1001-1500, 80.00%]: ce = 0.01989119 * 16000; err = 0.00543750 * 16000; time = 1.0414s; samplesPerSecond = 15363.9
05/13/2016 15:10:38: Finished Epoch[ 2 of 3]: [Training] ce = 0.02009503 * 60000; err = 0.00623333 * 60000; totalSamplesSeen = 120000; learningRatePerSample = 0.003125; epochTime=3.92922s
05/13/2016 15:10:38: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/Models/03_ConvBatchNorm.2'
05/13/2016 15:10:38: Starting Epoch 3: learning rate per sample = 0.003125 effective momentum = 0.900000 momentum as time constant = 303.7 samples
05/13/2016 15:10:38: Starting minibatch loop.
05/13/2016 15:10:39: Epoch[ 3 of 3]-Minibatch[ 1- 500, 26.67%]: ce = 0.01173781 * 16000; err = 0.00306250 * 16000; time = 1.0390s; samplesPerSecond = 15400.0
05/13/2016 15:10:40: Epoch[ 3 of 3]-Minibatch[ 501-1000, 53.33%]: ce = 0.01463517 * 16000; err = 0.00431250 * 16000; time = 1.0397s; samplesPerSecond = 15388.4
05/13/2016 15:10:41: Epoch[ 3 of 3]-Minibatch[1001-1500, 80.00%]: ce = 0.01582476 * 16000; err = 0.00493750 * 16000; time = 1.0373s; samplesPerSecond = 15425.2
05/13/2016 15:10:42: Finished Epoch[ 3 of 3]: [Training] ce = 0.01382984 * 60000; err = 0.00401667 * 60000; totalSamplesSeen = 180000; learningRatePerSample = 0.003125; epochTime=3.9054s
05/13/2016 15:10:42: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/MNIST_03_ConvBatchNorm@release_gpu/Models/03_ConvBatchNorm'
05/13/2016 15:10:42: CNTKCommandTrainEnd: train
05/13/2016 15:10:42: Action "train" complete.
05/13/2016 15:10:42: ##############################################################################
05/13/2016 15:10:42: # #
05/13/2016 15:10:42: # Action "test" #
05/13/2016 15:10:42: # #
05/13/2016 15:10:42: ##############################################################################
Post-processing network...
3 roots:
ce = CrossEntropyWithSoftmax()
err = ErrorPrediction()
ol.z = Plus()
Validating network. 36 nodes to process in pass 1.
Validating --> labels = InputValue() : -> [10 x *1]
Validating --> ol.W = LearnableParameter() : -> [10 x 128]
Validating --> h1.W = LearnableParameter() : -> [128 x 7 x 7 x 32]
Validating --> conv2.c.W = LearnableParameter() : -> [32 x 400]
Validating --> conv1.c.W = LearnableParameter() : -> [16 x 25]
Validating --> featScale = LearnableParameter() : -> [1 x 1]
Validating --> features = InputValue() : -> [28 x 28 x 1 x *1]
Validating --> featScaled = ElementTimes (featScale, features) : [1 x 1], [28 x 28 x 1 x *1] -> [28 x 28 x 1 x *1]
Validating --> conv1.c.c.c = Convolution (conv1.c.W, featScaled) : [16 x 25], [28 x 28 x 1 x *1] -> [28 x 28 x 16 x *1]
Validating --> conv1.c.c.sc = LearnableParameter() : -> [16 x 1]
Validating --> conv1.c.c.b = LearnableParameter() : -> [16 x 1]
Validating --> conv1.c.c.m = LearnableParameter() : -> [16 x 1]
Validating --> conv1.c.c.isd = LearnableParameter() : -> [16 x 1]
Validating --> conv1.c.c.y = BatchNormalization (conv1.c.c.c, conv1.c.c.sc, conv1.c.c.b, conv1.c.c.m, conv1.c.c.isd) : [28 x 28 x 16 x *1], [16 x 1], [16 x 1], [16 x 1], [16 x 1] -> [28 x 28 x 16 x *1]
Validating --> conv1.y = RectifiedLinear (conv1.c.c.y) : [28 x 28 x 16 x *1] -> [28 x 28 x 16 x *1]
Validating --> pool1 = MaxPooling (conv1.y) : [28 x 28 x 16 x *1] -> [14 x 14 x 16 x *1]
Validating --> conv2.c.c.c = Convolution (conv2.c.W, pool1) : [32 x 400], [14 x 14 x 16 x *1] -> [14 x 14 x 32 x *1]
Validating --> conv2.c.c.sc = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.b = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.m = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.isd = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.y = BatchNormalization (conv2.c.c.c, conv2.c.c.sc, conv2.c.c.b, conv2.c.c.m, conv2.c.c.isd) : [14 x 14 x 32 x *1], [32 x 1], [32 x 1], [32 x 1], [32 x 1] -> [14 x 14 x 32 x *1]
Validating --> conv2.y = RectifiedLinear (conv2.c.c.y) : [14 x 14 x 32 x *1] -> [14 x 14 x 32 x *1]
Validating --> pool2 = MaxPooling (conv2.y) : [14 x 14 x 32 x *1] -> [7 x 7 x 32 x *1]
Validating --> h1.t = Times (h1.W, pool2) : [128 x 7 x 7 x 32], [7 x 7 x 32 x *1] -> [128 x *1]
Validating --> h1.sc = LearnableParameter() : -> [128 x 1]
Validating --> h1.b = LearnableParameter() : -> [128 x 1]
Validating --> h1.m = LearnableParameter() : -> [128 x 1]
Validating --> h1.isd = LearnableParameter() : -> [128 x 1]
Validating --> h1.bn = BatchNormalization (h1.t, h1.sc, h1.b, h1.m, h1.isd) : [128 x *1], [128 x 1], [128 x 1], [128 x 1], [128 x 1] -> [128 x *1]
Validating --> h1.y = RectifiedLinear (h1.bn) : [128 x *1] -> [128 x *1]
Validating --> ol.t = Times (ol.W, h1.y) : [10 x 128], [128 x *1] -> [10 x *1]
Validating --> ol.b = LearnableParameter() : -> [10 x 1]
Validating --> ol.z = Plus (ol.t, ol.b) : [10 x *1], [10 x 1] -> [10 x 1 x *1]
Validating --> ce = CrossEntropyWithSoftmax (labels, ol.z) : [10 x *1], [10 x 1 x *1] -> [1]
Validating --> err = ErrorPrediction (labels, ol.z) : [10 x *1], [10 x 1 x *1] -> [1]
Validating network. 16 nodes to process in pass 2.
Validating network, final pass.
Using cuDNN convolution engine for geometry: Input: 28 x 28 x 1, Output: 28 x 28 x 16, Kernel: 5 x 5 x 1, Map: 1 x 1 x 16, Stride: 1 x 1 x 1, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
Using cuDNN convolution engine for geometry: Input: 28 x 28 x 16, Output: 14 x 14 x 16, Kernel: 2 x 2 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 14 x 14 x 16, Output: 14 x 14 x 32, Kernel: 5 x 5 x 16, Map: 1 x 1 x 32, Stride: 1 x 1 x 16, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
Using cuDNN convolution engine for geometry: Input: 14 x 14 x 32, Output: 7 x 7 x 32, Kernel: 2 x 2 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
20 out of 36 nodes do not share the minibatch layout with the input data.
Post-processing network complete.
evalNodeNames are not specified, using all the default evalnodes and training criterion nodes.
Allocating matrices for forward and/or backward propagation.
Memory Sharing Structure:
(nil): {[ce Gradient[1]] [conv1.c.W Gradient[16 x 25]] [conv1.c.c.b Gradient[16 x 1]] [conv1.c.c.c Gradient[28 x 28 x 16 x *1]] [conv1.c.c.isd Gradient[16 x 1]] [conv1.c.c.m Gradient[16 x 1]] [conv1.c.c.sc Gradient[16 x 1]] [conv1.c.c.y Gradient[28 x 28 x 16 x *1]] [conv1.y Gradient[28 x 28 x 16 x *1]] [conv2.c.W Gradient[32 x 400]] [conv2.c.c.b Gradient[32 x 1]] [conv2.c.c.c Gradient[14 x 14 x 32 x *1]] [conv2.c.c.isd Gradient[32 x 1]] [conv2.c.c.m Gradient[32 x 1]] [conv2.c.c.sc Gradient[32 x 1]] [conv2.c.c.y Gradient[14 x 14 x 32 x *1]] [conv2.y Gradient[14 x 14 x 32 x *1]] [err Gradient[1]] [featScale Gradient[1 x 1]] [featScaled Gradient[28 x 28 x 1 x *1]] [features Gradient[28 x 28 x 1 x *1]] [h1.W Gradient[128 x 7 x 7 x 32]] [h1.b Gradient[128 x 1]] [h1.bn Gradient[128 x *1]] [h1.isd Gradient[128 x 1]] [h1.m Gradient[128 x 1]] [h1.sc Gradient[128 x 1]] [h1.t Gradient[128 x *1]] [h1.y Gradient[128 x *1]] [labels Gradient[10 x *1]] [ol.W Gradient[10 x 128]] [ol.b Gradient[10 x 1]] [ol.t Gradient[10 x *1]] [ol.z Gradient[10 x 1 x *1]] [pool1 Gradient[14 x 14 x 16 x *1]] [pool2 Gradient[7 x 7 x 32 x *1]] }
0x7f50cab10a28: {[h1.sc Value[128 x 1]] }
0x7f50cab11988: {[h1.W Value[128 x 7 x 7 x 32]] }
0x7f50cab132e8: {[labels Value[10 x *1]] }
0x7f50cab13968: {[ol.b Value[10 x 1]] }
0x7f50cab14c88: {[h1.b Value[128 x 1]] }
0x7f50cab15368: {[h1.isd Value[128 x 1]] }
0x7f50cab15da8: {[h1.m Value[128 x 1]] }
0x7f50cab160c8: {[conv1.c.c.isd Value[16 x 1]] }
0x7f50cab17e68: {[ol.W Value[10 x 128]] }
0x7f50cab1ac98: {[ce Value[1]] }
0x7f50cab1c4f8: {[err Value[1]] }
0x7f50cabd0b58: {[conv1.c.c.c Value[28 x 28 x 16 x *1]] }
0x7f50cabd0e98: {[featScaled Value[28 x 28 x 1 x *1]] }
0x7f50cabd1148: {[conv1.c.c.y Value[28 x 28 x 16 x *1]] }
0x7f50cabd1f48: {[conv1.y Value[28 x 28 x 16 x *1]] }
0x7f50cabd2108: {[pool1 Value[14 x 14 x 16 x *1]] }
0x7f50cabd22c8: {[conv2.c.c.c Value[14 x 14 x 32 x *1]] }
0x7f50cabd2648: {[conv2.c.c.y Value[14 x 14 x 32 x *1]] }
0x7f50cabd2b88: {[conv2.y Value[14 x 14 x 32 x *1]] }
0x7f50cabd2d48: {[pool2 Value[7 x 7 x 32 x *1]] }
0x7f50cabd2f08: {[h1.t Value[128 x *1]] }
0x7f50cabd9558: {[h1.bn Value[128 x *1]] }
0x7f50cabd9a98: {[h1.y Value[128 x *1]] }
0x7f50cabd9c58: {[ol.t Value[10 x *1]] }
0x7f50cabd9e18: {[ol.z Value[10 x 1 x *1]] }
0x7f50cad85a38: {[conv1.c.c.b Value[16 x 1]] }
0x7f50d5601148: {[conv2.c.c.sc Value[32 x 1]] }
0x7f50d5601ea8: {[conv2.c.W Value[32 x 400]] }
0x7f50d5602728: {[conv1.c.W Value[16 x 25]] }
0x7f50d5602e58: {[conv2.c.c.b Value[32 x 1]] }
0x7f50d5603b28: {[conv1.c.c.sc Value[16 x 1]] }
0x7f50d56045d8: {[conv1.c.c.m Value[16 x 1]] }
0x7f50d5606dd8: {[conv2.c.c.isd Value[32 x 1]] }
0x7f50d5608478: {[conv2.c.c.m Value[32 x 1]] }
0x7f50d5609d38: {[featScale Value[1 x 1]] }
0x7f50d560a658: {[features Value[28 x 28 x 1 x *1]] }
05/13/2016 15:10:47: Final Results: Minibatch[1-313]: err = 0.00660000 * 10000; ce = 0.02083102 * 10000; perplexity = 1.02104950
05/13/2016 15:10:47: Action "test" complete.
05/13/2016 15:10:47: __COMPLETED__

579
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm/baseline.windows.txt Normal file
Просмотреть файл

@ -0,0 +1,579 @@
=== Running /cygdrive/c/jenkins/workspace/CNTK-Test-Windows-W1/x64/release/cntk.exe configFile=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/Config/03_ConvBatchNorm.cntk currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu\TestData RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu\TestData ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu DeviceId=0 timestamping=true train=[SGD=[maxEpochs=3]] imageLayout="cudnn"
-------------------------------------------------------------------
Build info:
Built time: May 13 2016 08:06:01
Last modified date: Thu May 12 07:31:50 2016
Build type: Release
Build target: GPU
With 1bit-SGD: no
CUDA_PATH: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5
CUB_PATH: c:\src\cub-1.4.1
CUDNN_PATH: c:\NVIDIA\cudnn-4.0\cuda
Build Branch: HEAD
Build SHA1: 35fadc316f045d843bbd9b85061250a959268787
Built by svcphil on Philly-Pool3
Build Path: c:\Jenkins\workspace\CNTK-Build-Windows\Source\CNTK\
-------------------------------------------------------------------
Changed current directory to C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu\TestData
05/13/2016 08:16:56: -------------------------------------------------------------------
05/13/2016 08:16:56: Build info:
05/13/2016 08:16:56: Built time: May 13 2016 08:06:01
05/13/2016 08:16:56: Last modified date: Thu May 12 07:31:50 2016
05/13/2016 08:16:56: Build type: Release
05/13/2016 08:16:56: Build target: GPU
05/13/2016 08:16:56: With 1bit-SGD: no
05/13/2016 08:16:56: CUDA_PATH: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5
05/13/2016 08:16:56: CUB_PATH: c:\src\cub-1.4.1
05/13/2016 08:16:56: CUDNN_PATH: c:\NVIDIA\cudnn-4.0\cuda
05/13/2016 08:16:56: Build Branch: HEAD
05/13/2016 08:16:56: Build SHA1: 35fadc316f045d843bbd9b85061250a959268787
05/13/2016 08:16:56: Built by svcphil on Philly-Pool3
05/13/2016 08:16:56: Build Path: c:\Jenkins\workspace\CNTK-Build-Windows\Source\CNTK\
05/13/2016 08:16:56: -------------------------------------------------------------------
05/13/2016 08:16:56: Running on Philly-Pool2 at 2016/05/13 08:16:56
05/13/2016 08:16:56: Command line:
C:\jenkins\workspace\CNTK-Test-Windows-W1\x64\release\cntk.exe configFile=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/Config/03_ConvBatchNorm.cntk currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu\TestData RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu\TestData ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu DeviceId=0 timestamping=true train=[SGD=[maxEpochs=3]] imageLayout="cudnn"
05/13/2016 08:16:56: >>>>>>>>>>>>>>>>>>>> RAW CONFIG (VARIABLES NOT RESOLVED) >>>>>>>>>>>>>>>>>>>>
05/13/2016 08:16:56: RootDir = ".."
ConfigDir = "$RootDir$/Config"
DataDir = "$RootDir$/Data"
OutputDir = "$RootDir$/Output"
ModelDir = "$OutputDir$/Models"
deviceId = 0
imageLayout = "cudnn"
command = train:test
precision = "float"
modelPath = "$ModelDir$/03_ConvBatchNorm"
ndlMacros = "$ConfigDir$/Macros.ndl"
traceLevel=1
numMBsToShowResult=500
initOnCPUOnly=true
train = [
action = "train"
NDLNetworkBuilder = [
networkDescription = "$ConfigDir$/03_ConvBatchNorm.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.5:0.1
momentumPerMB = 0.9
maxEpochs = 2
batchNormalizationBlendTimeConstant=0:1#INF
]
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
test = [
action = "test"
minibatchSize = 32
modelPath=$ModelDir$/03_ConvBatchNorm
NDLNetworkBuilder = [
networkDescription = "$ConfigDir$/03_ConvBatchNorm.ndl"
]
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu\TestData
RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu
DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu\TestData
ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config
OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu
DeviceId=0
timestamping=true
train=[SGD=[maxEpochs=3]]
imageLayout="cudnn"
05/13/2016 08:16:56: <<<<<<<<<<<<<<<<<<<< RAW CONFIG (VARIABLES NOT RESOLVED) <<<<<<<<<<<<<<<<<<<<
05/13/2016 08:16:56: >>>>>>>>>>>>>>>>>>>> RAW CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>
05/13/2016 08:16:56: RootDir = ".."
ConfigDir = "../Config"
DataDir = "../Data"
OutputDir = "../Output"
ModelDir = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu/Models"
deviceId = 0
imageLayout = "cudnn"
command = train:test
precision = "float"
modelPath = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu/Models/03_ConvBatchNorm"
ndlMacros = "C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/Macros.ndl"
traceLevel=1
numMBsToShowResult=500
initOnCPUOnly=true
train = [
action = "train"
NDLNetworkBuilder = [
networkDescription = "C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/03_ConvBatchNorm.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.5:0.1
momentumPerMB = 0.9
maxEpochs = 2
batchNormalizationBlendTimeConstant=0:1#INF
]
reader = [
readerType = "CNTKTextFormatReader"
file = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu\TestData/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
test = [
action = "test"
minibatchSize = 32
modelPath=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu/Models/03_ConvBatchNorm
NDLNetworkBuilder = [
networkDescription = "C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/03_ConvBatchNorm.ndl"
]
reader = [
readerType = "CNTKTextFormatReader"
file = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu\TestData/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu\TestData
RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu
DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu\TestData
ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config
OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu
DeviceId=0
timestamping=true
train=[SGD=[maxEpochs=3]]
imageLayout="cudnn"
05/13/2016 08:16:56: <<<<<<<<<<<<<<<<<<<< RAW CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<
05/13/2016 08:16:56: >>>>>>>>>>>>>>>>>>>> PROCESSED CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>
configparameters: 03_ConvBatchNorm.cntk:command=train:test
configparameters: 03_ConvBatchNorm.cntk:ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config
configparameters: 03_ConvBatchNorm.cntk:currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu\TestData
configparameters: 03_ConvBatchNorm.cntk:DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu\TestData
configparameters: 03_ConvBatchNorm.cntk:deviceId=0
configparameters: 03_ConvBatchNorm.cntk:imageLayout=cudnn
configparameters: 03_ConvBatchNorm.cntk:initOnCPUOnly=true
configparameters: 03_ConvBatchNorm.cntk:ModelDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu/Models
configparameters: 03_ConvBatchNorm.cntk:modelPath=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu/Models/03_ConvBatchNorm
configparameters: 03_ConvBatchNorm.cntk:ndlMacros=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/Macros.ndl
configparameters: 03_ConvBatchNorm.cntk:numMBsToShowResult=500
configparameters: 03_ConvBatchNorm.cntk:OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu
configparameters: 03_ConvBatchNorm.cntk:precision=float
configparameters: 03_ConvBatchNorm.cntk:RootDir=..
configparameters: 03_ConvBatchNorm.cntk:RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu
configparameters: 03_ConvBatchNorm.cntk:test=[
action = "test"
minibatchSize = 32
modelPath=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu/Models/03_ConvBatchNorm
NDLNetworkBuilder = [
networkDescription = "C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/03_ConvBatchNorm.ndl"
]
reader = [
readerType = "CNTKTextFormatReader"
file = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu\TestData/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
configparameters: 03_ConvBatchNorm.cntk:timestamping=true
configparameters: 03_ConvBatchNorm.cntk:traceLevel=1
configparameters: 03_ConvBatchNorm.cntk:train=[
action = "train"
NDLNetworkBuilder = [
networkDescription = "C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\MNIST\Config/03_ConvBatchNorm.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.5:0.1
momentumPerMB = 0.9
maxEpochs = 2
batchNormalizationBlendTimeConstant=0:1#INF
]
reader = [
readerType = "CNTKTextFormatReader"
file = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu\TestData/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
] [SGD=[maxEpochs=3]]
05/13/2016 08:16:56: <<<<<<<<<<<<<<<<<<<< PROCESSED CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<
05/13/2016 08:16:56: Commands: train test
05/13/2016 08:16:56: Precision = "float"
05/13/2016 08:16:56: CNTKModelPath: C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu/Models/03_ConvBatchNorm
05/13/2016 08:16:56: CNTKCommandTrainInfo: train : 3
05/13/2016 08:16:56: CNTKCommandTrainInfo: CNTKNoMoreCommands_Total : 3
05/13/2016 08:16:56: ##############################################################################
05/13/2016 08:16:56: # #
05/13/2016 08:16:56: # Action "train" #
05/13/2016 08:16:56: # #
05/13/2016 08:16:56: ##############################################################################
05/13/2016 08:16:56: CNTKCommandTrainBegin: train
NDLBuilder Using GPU 0
05/13/2016 08:16:57: Creating virgin network.
Post-processing network...
3 roots:
ce = CrossEntropyWithSoftmax()
err = ErrorPrediction()
ol.z = Plus()
Validating network. 36 nodes to process in pass 1.
Validating --> labels = InputValue() : -> [10 x *]
Validating --> ol.W = LearnableParameter() : -> [10 x 128]
Validating --> h1.W = LearnableParameter() : -> [128 x 1568]
Validating --> conv2.c.W = LearnableParameter() : -> [32 x 400]
Validating --> conv1.c.W = LearnableParameter() : -> [16 x 25]
Validating --> featScale = LearnableParameter() : -> [1 x 1]
Validating --> features = InputValue() : -> [28 x 28 x 1 x *]
Validating --> featScaled = ElementTimes (featScale, features) : [1 x 1], [28 x 28 x 1 x *] -> [28 x 28 x 1 x *]
Validating --> conv1.c.c.c = Convolution (conv1.c.W, featScaled) : [16 x 25], [28 x 28 x 1 x *] -> [28 x 28 x 16 x *]
Validating --> conv1.c.c.sc = LearnableParameter() : -> [16 x 1]
Validating --> conv1.c.c.b = LearnableParameter() : -> [16 x 1]
Validating --> conv1.c.c.m = LearnableParameter() : -> [16 x 1]
Validating --> conv1.c.c.isd = LearnableParameter() : -> [16 x 1]
Validating --> conv1.c.c.y = BatchNormalization (conv1.c.c.c, conv1.c.c.sc, conv1.c.c.b, conv1.c.c.m, conv1.c.c.isd) : [28 x 28 x 16 x *], [16 x 1], [16 x 1], [16 x 1], [16 x 1] -> [28 x 28 x 16 x *]
Validating --> conv1.y = RectifiedLinear (conv1.c.c.y) : [28 x 28 x 16 x *] -> [28 x 28 x 16 x *]
Validating --> pool1 = MaxPooling (conv1.y) : [28 x 28 x 16 x *] -> [14 x 14 x 16 x *]
Validating --> conv2.c.c.c = Convolution (conv2.c.W, pool1) : [32 x 400], [14 x 14 x 16 x *] -> [14 x 14 x 32 x *]
Validating --> conv2.c.c.sc = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.b = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.m = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.isd = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.y = BatchNormalization (conv2.c.c.c, conv2.c.c.sc, conv2.c.c.b, conv2.c.c.m, conv2.c.c.isd) : [14 x 14 x 32 x *], [32 x 1], [32 x 1], [32 x 1], [32 x 1] -> [14 x 14 x 32 x *]
Validating --> conv2.y = RectifiedLinear (conv2.c.c.y) : [14 x 14 x 32 x *] -> [14 x 14 x 32 x *]
Validating --> pool2 = MaxPooling (conv2.y) : [14 x 14 x 32 x *] -> [7 x 7 x 32 x *]
h1.t Times operation: For legacy compatibility, the sample layout of left input (h1.W LearnableParameter operation) was patched to [128 x 7 x 7 x 32] (from [128 x 1568])
Validating --> h1.t = Times (h1.W, pool2) : [128 x 7 x 7 x 32], [7 x 7 x 32 x *] -> [128 x *]
Validating --> h1.sc = LearnableParameter() : -> [128 x 1]
Validating --> h1.b = LearnableParameter() : -> [128 x 1]
Validating --> h1.m = LearnableParameter() : -> [128 x 1]
Validating --> h1.isd = LearnableParameter() : -> [128 x 1]
Validating --> h1.bn = BatchNormalization (h1.t, h1.sc, h1.b, h1.m, h1.isd) : [128 x *], [128 x 1], [128 x 1], [128 x 1], [128 x 1] -> [128 x *]
Validating --> h1.y = RectifiedLinear (h1.bn) : [128 x *] -> [128 x *]
Validating --> ol.t = Times (ol.W, h1.y) : [10 x 128], [128 x *] -> [10 x *]
Validating --> ol.b = LearnableParameter() : -> [10 x 1]
Validating --> ol.z = Plus (ol.t, ol.b) : [10 x *], [10 x 1] -> [10 x 1 x *]
Validating --> ce = CrossEntropyWithSoftmax (labels, ol.z) : [10 x *], [10 x 1 x *] -> [1]
Validating --> err = ErrorPrediction (labels, ol.z) : [10 x *], [10 x 1 x *] -> [1]
Validating network. 16 nodes to process in pass 2.
Validating network, final pass.
Using cuDNN convolution engine for geometry: Input: 28 x 28 x 1, Output: 28 x 28 x 16, Kernel: 5 x 5 x 1, Map: 1 x 1 x 16, Stride: 1 x 1 x 1, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
Using cuDNN convolution engine for geometry: Input: 28 x 28 x 16, Output: 14 x 14 x 16, Kernel: 2 x 2 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 14 x 14 x 16, Output: 14 x 14 x 32, Kernel: 5 x 5 x 16, Map: 1 x 1 x 32, Stride: 1 x 1 x 16, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
Using cuDNN convolution engine for geometry: Input: 14 x 14 x 32, Output: 7 x 7 x 32, Kernel: 2 x 2 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
20 out of 36 nodes do not share the minibatch layout with the input data.
Post-processing network complete.
05/13/2016 08:16:58: Created model with 36 nodes on GPU 0.
05/13/2016 08:16:58: Training criterion node(s):
05/13/2016 08:16:58: ce = CrossEntropyWithSoftmax
05/13/2016 08:16:58: Evaluation criterion node(s):
05/13/2016 08:16:58: err = ErrorPrediction
Allocating matrices for forward and/or backward propagation.
Memory Sharing Structure:
0000000000000000: {[conv1.c.c.isd Gradient[16 x 1]] [conv1.c.c.m Gradient[16 x 1]] [conv2.c.c.isd Gradient[32 x 1]] [conv2.c.c.m Gradient[32 x 1]] [err Gradient[1]] [featScale Gradient[1 x 1]] [featScaled Gradient[28 x 28 x 1 x *]] [features Gradient[28 x 28 x 1 x *]] [h1.isd Gradient[128 x 1]] [h1.m Gradient[128 x 1]] [labels Gradient[10 x *]] }
00000093B2DC5E20: {[features Value[28 x 28 x 1 x *]] }
00000093CB445890: {[ol.W Value[10 x 128]] }
00000093CB446290: {[ol.b Value[10 x 1]] }
00000093CB5EF4D0: {[conv2.c.c.m Value[32 x 1]] }
00000093CB5EF570: {[conv1.c.W Value[16 x 25]] }
00000093CB5EF610: {[h1.sc Value[128 x 1]] }
00000093CB5EF9D0: {[conv2.c.c.b Value[32 x 1]] }
00000093CB5EFBB0: {[h1.b Value[128 x 1]] }
00000093CB5EFCF0: {[h1.isd Value[128 x 1]] }
00000093CB5EFD90: {[h1.m Value[128 x 1]] }
00000093CB5F03D0: {[conv1.c.c.b Value[16 x 1]] }
00000093CB5F0470: {[conv1.c.c.sc Value[16 x 1]] }
00000093CB5F05B0: {[conv1.c.c.isd Value[16 x 1]] }
00000093CB5F06F0: {[conv2.c.W Value[32 x 400]] }
00000093CB5F0830: {[conv2.c.c.sc Value[32 x 1]] }
00000093CB5F08D0: {[conv2.c.c.isd Value[32 x 1]] }
00000093CB5F0970: {[labels Value[10 x *]] }
00000093CB5F0BF0: {[conv1.c.c.m Value[16 x 1]] }
00000093CB5F0D30: {[featScale Value[1 x 1]] }
00000093CB5F0DD0: {[h1.W Value[128 x 7 x 7 x 32]] }
00000093D1AAE180: {[conv2.c.c.b Gradient[32 x 1]] }
00000093D1AAE360: {[ol.t Gradient[10 x *]] [pool1 Gradient[14 x 14 x 16 x *]] [pool2 Gradient[7 x 7 x 32 x *]] }
00000093D1AAE400: {[h1.W Gradient[128 x 7 x 7 x 32]] }
00000093D1AAE5E0: {[conv1.c.c.c Gradient[28 x 28 x 16 x *]] [conv1.y Value[28 x 28 x 16 x *]] }
00000093D1AAE680: {[h1.b Gradient[128 x 1]] }
00000093D1AAE9A0: {[err Value[1]] }
00000093D1AAED60: {[ol.z Value[10 x 1 x *]] }
00000093D1AAEE00: {[ce Value[1]] }
00000093D1AAEF40: {[conv1.c.c.y Value[28 x 28 x 16 x *]] }
00000093D1AAF080: {[ol.b Gradient[10 x 1]] }
00000093D1AAF120: {[conv1.c.W Gradient[16 x 25]] [conv2.c.c.c Value[14 x 14 x 32 x *]] }
00000093D1AAF1C0: {[h1.bn Gradient[128 x *]] [ol.t Value[10 x *]] }
00000093D1AAF260: {[conv1.c.c.y Gradient[28 x 28 x 16 x *]] [pool1 Value[14 x 14 x 16 x *]] }
00000093D1AAF440: {[ol.W Gradient[10 x 128]] [ol.z Gradient[10 x 1 x *]] }
00000093D1AAFB20: {[h1.bn Value[128 x *]] }
00000093D1AAFDA0: {[conv1.c.c.c Value[28 x 28 x 16 x *]] }
00000093D1AAFE40: {[conv2.c.c.y Value[14 x 14 x 32 x *]] }
00000093D1AAFF80: {[conv2.c.c.y Gradient[14 x 14 x 32 x *]] [pool2 Value[7 x 7 x 32 x *]] }
00000093D1AB0020: {[conv2.c.W Gradient[32 x 400]] [h1.t Gradient[128 x *]] [h1.y Value[128 x *]] }
00000093D1AB00C0: {[h1.sc Gradient[128 x 1]] [h1.y Gradient[128 x *]] }
00000093D1AB0200: {[ce Gradient[1]] }
00000093D1AB03E0: {[conv1.c.c.sc Gradient[16 x 1]] [conv1.y Gradient[28 x 28 x 16 x *]] }
00000093D1AB0480: {[conv1.c.c.b Gradient[16 x 1]] [conv2.c.c.c Gradient[14 x 14 x 32 x *]] [conv2.y Value[14 x 14 x 32 x *]] }
00000093D1AB0660: {[featScaled Value[28 x 28 x 1 x *]] }
00000093D1AB0700: {[conv2.c.c.sc Gradient[32 x 1]] [conv2.y Gradient[14 x 14 x 32 x *]] [h1.t Value[128 x *]] }
05/13/2016 08:16:58: No PreCompute nodes found, skipping PreCompute step.
05/13/2016 08:16:58: Starting Epoch 1: learning rate per sample = 0.015625 effective momentum = 0.900000 momentum as time constant = 303.7 samples
05/13/2016 08:16:58: Starting minibatch loop.
05/13/2016 08:17:02: Epoch[ 1 of 3]-Minibatch[ 1- 500, 26.67%]: ce = 0.17330922 * 16000; err = 0.05325000 * 16000; time = 4.3656s; samplesPerSecond = 3665.0
05/13/2016 08:17:04: Epoch[ 1 of 3]-Minibatch[ 501-1000, 53.33%]: ce = 0.07897408 * 16000; err = 0.02456250 * 16000; time = 1.7980s; samplesPerSecond = 8899.0
05/13/2016 08:17:06: Epoch[ 1 of 3]-Minibatch[1001-1500, 80.00%]: ce = 0.06288062 * 16000; err = 0.02012500 * 16000; time = 1.7989s; samplesPerSecond = 8894.1
05/13/2016 08:17:07: Finished Epoch[ 1 of 3]: [Training] ce = 0.09585953 * 60000; err = 0.02956667 * 60000; totalSamplesSeen = 60000; learningRatePerSample = 0.015625; epochTime=9.34707s
05/13/2016 08:17:07: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu/Models/03_ConvBatchNorm.1'
Setting batch normalization blend time constant to 1.#INF.
05/13/2016 08:17:07: Starting Epoch 2: learning rate per sample = 0.003125 effective momentum = 0.900000 momentum as time constant = 303.7 samples
05/13/2016 08:17:07: Starting minibatch loop.
05/13/2016 08:17:09: Epoch[ 2 of 3]-Minibatch[ 1- 500, 26.67%]: ce = 0.02381749 * 16000; err = 0.00700000 * 16000; time = 1.7975s; samplesPerSecond = 8901.2
05/13/2016 08:17:11: Epoch[ 2 of 3]-Minibatch[ 501-1000, 53.33%]: ce = 0.02147904 * 16000; err = 0.00687500 * 16000; time = 1.7971s; samplesPerSecond = 8903.3
05/13/2016 08:17:13: Epoch[ 2 of 3]-Minibatch[1001-1500, 80.00%]: ce = 0.01875302 * 16000; err = 0.00581250 * 16000; time = 1.7965s; samplesPerSecond = 8906.1
05/13/2016 08:17:14: Finished Epoch[ 2 of 3]: [Training] ce = 0.02042361 * 60000; err = 0.00626667 * 60000; totalSamplesSeen = 120000; learningRatePerSample = 0.003125; epochTime=6.7551s
05/13/2016 08:17:14: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu/Models/03_ConvBatchNorm.2'
05/13/2016 08:17:14: Starting Epoch 3: learning rate per sample = 0.003125 effective momentum = 0.900000 momentum as time constant = 303.7 samples
05/13/2016 08:17:14: Starting minibatch loop.
05/13/2016 08:17:16: Epoch[ 3 of 3]-Minibatch[ 1- 500, 26.67%]: ce = 0.01552748 * 16000; err = 0.00400000 * 16000; time = 1.7980s; samplesPerSecond = 8899.0
05/13/2016 08:17:18: Epoch[ 3 of 3]-Minibatch[ 501-1000, 53.33%]: ce = 0.01295741 * 16000; err = 0.00356250 * 16000; time = 1.7961s; samplesPerSecond = 8908.2
05/13/2016 08:17:20: Epoch[ 3 of 3]-Minibatch[1001-1500, 80.00%]: ce = 0.01382423 * 16000; err = 0.00393750 * 16000; time = 1.7964s; samplesPerSecond = 8906.7
05/13/2016 08:17:21: Finished Epoch[ 3 of 3]: [Training] ce = 0.01415997 * 60000; err = 0.00391667 * 60000; totalSamplesSeen = 180000; learningRatePerSample = 0.003125; epochTime=6.75556s
05/13/2016 08:17:21: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\MNIST_03_ConvBatchNorm@release_gpu/Models/03_ConvBatchNorm'
05/13/2016 08:17:21: CNTKCommandTrainEnd: train
05/13/2016 08:17:21: Action "train" complete.
05/13/2016 08:17:21: ##############################################################################
05/13/2016 08:17:21: # #
05/13/2016 08:17:21: # Action "test" #
05/13/2016 08:17:21: # #
05/13/2016 08:17:21: ##############################################################################
Post-processing network...
3 roots:
ce = CrossEntropyWithSoftmax()
err = ErrorPrediction()
ol.z = Plus()
Validating network. 36 nodes to process in pass 1.
Validating --> labels = InputValue() : -> [10 x *1]
Validating --> ol.W = LearnableParameter() : -> [10 x 128]
Validating --> h1.W = LearnableParameter() : -> [128 x 7 x 7 x 32]
Validating --> conv2.c.W = LearnableParameter() : -> [32 x 400]
Validating --> conv1.c.W = LearnableParameter() : -> [16 x 25]
Validating --> featScale = LearnableParameter() : -> [1 x 1]
Validating --> features = InputValue() : -> [28 x 28 x 1 x *1]
Validating --> featScaled = ElementTimes (featScale, features) : [1 x 1], [28 x 28 x 1 x *1] -> [28 x 28 x 1 x *1]
Validating --> conv1.c.c.c = Convolution (conv1.c.W, featScaled) : [16 x 25], [28 x 28 x 1 x *1] -> [28 x 28 x 16 x *1]
Validating --> conv1.c.c.sc = LearnableParameter() : -> [16 x 1]
Validating --> conv1.c.c.b = LearnableParameter() : -> [16 x 1]
Validating --> conv1.c.c.m = LearnableParameter() : -> [16 x 1]
Validating --> conv1.c.c.isd = LearnableParameter() : -> [16 x 1]
Validating --> conv1.c.c.y = BatchNormalization (conv1.c.c.c, conv1.c.c.sc, conv1.c.c.b, conv1.c.c.m, conv1.c.c.isd) : [28 x 28 x 16 x *1], [16 x 1], [16 x 1], [16 x 1], [16 x 1] -> [28 x 28 x 16 x *1]
Validating --> conv1.y = RectifiedLinear (conv1.c.c.y) : [28 x 28 x 16 x *1] -> [28 x 28 x 16 x *1]
Validating --> pool1 = MaxPooling (conv1.y) : [28 x 28 x 16 x *1] -> [14 x 14 x 16 x *1]
Validating --> conv2.c.c.c = Convolution (conv2.c.W, pool1) : [32 x 400], [14 x 14 x 16 x *1] -> [14 x 14 x 32 x *1]
Validating --> conv2.c.c.sc = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.b = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.m = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.isd = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.y = BatchNormalization (conv2.c.c.c, conv2.c.c.sc, conv2.c.c.b, conv2.c.c.m, conv2.c.c.isd) : [14 x 14 x 32 x *1], [32 x 1], [32 x 1], [32 x 1], [32 x 1] -> [14 x 14 x 32 x *1]
Validating --> conv2.y = RectifiedLinear (conv2.c.c.y) : [14 x 14 x 32 x *1] -> [14 x 14 x 32 x *1]
Validating --> pool2 = MaxPooling (conv2.y) : [14 x 14 x 32 x *1] -> [7 x 7 x 32 x *1]
Validating --> h1.t = Times (h1.W, pool2) : [128 x 7 x 7 x 32], [7 x 7 x 32 x *1] -> [128 x *1]
Validating --> h1.sc = LearnableParameter() : -> [128 x 1]
Validating --> h1.b = LearnableParameter() : -> [128 x 1]
Validating --> h1.m = LearnableParameter() : -> [128 x 1]
Validating --> h1.isd = LearnableParameter() : -> [128 x 1]
Validating --> h1.bn = BatchNormalization (h1.t, h1.sc, h1.b, h1.m, h1.isd) : [128 x *1], [128 x 1], [128 x 1], [128 x 1], [128 x 1] -> [128 x *1]
Validating --> h1.y = RectifiedLinear (h1.bn) : [128 x *1] -> [128 x *1]
Validating --> ol.t = Times (ol.W, h1.y) : [10 x 128], [128 x *1] -> [10 x *1]
Validating --> ol.b = LearnableParameter() : -> [10 x 1]
Validating --> ol.z = Plus (ol.t, ol.b) : [10 x *1], [10 x 1] -> [10 x 1 x *1]
Validating --> ce = CrossEntropyWithSoftmax (labels, ol.z) : [10 x *1], [10 x 1 x *1] -> [1]
Validating --> err = ErrorPrediction (labels, ol.z) : [10 x *1], [10 x 1 x *1] -> [1]
Validating network. 16 nodes to process in pass 2.
Validating network, final pass.
Using cuDNN convolution engine for geometry: Input: 28 x 28 x 1, Output: 28 x 28 x 16, Kernel: 5 x 5 x 1, Map: 1 x 1 x 16, Stride: 1 x 1 x 1, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
Using cuDNN convolution engine for geometry: Input: 28 x 28 x 16, Output: 14 x 14 x 16, Kernel: 2 x 2 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 14 x 14 x 16, Output: 14 x 14 x 32, Kernel: 5 x 5 x 16, Map: 1 x 1 x 32, Stride: 1 x 1 x 16, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
Using cuDNN convolution engine for geometry: Input: 14 x 14 x 32, Output: 7 x 7 x 32, Kernel: 2 x 2 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
20 out of 36 nodes do not share the minibatch layout with the input data.
Post-processing network complete.
evalNodeNames are not specified, using all the default evalnodes and training criterion nodes.
Allocating matrices for forward and/or backward propagation.
Memory Sharing Structure:
0000000000000000: {[ce Gradient[1]] [conv1.c.W Gradient[16 x 25]] [conv1.c.c.b Gradient[16 x 1]] [conv1.c.c.c Gradient[28 x 28 x 16 x *1]] [conv1.c.c.isd Gradient[16 x 1]] [conv1.c.c.m Gradient[16 x 1]] [conv1.c.c.sc Gradient[16 x 1]] [conv1.c.c.y Gradient[28 x 28 x 16 x *1]] [conv1.y Gradient[28 x 28 x 16 x *1]] [conv2.c.W Gradient[32 x 400]] [conv2.c.c.b Gradient[32 x 1]] [conv2.c.c.c Gradient[14 x 14 x 32 x *1]] [conv2.c.c.isd Gradient[32 x 1]] [conv2.c.c.m Gradient[32 x 1]] [conv2.c.c.sc Gradient[32 x 1]] [conv2.c.c.y Gradient[14 x 14 x 32 x *1]] [conv2.y Gradient[14 x 14 x 32 x *1]] [err Gradient[1]] [featScale Gradient[1 x 1]] [featScaled Gradient[28 x 28 x 1 x *1]] [features Gradient[28 x 28 x 1 x *1]] [h1.W Gradient[128 x 7 x 7 x 32]] [h1.b Gradient[128 x 1]] [h1.bn Gradient[128 x *1]] [h1.isd Gradient[128 x 1]] [h1.m Gradient[128 x 1]] [h1.sc Gradient[128 x 1]] [h1.t Gradient[128 x *1]] [h1.y Gradient[128 x *1]] [labels Gradient[10 x *1]] [ol.W Gradient[10 x 128]] [ol.b Gradient[10 x 1]] [ol.t Gradient[10 x *1]] [ol.z Gradient[10 x 1 x *1]] [pool1 Gradient[14 x 14 x 16 x *1]] [pool2 Gradient[7 x 7 x 32 x *1]] }
00000093D1AAEE00: {[pool2 Value[7 x 7 x 32 x *1]] }
00000093D1AAEEA0: {[conv2.c.c.y Value[14 x 14 x 32 x *1]] }
00000093D1AAF300: {[h1.y Value[128 x *1]] }
00000093D1AAF580: {[conv2.y Value[14 x 14 x 32 x *1]] }
00000093D1AAF760: {[h1.bn Value[128 x *1]] }
00000093D1AAF940: {[ol.t Value[10 x *1]] }
00000093D1AB0200: {[h1.t Value[128 x *1]] }
00000093D1AB0700: {[ol.z Value[10 x 1 x *1]] }
00000093D1CA31E0: {[conv2.c.c.sc Value[32 x 1]] }
00000093D1CA3320: {[conv2.c.c.isd Value[32 x 1]] }
00000093D1CA35A0: {[conv2.c.W Value[32 x 400]] }
00000093D1CA3BE0: {[conv2.c.c.m Value[32 x 1]] }
00000093D1CA3FA0: {[conv2.c.c.b Value[32 x 1]] }
00000093D1CA40E0: {[featScale Value[1 x 1]] }
00000093D1CA4720: {[features Value[28 x 28 x 1 x *1]] }
00000093D1CA4C20: {[conv1.c.c.isd Value[16 x 1]] }
00000093D1CA5440: {[conv1.c.c.m Value[16 x 1]] }
00000093D1CA54E0: {[conv1.c.c.b Value[16 x 1]] }
00000093D1CA5580: {[conv1.c.c.sc Value[16 x 1]] }
00000093D1CA5620: {[conv1.c.W Value[16 x 25]] }
00000093D1CA5A80: {[ol.b Value[10 x 1]] }
00000093D1CA5B20: {[ol.W Value[10 x 128]] }
00000093D1CA60C0: {[err Value[1]] }
00000093D1CA6160: {[h1.W Value[128 x 7 x 7 x 32]] }
00000093D1CA6200: {[conv1.c.c.c Value[28 x 28 x 16 x *1]] }
00000093D1CA62A0: {[h1.isd Value[128 x 1]] }
00000093D1CA6340: {[ce Value[1]] }
00000093D1CA6480: {[h1.b Value[128 x 1]] }
00000093D1CA6660: {[conv1.c.c.y Value[28 x 28 x 16 x *1]] }
00000093D1CA6700: {[featScaled Value[28 x 28 x 1 x *1]] }
00000093D1CA68E0: {[h1.sc Value[128 x 1]] }
00000093D1CA6980: {[conv1.y Value[28 x 28 x 16 x *1]] }
00000093D1CA6AC0: {[pool1 Value[14 x 14 x 16 x *1]] }
00000093D1CA6B60: {[conv2.c.c.c Value[14 x 14 x 32 x *1]] }
00000093D1CA6D40: {[h1.m Value[128 x 1]] }
00000093D1CA6DE0: {[labels Value[10 x *1]] }
05/13/2016 08:17:32: Final Results: Minibatch[1-313]: err = 0.00710000 * 10000; ce = 0.02063067 * 10000; perplexity = 1.02084496
05/13/2016 08:17:32: Action "test" complete.
05/13/2016 08:17:32: __COMPLETED__

11
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm/run-test Executable file
Просмотреть файл

@ -0,0 +1,11 @@
#!/bin/bash
. $TEST_DIR/../run-test-common
cntkrun $CNTKTextConfigDir/03_ConvBatchNorm.cntk "train=[SGD=[maxEpochs=3]] imageLayout=\"$imageLayout\""
ExitCode=$?
# Delete the test data if copied
[[ "$Copied" -eq "1" ]] && rm -rf "$DataDir"
exit $ExitCode

39
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/03_ConvBatchNorm/testcases.yml Normal file
Просмотреть файл

@ -0,0 +1,39 @@
dataDir: ../../../../../../../Examples/Image/MNIST/Data
tags:
# In BVT, run Release GPU
- bvt-e (build_sku=='gpu') and (device=='gpu') and (flavor=='release')
# In Nightly, additionally run Debug GPU
# Note: Batch normalization training on CPU is not yet implemented.
- nightly-e (build_sku=='gpu') and (device=='gpu')
testCases:
CNTK Run must be completed:
patterns:
- __COMPLETED__
Must train epochs in exactly same order and parameters:
patterns:
- Starting Epoch {{integer}}
- learning rate per sample = {{float}}
- momentum = {{float}}
Epochs (with low train loss) must be finished with expected results:
patterns:
- Finished Epoch[{{integer}} of {{integer}}]
- ce = {{float,tolerance=0.05}}
- err = {{float,tolerance=0.005}}
- totalSamplesSeen = {{integer}}
- learningRatePerSample = {{float,tolerance=0.1%}}
Per-minibatch (with low train loss) training results must match:
patterns:
- Epoch[{{integer}} of {{integer}}]-Minibatch[{{integer}}-{{integer}}
- ce = {{float,tolerance=0.05}} * {{integer}}
- err = {{float,tolerance=0.005}} * {{integer}}
Final test results must match:
patterns:
- "Final Results: Minibatch[{{integer}}-{{integer}}]"
- err = {{float,tolerance=0.005}} * {{integer}}
- ce = {{float,tolerance=0.05}} * {{integer}}

88
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/Config/01_OneHidden.cntk Normal file
Просмотреть файл

@ -0,0 +1,88 @@
# Parameters can be overwritten on the command line
# for example: cntk configFile=myConfigFile RootDir=../..
# For running from Visual Studio add
# currentDirectory=$(SolutionDir)/<path to corresponding data folder>
RootDir = ".."
ConfigDir = "$RootDir$/Config"
DataDir = "$RootDir$/Data"
OutputDir = "$RootDir$/Output"
ModelDir = "$OutputDir$/Models"
deviceId = 0
imageLayout = "cudnn"
# override the above as follows when running on CPU:
# deviceId = -1
command = train:test
precision = "float"
modelPath = "$ModelDir$/01_OneHidden"
ndlMacros = "$ConfigDir$/Macros.ndl"
# uncomment the following line to write logs to a file
# stderr = "$OutputDir$/01_OneHidden_out"
traceLevel=1
numMBsToShowResult=500
# If set to true, always initialize the network on CPU, making initialization consistent across CPU and GPU targets (for testing).
initOnCPUOnly=true
#######################################
# TRAINING CONFIG #
#######################################
train = [
action = "train"
NDLNetworkBuilder = [
networkDescription = "$ConfigDir$/01_OneHidden.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.1
momentumPerMB = 0
maxEpochs = 30
]
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
#######################################
# TEST CONFIG #
#######################################
test = [
action = "test"
minibatchSize = 16
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]

95
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/Config/02_Convolution.cntk Normal file
Просмотреть файл

@ -0,0 +1,95 @@
# Parameters can be overwritten on the command line
# for example: cntk configFile=myConfigFile RootDir=../..
# For running from Visual Studio add
# currentDirectory=$(SolutionDir)/<path to corresponding data folder>
RootDir = ".."
ConfigDir = "$RootDir$/Config"
DataDir = "$RootDir$/Data"
OutputDir = "$RootDir$/Output"
ModelDir = "$OutputDir$/Models"
deviceId = 0
imageLayout = "cudnn"
# Override the above as follows when running on CPU:
# deviceId = -1
# Note: Compared to GPU, this runs very slow.
command = train:test
precision = "float"
modelPath = "$ModelDir$/02_Convolution"
ndlMacros = "$ConfigDir$/Macros.ndl"
# uncomment the following line to write logs to a file
# stderr = "$OutputDir$/02_Convolution_out"
traceLevel=1
numMBsToShowResult=500
prefetch=true
# If set to true, always initialize the network on CPU, making initialization consistent across CPU and GPU targets (for testing).
initOnCPUOnly=true
#######################################
# TRAINING CONFIG #
#######################################
train = [
action = "train"
NDLNetworkBuilder = [
networkDescription = "$ConfigDir$/02_Convolution.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.1*5:0.3
momentumPerMB = 0*10:0.7
maxEpochs = 15
]
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
#######################################
# TEST CONFIG #
#######################################
test = [
action = test
minibatchSize = 16
NDLNetworkBuilder = [
networkDescription = "$ConfigDir$/02_Convolution.ndl"
]
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]

98
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/Config/03_ConvBatchNorm.cntk Normal file
Просмотреть файл

@ -0,0 +1,98 @@
# Parameters can be overwritten on the command line
# for example: cntk configFile=myConfigFile RootDir=../..
# For running from Visual Studio add
# currentDirectory=$(SolutionDir)/<path to corresponding data folder>
RootDir = ".."
ConfigDir = "$RootDir$/Config"
DataDir = "$RootDir$/Data"
OutputDir = "$RootDir$/Output"
ModelDir = "$OutputDir$/Models"
deviceId = 0
imageLayout = "cudnn"
# Note: Batch normalization training on CPU is not yet implemented.
# When it is, override the above as follows when running on CPU:
# deviceId = -1
command = train:test
precision = "float"
modelPath = "$ModelDir$/03_ConvBatchNorm"
ndlMacros = "$ConfigDir$/Macros.ndl"
# uncomment the following line to write logs to a file
# stderr = "$OutputDir$/03_ConvBatchNorm_out"
traceLevel=1
numMBsToShowResult=500
# If set to true, always initialize the network on CPU, making initialization consistent across CPU and GPU targets (for testing).
initOnCPUOnly=true
#######################################
# TRAINING CONFIG #
#######################################
train = [
action = "train"
NDLNetworkBuilder = [
networkDescription = "$ConfigDir$/03_ConvBatchNorm.ndl"
]
SGD = [
epochSize = 60000
minibatchSize = 32
learningRatesPerMB = 0.5:0.1
momentumPerMB = 0.9
maxEpochs = 2
#batchNormalizationTimeConstant=0 # Set through NDL
batchNormalizationBlendTimeConstant=0:1#INF
]
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Train-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
#######################################
# TEST CONFIG #
#######################################
test = [
action = "test"
minibatchSize = 32
modelPath=$ModelDir$/03_ConvBatchNorm
NDLNetworkBuilder = [
networkDescription = "$ConfigDir$/03_ConvBatchNorm.ndl"
]
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Test-28x28_cntk_text.txt"
input = [
features = [
dim = 784
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]

29
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/run-test-common Executable file
Просмотреть файл

@ -0,0 +1,29 @@
#!/bin/bash
. $TEST_ROOT_DIR/run-test-common
ConfigDir=$TEST_DIR/../../../../../../../Examples/Image/MNIST/Config
CNTKTextConfigDir=../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/MNIST/Config
if [[ ! -d $TEST_DATA_DIR || ! -e $TEST_DATA_DIR/Train-28x28_cntk_text.txt || ! -e $TEST_DATA_DIR/Test-28x28_cntk_text.txt ]]; then
# Cannot find test data locally.
# Try external test data directory (not part of the CNTK repository) as an alternative.
if [[ -d "$CNTK_EXTERNAL_TESTDATA_SOURCE_DIRECTORY" ]]; then
if [ "$OS" == "Windows_NT" ]; then
DataSourceDir=`cygpath -au $CNTK_EXTERNAL_TESTDATA_SOURCE_DIRECTORY`/Image/MNIST/v0
else
DataSourceDir=$CNTK_EXTERNAL_TESTDATA_SOURCE_DIRECTORY/Image/MNIST/v0
fi
# Copy the test data to the test run directory
DataDir=$TEST_RUN_DIR/TestData
mkdir $DataDir
cp -R $DataSourceDir/*_cntk_text.txt $DataDir || exit $?
Copied=1
else
echo Error: cannot find data. Please see Examples/Image/MNIST/README.md for instructions to get it.
exit 1
fi
fi
imageLayout=cudnn

13
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution/README.txt Normal file
Просмотреть файл

@ -0,0 +1,13 @@
Test runtimes
Windows:
Running test CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution (debug cpu) - [OK] 2250.09 sec
Running test CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution (debug gpu) - [OK] 135.24 sec
Running test CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution (release cpu) - [OK] 186.46 sec
Running test CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution (release gpu) - [OK] 25.65 sec
Linux:
Running test CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution (debug cpu) - [OK] 469.17 sec
Running test CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution (debug gpu) - [OK] 30.43 sec
Running test CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution (release cpu) - [OK] 29.58 sec
Running test CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution (release gpu) - [OK] 10.27 sec

605
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution/baseline.linux.txt Normal file
Просмотреть файл

@ -0,0 +1,605 @@
=== Running /home/philly/jenkins/workspace/CNTK-Test-Linux-W1/build/gpu/release/bin/cntk configFile=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10/../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/Config/01_Conv.cntk currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/TestData RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/TestData ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10 OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu DeviceId=0 timestamping=true Train=[SGD=[maxEpochs=10]] Train=[SGD=[epochSize=100]] stderr=-
-------------------------------------------------------------------
Build info:
Built time: May 13 2016 14:50:25
Last modified date: Thu May 12 14:00:37 2016
Build type: release
Build target: GPU
With 1bit-SGD: no
Math lib: acml
CUDA_PATH: /usr/local/cuda-7.5
CUB_PATH: /usr/local/cub-1.4.1
CUDNN_PATH: /usr/local/cudnn-4.0
Build Branch: HEAD
Build SHA1: 35fadc316f045d843bbd9b85061250a959268787
Built by philly on d8dc82703b0f
Build Path: /home/philly/jenkins/workspace/CNTK-Build-Linux
-------------------------------------------------------------------
Changed current directory to /tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/TestData
05/13/2016 15:10:47: Redirecting stderr to file -_Train_Test.log
05/13/2016 15:10:47: -------------------------------------------------------------------
05/13/2016 15:10:47: Build info:
05/13/2016 15:10:47: Built time: May 13 2016 14:50:25
05/13/2016 15:10:47: Last modified date: Thu May 12 14:00:37 2016
05/13/2016 15:10:47: Build type: release
05/13/2016 15:10:47: Build target: GPU
05/13/2016 15:10:47: With 1bit-SGD: no
05/13/2016 15:10:47: Math lib: acml
05/13/2016 15:10:47: CUDA_PATH: /usr/local/cuda-7.5
05/13/2016 15:10:47: CUB_PATH: /usr/local/cub-1.4.1
05/13/2016 15:10:47: CUDNN_PATH: /usr/local/cudnn-4.0
05/13/2016 15:10:47: Build Branch: HEAD
05/13/2016 15:10:47: Build SHA1: 35fadc316f045d843bbd9b85061250a959268787
05/13/2016 15:10:47: Built by philly on d8dc82703b0f
05/13/2016 15:10:47: Build Path: /home/philly/jenkins/workspace/CNTK-Build-Linux
05/13/2016 15:10:47: -------------------------------------------------------------------
05/13/2016 15:10:47: Running on localhost at 2016/05/13 15:10:47
05/13/2016 15:10:47: Command line:
/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/build/gpu/release/bin/cntk configFile=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10/../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/Config/01_Conv.cntk currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/TestData RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/TestData ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10 OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu DeviceId=0 timestamping=true Train=[SGD=[maxEpochs=10]] Train=[SGD=[epochSize=100]] stderr=-
05/13/2016 15:10:47: >>>>>>>>>>>>>>>>>>>> RAW CONFIG (VARIABLES NOT RESOLVED) >>>>>>>>>>>>>>>>>>>>
05/13/2016 15:10:47: RootDir = "."
ConfigDir = "$RootDir$"
DataDir = "$RootDir$"
OutputDir = "$RootDir$/Output"
ModelDir = "$OutputDir$/Models"
ndlMacros = "$ConfigDir$/Macros.ndl"
precision = "float"
deviceId = 0
imageLayout = "cudnn"
initOnCPUOnly=true
prefetch = "true"
command = Train:Test
modelPath = "$ModelDir$/01_Convolution"
stderr = "$OutputDir$/01_Conv"
traceLevel = 1
numMBsToShowResult = 500
Train = [
action = "train"
NDLNetworkBuilder = [
networkDescription = "$ConfigDir$/01_Convolution.ndl"
]
SGD = [
epochSize = 49984
minibatchSize = 64
learningRatesPerMB = 0.01*10:0.003*10:0.001
momentumPerMB = 0.9*20:0.99
maxEpochs = 30
L2RegWeight = 0.03
dropoutRate = 0*5:0.5
]
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Train_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
Test = [
action = "test"
minibatchSize = 16
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Test_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/TestData
RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu
DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/TestData
ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10
OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu
DeviceId=0
timestamping=true
Train=[SGD=[maxEpochs=10]]
Train=[SGD=[epochSize=100]]
stderr=-
05/13/2016 15:10:47: <<<<<<<<<<<<<<<<<<<< RAW CONFIG (VARIABLES NOT RESOLVED) <<<<<<<<<<<<<<<<<<<<
05/13/2016 15:10:47: >>>>>>>>>>>>>>>>>>>> RAW CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>
05/13/2016 15:10:47: RootDir = "."
ConfigDir = "."
DataDir = "."
OutputDir = "./Output"
ModelDir = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/Models"
ndlMacros = "/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10/Macros.ndl"
precision = "float"
deviceId = 0
imageLayout = "cudnn"
initOnCPUOnly=true
prefetch = "true"
command = Train:Test
modelPath = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution"
stderr = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/01_Conv"
traceLevel = 1
numMBsToShowResult = 500
Train = [
action = "train"
NDLNetworkBuilder = [
networkDescription = "/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10/01_Convolution.ndl"
]
SGD = [
epochSize = 49984
minibatchSize = 64
learningRatesPerMB = 0.01*10:0.003*10:0.001
momentumPerMB = 0.9*20:0.99
maxEpochs = 30
L2RegWeight = 0.03
dropoutRate = 0*5:0.5
]
reader = [
readerType = "CNTKTextFormatReader"
file = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/TestData/Train_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
Test = [
action = "test"
minibatchSize = 16
reader = [
readerType = "CNTKTextFormatReader"
file = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/TestData/Test_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/TestData
RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu
DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/TestData
ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10
OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu
DeviceId=0
timestamping=true
Train=[SGD=[maxEpochs=10]]
Train=[SGD=[epochSize=100]]
stderr=-
05/13/2016 15:10:47: <<<<<<<<<<<<<<<<<<<< RAW CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<
05/13/2016 15:10:47: >>>>>>>>>>>>>>>>>>>> PROCESSED CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>
configparameters: 01_Conv.cntk:command=Train:Test
configparameters: 01_Conv.cntk:ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10
configparameters: 01_Conv.cntk:currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/TestData
configparameters: 01_Conv.cntk:DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/TestData
configparameters: 01_Conv.cntk:deviceId=0
configparameters: 01_Conv.cntk:imageLayout=cudnn
configparameters: 01_Conv.cntk:initOnCPUOnly=true
configparameters: 01_Conv.cntk:ModelDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/Models
configparameters: 01_Conv.cntk:modelPath=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution
configparameters: 01_Conv.cntk:ndlMacros=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10/Macros.ndl
configparameters: 01_Conv.cntk:numMBsToShowResult=500
configparameters: 01_Conv.cntk:OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu
configparameters: 01_Conv.cntk:precision=float
configparameters: 01_Conv.cntk:prefetch=true
configparameters: 01_Conv.cntk:RootDir=.
configparameters: 01_Conv.cntk:RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu
configparameters: 01_Conv.cntk:stderr=-
configparameters: 01_Conv.cntk:Test=[
action = "test"
minibatchSize = 16
reader = [
readerType = "CNTKTextFormatReader"
file = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/TestData/Test_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
configparameters: 01_Conv.cntk:timestamping=true
configparameters: 01_Conv.cntk:traceLevel=1
configparameters: 01_Conv.cntk:Train=[
action = "train"
NDLNetworkBuilder = [
networkDescription = "/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10/01_Convolution.ndl"
]
SGD = [
epochSize = 49984
minibatchSize = 64
learningRatesPerMB = 0.01*10:0.003*10:0.001
momentumPerMB = 0.9*20:0.99
maxEpochs = 30
L2RegWeight = 0.03
dropoutRate = 0*5:0.5
]
reader = [
readerType = "CNTKTextFormatReader"
file = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/TestData/Train_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
] [SGD=[maxEpochs=10]] [SGD=[epochSize=100]]
05/13/2016 15:10:47: <<<<<<<<<<<<<<<<<<<< PROCESSED CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<
05/13/2016 15:10:47: Commands: Train Test
05/13/2016 15:10:47: Precision = "float"
05/13/2016 15:10:47: CNTKModelPath: /tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution
05/13/2016 15:10:47: CNTKCommandTrainInfo: Train : 10
05/13/2016 15:10:47: CNTKCommandTrainInfo: CNTKNoMoreCommands_Total : 10
05/13/2016 15:10:47: ##############################################################################
05/13/2016 15:10:47: # #
05/13/2016 15:10:47: # Action "train" #
05/13/2016 15:10:47: # #
05/13/2016 15:10:47: ##############################################################################
05/13/2016 15:10:47: CNTKCommandTrainBegin: Train
NDLBuilder Using GPU 0
05/13/2016 15:10:47: Creating virgin network.
SetGaussianRandomValue (GPU): creating curand object with seed 1, sizeof(ElemType)==4
Post-processing network...
3 roots:
CE = CrossEntropyWithSoftmax()
Err = ErrorPrediction()
OutputNodes.z = Plus()
Validating network. 34 nodes to process in pass 1.
Validating --> labels = InputValue() : -> [10 x *]
Validating --> OutputNodes.W = LearnableParameter() : -> [10 x 64]
Validating --> h1.W = LearnableParameter() : -> [64 x 3 x 3 x 64]
Validating --> conv3_act.W = LearnableParameter() : -> [64 x 800]
Validating --> conv2_act.W = LearnableParameter() : -> [32 x 800]
Validating --> conv1_act.W = LearnableParameter() : -> [32 x 75]
Validating --> features = InputValue() : -> [32 x 32 x 3 x *]
Validating --> featOffs = LearnableParameter() : -> [1 x 1]
Validating --> featScaled = Minus (features, featOffs) : [32 x 32 x 3 x *], [1 x 1] -> [32 x 32 x 3 x *]
Validating --> conv1_act.c = Convolution (conv1_act.W, featScaled) : [32 x 75], [32 x 32 x 3 x *] -> [32 x 32 x 32 x *]
Validating --> conv1_act.b = LearnableParameter() : -> [1 x 1 x 32]
Validating --> conv1_act.p = Plus (conv1_act.c, conv1_act.b) : [32 x 32 x 32 x *], [1 x 1 x 32] -> [32 x 32 x 32 x *]
Validating --> conv1_act.y = RectifiedLinear (conv1_act.p) : [32 x 32 x 32 x *] -> [32 x 32 x 32 x *]
Validating --> pool1 = MaxPooling (conv1_act.y) : [32 x 32 x 32 x *] -> [15 x 15 x 32 x *]
Validating --> conv2_act.c = Convolution (conv2_act.W, pool1) : [32 x 800], [15 x 15 x 32 x *] -> [15 x 15 x 32 x *]
Validating --> conv2_act.b = LearnableParameter() : -> [1 x 1 x 32]
Validating --> conv2_act.p = Plus (conv2_act.c, conv2_act.b) : [15 x 15 x 32 x *], [1 x 1 x 32] -> [15 x 15 x 32 x *]
Validating --> conv2_act.y = RectifiedLinear (conv2_act.p) : [15 x 15 x 32 x *] -> [15 x 15 x 32 x *]
Validating --> pool2 = MaxPooling (conv2_act.y) : [15 x 15 x 32 x *] -> [7 x 7 x 32 x *]
Validating --> conv3_act.c = Convolution (conv3_act.W, pool2) : [64 x 800], [7 x 7 x 32 x *] -> [7 x 7 x 64 x *]
Validating --> conv3_act.b = LearnableParameter() : -> [1 x 1 x 64]
Validating --> conv3_act.p = Plus (conv3_act.c, conv3_act.b) : [7 x 7 x 64 x *], [1 x 1 x 64] -> [7 x 7 x 64 x *]
Validating --> conv3_act.y = RectifiedLinear (conv3_act.p) : [7 x 7 x 64 x *] -> [7 x 7 x 64 x *]
Validating --> pool3 = MaxPooling (conv3_act.y) : [7 x 7 x 64 x *] -> [3 x 3 x 64 x *]
Validating --> h1.t = Times (h1.W, pool3) : [64 x 3 x 3 x 64], [3 x 3 x 64 x *] -> [64 x *]
Validating --> h1.b = LearnableParameter() : -> [64 x 1]
Validating --> h1.z = Plus (h1.t, h1.b) : [64 x *], [64 x 1] -> [64 x 1 x *]
Validating --> h1.y = RectifiedLinear (h1.z) : [64 x 1 x *] -> [64 x 1 x *]
Validating --> h1_d = Dropout (h1.y) : [64 x 1 x *] -> [64 x 1 x *]
Validating --> OutputNodes.t = Times (OutputNodes.W, h1_d) : [10 x 64], [64 x 1 x *] -> [10 x 1 x *]
Validating --> OutputNodes.b = LearnableParameter() : -> [10]
Validating --> OutputNodes.z = Plus (OutputNodes.t, OutputNodes.b) : [10 x 1 x *], [10] -> [10 x 1 x *]
Validating --> CE = CrossEntropyWithSoftmax (labels, OutputNodes.z) : [10 x *], [10 x 1 x *] -> [1]
Validating --> Err = ErrorPrediction (labels, OutputNodes.z) : [10 x *], [10 x 1 x *] -> [1]
Validating network. 21 nodes to process in pass 2.
Validating network, final pass.
Using cuDNN convolution engine for geometry: Input: 32 x 32 x 3, Output: 32 x 32 x 32, Kernel: 5 x 5 x 3, Map: 1 x 1 x 32, Stride: 1 x 1 x 3, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 32 x 32 x 32, Output: 15 x 15 x 32, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 15 x 15 x 32, Output: 15 x 15 x 32, Kernel: 5 x 5 x 32, Map: 1 x 1 x 32, Stride: 1 x 1 x 32, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 15 x 15 x 32, Output: 7 x 7 x 32, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 7 x 7 x 32, Output: 7 x 7 x 64, Kernel: 5 x 5 x 32, Map: 1 x 1 x 64, Stride: 1 x 1 x 32, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 7 x 7 x 64, Output: 3 x 3 x 64, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
13 out of 34 nodes do not share the minibatch layout with the input data.
Post-processing network complete.
05/13/2016 15:10:48: Created model with 34 nodes on GPU 0.
05/13/2016 15:10:48: Training criterion node(s):
05/13/2016 15:10:48: CE = CrossEntropyWithSoftmax
05/13/2016 15:10:48: Evaluation criterion node(s):
05/13/2016 15:10:48: Err = ErrorPrediction
Allocating matrices for forward and/or backward propagation.
Memory Sharing Structure:
(nil): {[Err Gradient[1]] [featOffs Gradient[1 x 1]] [featScaled Gradient[32 x 32 x 3 x *]] [features Gradient[32 x 32 x 3 x *]] [labels Gradient[10 x *]] }
0x2485d28: {[OutputNodes.z Value[10 x 1 x *]] }
0x2485ee8: {[CE Value[1]] }
0x2486168: {[conv1_act.W Gradient[32 x 75]] [conv1_act.p Value[32 x 32 x 32 x *]] }
0x2486328: {[conv1_act.c Gradient[32 x 32 x 32 x *]] [conv1_act.y Value[32 x 32 x 32 x *]] }
0x24864e8: {[conv1_act.p Gradient[32 x 32 x 32 x *]] [pool1 Value[15 x 15 x 32 x *]] }
0x249a638: {[features Value[32 x 32 x 3 x *]] }
0x2975298: {[conv1_act.b Value[1 x 1 x 32]] }
0x2976b48: {[conv2_act.W Value[32 x 800]] }
0x2977ae8: {[conv2_act.b Value[1 x 1 x 32]] }
0x2979668: {[conv3_act.W Value[64 x 800]] }
0x2979f08: {[conv3_act.b Value[1 x 1 x 64]] }
0x297bae8: {[h1.W Value[64 x 3 x 3 x 64]] }
0x297c538: {[h1.b Value[64 x 1]] }
0x297d5c8: {[OutputNodes.W Value[10 x 64]] }
0x297ea98: {[OutputNodes.b Value[10]] }
0x2dd1458: {[featOffs Value[1 x 1]] }
0x2dd2678: {[labels Value[10 x *]] }
0x2dd2eb8: {[conv1_act.W Value[32 x 75]] }
0x7a59dd8: {[Err Value[1]] }
0x7a5d378: {[featScaled Value[32 x 32 x 3 x *]] }
0x7a5d6d8: {[conv1_act.c Value[32 x 32 x 32 x *]] }
0x7a5e478: {[conv2_act.c Value[15 x 15 x 32 x *]] }
0x7a5e638: {[conv1_act.b Gradient[1 x 1 x 32]] [conv1_act.y Gradient[32 x 32 x 32 x *]] }
0x7a5e7f8: {[conv2_act.W Gradient[32 x 800]] [conv2_act.p Value[15 x 15 x 32 x *]] }
0x7a7ade8: {[conv2_act.c Gradient[15 x 15 x 32 x *]] [conv2_act.y Value[15 x 15 x 32 x *]] }
0x7a7afa8: {[conv2_act.p Gradient[15 x 15 x 32 x *]] [pool1 Gradient[15 x 15 x 32 x *]] [pool2 Value[7 x 7 x 32 x *]] }
0x7a7b168: {[conv3_act.c Value[7 x 7 x 64 x *]] }
0x7a7b328: {[conv2_act.b Gradient[1 x 1 x 32]] [conv2_act.y Gradient[15 x 15 x 32 x *]] }
0x7a7b4e8: {[conv3_act.W Gradient[64 x 800]] [conv3_act.p Value[7 x 7 x 64 x *]] }
0x7a7b6a8: {[conv3_act.c Gradient[7 x 7 x 64 x *]] [conv3_act.y Value[7 x 7 x 64 x *]] }
0x7a7b868: {[conv3_act.p Gradient[7 x 7 x 64 x *]] [pool2 Gradient[7 x 7 x 32 x *]] [pool3 Value[3 x 3 x 64 x *]] }
0x7a7ba28: {[conv3_act.b Gradient[1 x 1 x 64]] [conv3_act.y Gradient[7 x 7 x 64 x *]] [h1.t Value[64 x *]] }
0x7a7bbe8: {[h1.W Gradient[64 x 3 x 3 x 64]] [h1.z Value[64 x 1 x *]] }
0x7a7bda8: {[h1.t Gradient[64 x *]] [h1.y Value[64 x 1 x *]] }
0x7a7bf68: {[h1_d Value[64 x 1 x *]] }
0x7a7c128: {[h1.z Gradient[64 x 1 x *]] [pool3 Gradient[3 x 3 x 64 x *]] }
0x7a7c2e8: {[OutputNodes.t Value[10 x 1 x *]] [h1.b Gradient[64 x 1]] [h1.y Gradient[64 x 1 x *]] }
0x7a7cdc8: {[CE Gradient[1]] }
0x7a7cf88: {[OutputNodes.W Gradient[10 x 64]] [OutputNodes.z Gradient[10 x 1 x *]] }
0x7a7d148: {[OutputNodes.t Gradient[10 x 1 x *]] }
0x7a7d308: {[OutputNodes.b Gradient[10]] }
0x7a7d4c8: {[h1_d Gradient[64 x 1 x *]] }
05/13/2016 15:10:48: No PreCompute nodes found, skipping PreCompute step.
05/13/2016 15:10:48: Starting Epoch 1: learning rate per sample = 0.000156 effective momentum = 0.900000 momentum as time constant = 607.4 samples
05/13/2016 15:10:48: Starting minibatch loop.
05/13/2016 15:10:51: Finished Epoch[ 1 of 10]: [Training] CE = 2.30242050 * 100; Err = 0.88000000 * 100; totalSamplesSeen = 100; learningRatePerSample = 0.00015625; epochTime=3.55904s
05/13/2016 15:10:51: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution.1'
05/13/2016 15:10:51: Starting Epoch 2: learning rate per sample = 0.000156 effective momentum = 0.900000 momentum as time constant = 607.4 samples
05/13/2016 15:10:51: Starting minibatch loop.
05/13/2016 15:10:51: Finished Epoch[ 2 of 10]: [Training] CE = 2.30175842 * 100; Err = 0.94000000 * 100; totalSamplesSeen = 200; learningRatePerSample = 0.00015625; epochTime=0.011903s
05/13/2016 15:10:51: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution.2'
05/13/2016 15:10:51: Starting Epoch 3: learning rate per sample = 0.000156 effective momentum = 0.900000 momentum as time constant = 607.4 samples
05/13/2016 15:10:51: Starting minibatch loop.
05/13/2016 15:10:51: Finished Epoch[ 3 of 10]: [Training] CE = 2.30054413 * 100; Err = 0.90000000 * 100; totalSamplesSeen = 300; learningRatePerSample = 0.00015625; epochTime=0.012701s
05/13/2016 15:10:51: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution.3'
05/13/2016 15:10:51: Starting Epoch 4: learning rate per sample = 0.000156 effective momentum = 0.900000 momentum as time constant = 607.4 samples
05/13/2016 15:10:51: Starting minibatch loop.
05/13/2016 15:10:51: Finished Epoch[ 4 of 10]: [Training] CE = 2.30022812 * 100; Err = 0.88000000 * 100; totalSamplesSeen = 400; learningRatePerSample = 0.00015625; epochTime=0.01144s
05/13/2016 15:10:51: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution.4'
05/13/2016 15:10:51: Starting Epoch 5: learning rate per sample = 0.000156 effective momentum = 0.900000 momentum as time constant = 607.4 samples
05/13/2016 15:10:51: Starting minibatch loop.
05/13/2016 15:10:51: Finished Epoch[ 5 of 10]: [Training] CE = 2.29579636 * 100; Err = 0.87000000 * 100; totalSamplesSeen = 500; learningRatePerSample = 0.00015625; epochTime=0.011529s
05/13/2016 15:10:51: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution.5'
Setting dropout rate to 0.5.
05/13/2016 15:10:51: Starting Epoch 6: learning rate per sample = 0.000156 effective momentum = 0.900000 momentum as time constant = 607.4 samples
05/13/2016 15:10:51: Starting minibatch loop.
(GPU): creating curand object with seed 5
05/13/2016 15:10:51: Finished Epoch[ 6 of 10]: [Training] CE = 2.30121231 * 100; Err = 0.84000000 * 100; totalSamplesSeen = 600; learningRatePerSample = 0.00015625; epochTime=0.012276s
05/13/2016 15:10:51: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution.6'
05/13/2016 15:10:51: Starting Epoch 7: learning rate per sample = 0.000156 effective momentum = 0.900000 momentum as time constant = 607.4 samples
05/13/2016 15:10:51: Starting minibatch loop.
(GPU): creating curand object with seed 6
05/13/2016 15:10:52: Finished Epoch[ 7 of 10]: [Training] CE = 2.28975647 * 100; Err = 0.93000000 * 100; totalSamplesSeen = 700; learningRatePerSample = 0.00015625; epochTime=0.011495s
05/13/2016 15:10:52: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution.7'
05/13/2016 15:10:52: Starting Epoch 8: learning rate per sample = 0.000156 effective momentum = 0.900000 momentum as time constant = 607.4 samples
05/13/2016 15:10:52: Starting minibatch loop.
(GPU): creating curand object with seed 7
05/13/2016 15:10:52: Finished Epoch[ 8 of 10]: [Training] CE = 2.29035095 * 100; Err = 0.91000000 * 100; totalSamplesSeen = 800; learningRatePerSample = 0.00015625; epochTime=0.012157s
05/13/2016 15:10:52: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution.8'
05/13/2016 15:10:52: Starting Epoch 9: learning rate per sample = 0.000156 effective momentum = 0.900000 momentum as time constant = 607.4 samples
05/13/2016 15:10:52: Starting minibatch loop.
(GPU): creating curand object with seed 8
05/13/2016 15:10:52: Finished Epoch[ 9 of 10]: [Training] CE = 2.29797729 * 100; Err = 0.87000000 * 100; totalSamplesSeen = 900; learningRatePerSample = 0.00015625; epochTime=0.011451s
05/13/2016 15:10:52: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution.9'
05/13/2016 15:10:52: Starting Epoch 10: learning rate per sample = 0.000156 effective momentum = 0.900000 momentum as time constant = 607.4 samples
05/13/2016 15:10:52: Starting minibatch loop.
(GPU): creating curand object with seed 9
05/13/2016 15:10:52: Finished Epoch[10 of 10]: [Training] CE = 2.29764435 * 100; Err = 0.87000000 * 100; totalSamplesSeen = 1000; learningRatePerSample = 0.00015625; epochTime=0.012689s
05/13/2016 15:10:52: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution'
05/13/2016 15:10:52: CNTKCommandTrainEnd: Train
05/13/2016 15:10:52: Action "train" complete.
05/13/2016 15:10:52: ##############################################################################
05/13/2016 15:10:52: # #
05/13/2016 15:10:52: # Action "test" #
05/13/2016 15:10:52: # #
05/13/2016 15:10:52: ##############################################################################
Post-processing network...
3 roots:
CE = CrossEntropyWithSoftmax()
Err = ErrorPrediction()
OutputNodes.z = Plus()
Validating network. 34 nodes to process in pass 1.
Validating --> labels = InputValue() : -> [10 x *1]
Validating --> OutputNodes.W = LearnableParameter() : -> [10 x 64]
Validating --> h1.W = LearnableParameter() : -> [64 x 3 x 3 x 64]
Validating --> conv3_act.W = LearnableParameter() : -> [64 x 800]
Validating --> conv2_act.W = LearnableParameter() : -> [32 x 800]
Validating --> conv1_act.W = LearnableParameter() : -> [32 x 75]
Validating --> features = InputValue() : -> [32 x 32 x 3 x *1]
Validating --> featOffs = LearnableParameter() : -> [1 x 1]
Validating --> featScaled = Minus (features, featOffs) : [32 x 32 x 3 x *1], [1 x 1] -> [32 x 32 x 3 x *1]
Validating --> conv1_act.c = Convolution (conv1_act.W, featScaled) : [32 x 75], [32 x 32 x 3 x *1] -> [32 x 32 x 32 x *1]
Validating --> conv1_act.b = LearnableParameter() : -> [1 x 1 x 32]
Validating --> conv1_act.p = Plus (conv1_act.c, conv1_act.b) : [32 x 32 x 32 x *1], [1 x 1 x 32] -> [32 x 32 x 32 x *1]
Validating --> conv1_act.y = RectifiedLinear (conv1_act.p) : [32 x 32 x 32 x *1] -> [32 x 32 x 32 x *1]
Validating --> pool1 = MaxPooling (conv1_act.y) : [32 x 32 x 32 x *1] -> [15 x 15 x 32 x *1]
Validating --> conv2_act.c = Convolution (conv2_act.W, pool1) : [32 x 800], [15 x 15 x 32 x *1] -> [15 x 15 x 32 x *1]
Validating --> conv2_act.b = LearnableParameter() : -> [1 x 1 x 32]
Validating --> conv2_act.p = Plus (conv2_act.c, conv2_act.b) : [15 x 15 x 32 x *1], [1 x 1 x 32] -> [15 x 15 x 32 x *1]
Validating --> conv2_act.y = RectifiedLinear (conv2_act.p) : [15 x 15 x 32 x *1] -> [15 x 15 x 32 x *1]
Validating --> pool2 = MaxPooling (conv2_act.y) : [15 x 15 x 32 x *1] -> [7 x 7 x 32 x *1]
Validating --> conv3_act.c = Convolution (conv3_act.W, pool2) : [64 x 800], [7 x 7 x 32 x *1] -> [7 x 7 x 64 x *1]
Validating --> conv3_act.b = LearnableParameter() : -> [1 x 1 x 64]
Validating --> conv3_act.p = Plus (conv3_act.c, conv3_act.b) : [7 x 7 x 64 x *1], [1 x 1 x 64] -> [7 x 7 x 64 x *1]
Validating --> conv3_act.y = RectifiedLinear (conv3_act.p) : [7 x 7 x 64 x *1] -> [7 x 7 x 64 x *1]
Validating --> pool3 = MaxPooling (conv3_act.y) : [7 x 7 x 64 x *1] -> [3 x 3 x 64 x *1]
Validating --> h1.t = Times (h1.W, pool3) : [64 x 3 x 3 x 64], [3 x 3 x 64 x *1] -> [64 x *1]
Validating --> h1.b = LearnableParameter() : -> [64 x 1]
Validating --> h1.z = Plus (h1.t, h1.b) : [64 x *1], [64 x 1] -> [64 x 1 x *1]
Validating --> h1.y = RectifiedLinear (h1.z) : [64 x 1 x *1] -> [64 x 1 x *1]
Validating --> h1_d = Dropout (h1.y) : [64 x 1 x *1] -> [64 x 1 x *1]
Validating --> OutputNodes.t = Times (OutputNodes.W, h1_d) : [10 x 64], [64 x 1 x *1] -> [10 x 1 x *1]
Validating --> OutputNodes.b = LearnableParameter() : -> [10]
Validating --> OutputNodes.z = Plus (OutputNodes.t, OutputNodes.b) : [10 x 1 x *1], [10] -> [10 x 1 x *1]
Validating --> CE = CrossEntropyWithSoftmax (labels, OutputNodes.z) : [10 x *1], [10 x 1 x *1] -> [1]
Validating --> Err = ErrorPrediction (labels, OutputNodes.z) : [10 x *1], [10 x 1 x *1] -> [1]
Validating network. 21 nodes to process in pass 2.
Validating network, final pass.
Using cuDNN convolution engine for geometry: Input: 32 x 32 x 3, Output: 32 x 32 x 32, Kernel: 5 x 5 x 3, Map: 1 x 1 x 32, Stride: 1 x 1 x 3, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 32 x 32 x 32, Output: 15 x 15 x 32, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 15 x 15 x 32, Output: 15 x 15 x 32, Kernel: 5 x 5 x 32, Map: 1 x 1 x 32, Stride: 1 x 1 x 32, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 15 x 15 x 32, Output: 7 x 7 x 32, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 7 x 7 x 32, Output: 7 x 7 x 64, Kernel: 5 x 5 x 32, Map: 1 x 1 x 64, Stride: 1 x 1 x 32, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 7 x 7 x 64, Output: 3 x 3 x 64, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
13 out of 34 nodes do not share the minibatch layout with the input data.
Post-processing network complete.
evalNodeNames are not specified, using all the default evalnodes and training criterion nodes.
Allocating matrices for forward and/or backward propagation.
Memory Sharing Structure:
(nil): {[CE Gradient[1]] [Err Gradient[1]] [OutputNodes.W Gradient[10 x 64]] [OutputNodes.b Gradient[10]] [OutputNodes.t Gradient[10 x 1 x *1]] [OutputNodes.z Gradient[10 x 1 x *1]] [conv1_act.W Gradient[32 x 75]] [conv1_act.b Gradient[1 x 1 x 32]] [conv1_act.c Gradient[32 x 32 x 32 x *1]] [conv1_act.p Gradient[32 x 32 x 32 x *1]] [conv1_act.y Gradient[32 x 32 x 32 x *1]] [conv2_act.W Gradient[32 x 800]] [conv2_act.b Gradient[1 x 1 x 32]] [conv2_act.c Gradient[15 x 15 x 32 x *1]] [conv2_act.p Gradient[15 x 15 x 32 x *1]] [conv2_act.y Gradient[15 x 15 x 32 x *1]] [conv3_act.W Gradient[64 x 800]] [conv3_act.b Gradient[1 x 1 x 64]] [conv3_act.c Gradient[7 x 7 x 64 x *1]] [conv3_act.p Gradient[7 x 7 x 64 x *1]] [conv3_act.y Gradient[7 x 7 x 64 x *1]] [featOffs Gradient[1 x 1]] [featScaled Gradient[32 x 32 x 3 x *1]] [features Gradient[32 x 32 x 3 x *1]] [h1.W Gradient[64 x 3 x 3 x 64]] [h1.b Gradient[64 x 1]] [h1.t Gradient[64 x *1]] [h1.y Gradient[64 x 1 x *1]] [h1.z Gradient[64 x 1 x *1]] [h1_d Gradient[64 x 1 x *1]] [labels Gradient[10 x *1]] [pool1 Gradient[15 x 15 x 32 x *1]] [pool2 Gradient[7 x 7 x 32 x *1]] [pool3 Gradient[3 x 3 x 64 x *1]] }
0x7fc883e04ba8: {[conv1_act.b Value[1 x 1 x 32]] }
0x7fc883e05fc8: {[conv1_act.W Value[32 x 75]] }
0x7fc883e06768: {[conv2_act.b Value[1 x 1 x 32]] }
0x7fc883e06928: {[conv2_act.W Value[32 x 800]] }
0x7fc883e085b8: {[conv3_act.b Value[1 x 1 x 64]] }
0x7fc883e09528: {[conv3_act.W Value[64 x 800]] }
0x7fc883e0b568: {[featOffs Value[1 x 1]] }
0x7fc883e0c1e8: {[features Value[32 x 32 x 3 x *1]] }
0x7fc883e0cc38: {[h1.b Value[64 x 1]] }
0x7fc883e0cf08: {[h1.W Value[64 x 3 x 3 x 64]] }
0x7fc883e0eb48: {[labels Value[10 x *1]] }
0x7fc883e0f558: {[OutputNodes.b Value[10]] }
0x7fc883e10068: {[OutputNodes.W Value[10 x 64]] }
0x7fc883e286b8: {[Err Value[1]] }
0x7fc883e2bd28: {[CE Value[1]] }
0x7fc883e2bfa8: {[conv1_act.y Value[32 x 32 x 32 x *1]] }
0x7fc883e54728: {[conv1_act.c Value[32 x 32 x 32 x *1]] }
0x7fc883e54a88: {[featScaled Value[32 x 32 x 3 x *1]] }
0x7fc883e54c18: {[conv1_act.p Value[32 x 32 x 32 x *1]] }
0x7fc883e71a78: {[pool1 Value[15 x 15 x 32 x *1]] }
0x7fc883e71c38: {[conv2_act.c Value[15 x 15 x 32 x *1]] }
0x7fc883e71fb8: {[conv2_act.p Value[15 x 15 x 32 x *1]] }
0x7fc883e72178: {[conv2_act.y Value[15 x 15 x 32 x *1]] }
0x7fc883e72338: {[pool2 Value[7 x 7 x 32 x *1]] }
0x7fc883e724f8: {[conv3_act.c Value[7 x 7 x 64 x *1]] }
0x7fc883e72878: {[conv3_act.p Value[7 x 7 x 64 x *1]] }
0x7fc883e72a38: {[conv3_act.y Value[7 x 7 x 64 x *1]] }
0x7fc883e72bf8: {[pool3 Value[3 x 3 x 64 x *1]] }
0x7fc883e72db8: {[h1.t Value[64 x *1]] }
0x7fc883e72f78: {[h1.z Value[64 x 1 x *1]] }
0x7fc883e73138: {[h1.y Value[64 x 1 x *1]] }
0x7fc883e732f8: {[h1_d Value[64 x 1 x *1]] }
0x7fc883e73678: {[OutputNodes.t Value[10 x 1 x *1]] }
0x7fc883e73838: {[OutputNodes.z Value[10 x 1 x *1]] }
05/13/2016 15:10:58: Final Results: Minibatch[1-625]: Err = 0.86430000 * 10000; CE = 2.28476029 * 10000; perplexity = 9.82333117
05/13/2016 15:10:58: Action "test" complete.
05/13/2016 15:10:58: __COMPLETED__

603
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution/baseline.windows.txt Normal file
Просмотреть файл

@ -0,0 +1,603 @@
=== Running /cygdrive/c/jenkins/workspace/CNTK-Test-Windows-W1/x64/release/cntk.exe configFile=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10/../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/Config/01_Conv.cntk currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu\TestData RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu\TestData ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10 OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu DeviceId=0 timestamping=true Train=[SGD=[maxEpochs=10]] Train=[SGD=[epochSize=100]] stderr=-
-------------------------------------------------------------------
Build info:
Built time: May 13 2016 08:06:01
Last modified date: Thu May 12 07:31:50 2016
Build type: Release
Build target: GPU
With 1bit-SGD: no
CUDA_PATH: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5
CUB_PATH: c:\src\cub-1.4.1
CUDNN_PATH: c:\NVIDIA\cudnn-4.0\cuda
Build Branch: HEAD
Build SHA1: 35fadc316f045d843bbd9b85061250a959268787
Built by svcphil on Philly-Pool3
Build Path: c:\Jenkins\workspace\CNTK-Build-Windows\Source\CNTK\
-------------------------------------------------------------------
Changed current directory to C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu\TestData
05/13/2016 08:17:50: Redirecting stderr to file -_Train_Test.log
05/13/2016 08:17:50: -------------------------------------------------------------------
05/13/2016 08:17:50: Build info:
05/13/2016 08:17:50: Built time: May 13 2016 08:06:01
05/13/2016 08:17:50: Last modified date: Thu May 12 07:31:50 2016
05/13/2016 08:17:50: Build type: Release
05/13/2016 08:17:50: Build target: GPU
05/13/2016 08:17:50: With 1bit-SGD: no
05/13/2016 08:17:50: CUDA_PATH: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5
05/13/2016 08:17:50: CUB_PATH: c:\src\cub-1.4.1
05/13/2016 08:17:50: CUDNN_PATH: c:\NVIDIA\cudnn-4.0\cuda
05/13/2016 08:17:50: Build Branch: HEAD
05/13/2016 08:17:50: Build SHA1: 35fadc316f045d843bbd9b85061250a959268787
05/13/2016 08:17:50: Built by svcphil on Philly-Pool3
05/13/2016 08:17:50: Build Path: c:\Jenkins\workspace\CNTK-Build-Windows\Source\CNTK\
05/13/2016 08:17:50: -------------------------------------------------------------------
05/13/2016 08:17:50: Running on Philly-Pool2 at 2016/05/13 08:17:50
05/13/2016 08:17:50: Command line:
C:\jenkins\workspace\CNTK-Test-Windows-W1\x64\release\cntk.exe configFile=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10/../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/Config/01_Conv.cntk currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu\TestData RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu\TestData ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10 OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu DeviceId=0 timestamping=true Train=[SGD=[maxEpochs=10]] Train=[SGD=[epochSize=100]] stderr=-
05/13/2016 08:17:50: >>>>>>>>>>>>>>>>>>>> RAW CONFIG (VARIABLES NOT RESOLVED) >>>>>>>>>>>>>>>>>>>>
05/13/2016 08:17:50: RootDir = "."
ConfigDir = "$RootDir$"
DataDir = "$RootDir$"
OutputDir = "$RootDir$/Output"
ModelDir = "$OutputDir$/Models"
ndlMacros = "$ConfigDir$/Macros.ndl"
precision = "float"
deviceId = 0
imageLayout = "cudnn"
initOnCPUOnly=true
prefetch = "true"
command = Train:Test
modelPath = "$ModelDir$/01_Convolution"
stderr = "$OutputDir$/01_Conv"
traceLevel = 1
numMBsToShowResult = 500
Train = [
action = "train"
NDLNetworkBuilder = [
networkDescription = "$ConfigDir$/01_Convolution.ndl"
]
SGD = [
epochSize = 49984
minibatchSize = 64
learningRatesPerMB = 0.01*10:0.003*10:0.001
momentumPerMB = 0.9*20:0.99
maxEpochs = 30
L2RegWeight = 0.03
dropoutRate = 0*5:0.5
]
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Train_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
Test = [
action = "test"
minibatchSize = 16
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Test_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu\TestData
RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu
DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu\TestData
ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10
OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu
DeviceId=0
timestamping=true
Train=[SGD=[maxEpochs=10]]
Train=[SGD=[epochSize=100]]
stderr=-
05/13/2016 08:17:50: <<<<<<<<<<<<<<<<<<<< RAW CONFIG (VARIABLES NOT RESOLVED) <<<<<<<<<<<<<<<<<<<<
05/13/2016 08:17:50: >>>>>>>>>>>>>>>>>>>> RAW CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>
05/13/2016 08:17:50: RootDir = "."
ConfigDir = "."
DataDir = "."
OutputDir = "./Output"
ModelDir = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu/Models"
ndlMacros = "C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10/Macros.ndl"
precision = "float"
deviceId = 0
imageLayout = "cudnn"
initOnCPUOnly=true
prefetch = "true"
command = Train:Test
modelPath = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution"
stderr = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu/01_Conv"
traceLevel = 1
numMBsToShowResult = 500
Train = [
action = "train"
NDLNetworkBuilder = [
networkDescription = "C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10/01_Convolution.ndl"
]
SGD = [
epochSize = 49984
minibatchSize = 64
learningRatesPerMB = 0.01*10:0.003*10:0.001
momentumPerMB = 0.9*20:0.99
maxEpochs = 30
L2RegWeight = 0.03
dropoutRate = 0*5:0.5
]
reader = [
readerType = "CNTKTextFormatReader"
file = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu\TestData/Train_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
Test = [
action = "test"
minibatchSize = 16
reader = [
readerType = "CNTKTextFormatReader"
file = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu\TestData/Test_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu\TestData
RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu
DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu\TestData
ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10
OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu
DeviceId=0
timestamping=true
Train=[SGD=[maxEpochs=10]]
Train=[SGD=[epochSize=100]]
stderr=-
05/13/2016 08:17:50: <<<<<<<<<<<<<<<<<<<< RAW CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<
05/13/2016 08:17:50: >>>>>>>>>>>>>>>>>>>> PROCESSED CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>
configparameters: 01_Conv.cntk:command=Train:Test
configparameters: 01_Conv.cntk:ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10
configparameters: 01_Conv.cntk:currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu\TestData
configparameters: 01_Conv.cntk:DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu\TestData
configparameters: 01_Conv.cntk:deviceId=0
configparameters: 01_Conv.cntk:imageLayout=cudnn
configparameters: 01_Conv.cntk:initOnCPUOnly=true
configparameters: 01_Conv.cntk:ModelDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu/Models
configparameters: 01_Conv.cntk:modelPath=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution
configparameters: 01_Conv.cntk:ndlMacros=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10/Macros.ndl
configparameters: 01_Conv.cntk:numMBsToShowResult=500
configparameters: 01_Conv.cntk:OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu
configparameters: 01_Conv.cntk:precision=float
configparameters: 01_Conv.cntk:prefetch=true
configparameters: 01_Conv.cntk:RootDir=.
configparameters: 01_Conv.cntk:RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu
configparameters: 01_Conv.cntk:stderr=-
configparameters: 01_Conv.cntk:Test=[
action = "test"
minibatchSize = 16
reader = [
readerType = "CNTKTextFormatReader"
file = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu\TestData/Test_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
configparameters: 01_Conv.cntk:timestamping=true
configparameters: 01_Conv.cntk:traceLevel=1
configparameters: 01_Conv.cntk:Train=[
action = "train"
NDLNetworkBuilder = [
networkDescription = "C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10/01_Convolution.ndl"
]
SGD = [
epochSize = 49984
minibatchSize = 64
learningRatesPerMB = 0.01*10:0.003*10:0.001
momentumPerMB = 0.9*20:0.99
maxEpochs = 30
L2RegWeight = 0.03
dropoutRate = 0*5:0.5
]
reader = [
readerType = "CNTKTextFormatReader"
file = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu\TestData/Train_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
] [SGD=[maxEpochs=10]] [SGD=[epochSize=100]]
05/13/2016 08:17:50: <<<<<<<<<<<<<<<<<<<< PROCESSED CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<
05/13/2016 08:17:50: Commands: Train Test
05/13/2016 08:17:50: Precision = "float"
05/13/2016 08:17:50: CNTKModelPath: C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution
05/13/2016 08:17:50: CNTKCommandTrainInfo: Train : 10
05/13/2016 08:17:50: CNTKCommandTrainInfo: CNTKNoMoreCommands_Total : 10
05/13/2016 08:17:50: ##############################################################################
05/13/2016 08:17:50: # #
05/13/2016 08:17:50: # Action "train" #
05/13/2016 08:17:50: # #
05/13/2016 08:17:50: ##############################################################################
05/13/2016 08:17:50: CNTKCommandTrainBegin: Train
NDLBuilder Using GPU 0
05/13/2016 08:17:52: Creating virgin network.
Microsoft::MSR::CNTK::GPUMatrix<ElemType>::SetGaussianRandomValue (GPU): creating curand object with seed 1, sizeof(ElemType)==4
Post-processing network...
3 roots:
CE = CrossEntropyWithSoftmax()
Err = ErrorPrediction()
OutputNodes.z = Plus()
Validating network. 34 nodes to process in pass 1.
Validating --> labels = InputValue() : -> [10 x *]
Validating --> OutputNodes.W = LearnableParameter() : -> [10 x 64]
Validating --> h1.W = LearnableParameter() : -> [64 x 3 x 3 x 64]
Validating --> conv3_act.W = LearnableParameter() : -> [64 x 800]
Validating --> conv2_act.W = LearnableParameter() : -> [32 x 800]
Validating --> conv1_act.W = LearnableParameter() : -> [32 x 75]
Validating --> features = InputValue() : -> [32 x 32 x 3 x *]
Validating --> featOffs = LearnableParameter() : -> [1 x 1]
Validating --> featScaled = Minus (features, featOffs) : [32 x 32 x 3 x *], [1 x 1] -> [32 x 32 x 3 x *]
Validating --> conv1_act.c = Convolution (conv1_act.W, featScaled) : [32 x 75], [32 x 32 x 3 x *] -> [32 x 32 x 32 x *]
Validating --> conv1_act.b = LearnableParameter() : -> [1 x 1 x 32]
Validating --> conv1_act.p = Plus (conv1_act.c, conv1_act.b) : [32 x 32 x 32 x *], [1 x 1 x 32] -> [32 x 32 x 32 x *]
Validating --> conv1_act.y = RectifiedLinear (conv1_act.p) : [32 x 32 x 32 x *] -> [32 x 32 x 32 x *]
Validating --> pool1 = MaxPooling (conv1_act.y) : [32 x 32 x 32 x *] -> [15 x 15 x 32 x *]
Validating --> conv2_act.c = Convolution (conv2_act.W, pool1) : [32 x 800], [15 x 15 x 32 x *] -> [15 x 15 x 32 x *]
Validating --> conv2_act.b = LearnableParameter() : -> [1 x 1 x 32]
Validating --> conv2_act.p = Plus (conv2_act.c, conv2_act.b) : [15 x 15 x 32 x *], [1 x 1 x 32] -> [15 x 15 x 32 x *]
Validating --> conv2_act.y = RectifiedLinear (conv2_act.p) : [15 x 15 x 32 x *] -> [15 x 15 x 32 x *]
Validating --> pool2 = MaxPooling (conv2_act.y) : [15 x 15 x 32 x *] -> [7 x 7 x 32 x *]
Validating --> conv3_act.c = Convolution (conv3_act.W, pool2) : [64 x 800], [7 x 7 x 32 x *] -> [7 x 7 x 64 x *]
Validating --> conv3_act.b = LearnableParameter() : -> [1 x 1 x 64]
Validating --> conv3_act.p = Plus (conv3_act.c, conv3_act.b) : [7 x 7 x 64 x *], [1 x 1 x 64] -> [7 x 7 x 64 x *]
Validating --> conv3_act.y = RectifiedLinear (conv3_act.p) : [7 x 7 x 64 x *] -> [7 x 7 x 64 x *]
Validating --> pool3 = MaxPooling (conv3_act.y) : [7 x 7 x 64 x *] -> [3 x 3 x 64 x *]
Validating --> h1.t = Times (h1.W, pool3) : [64 x 3 x 3 x 64], [3 x 3 x 64 x *] -> [64 x *]
Validating --> h1.b = LearnableParameter() : -> [64 x 1]
Validating --> h1.z = Plus (h1.t, h1.b) : [64 x *], [64 x 1] -> [64 x 1 x *]
Validating --> h1.y = RectifiedLinear (h1.z) : [64 x 1 x *] -> [64 x 1 x *]
Validating --> h1_d = Dropout (h1.y) : [64 x 1 x *] -> [64 x 1 x *]
Validating --> OutputNodes.t = Times (OutputNodes.W, h1_d) : [10 x 64], [64 x 1 x *] -> [10 x 1 x *]
Validating --> OutputNodes.b = LearnableParameter() : -> [10]
Validating --> OutputNodes.z = Plus (OutputNodes.t, OutputNodes.b) : [10 x 1 x *], [10] -> [10 x 1 x *]
Validating --> CE = CrossEntropyWithSoftmax (labels, OutputNodes.z) : [10 x *], [10 x 1 x *] -> [1]
Validating --> Err = ErrorPrediction (labels, OutputNodes.z) : [10 x *], [10 x 1 x *] -> [1]
Validating network. 21 nodes to process in pass 2.
Validating network, final pass.
Using cuDNN convolution engine for geometry: Input: 32 x 32 x 3, Output: 32 x 32 x 32, Kernel: 5 x 5 x 3, Map: 1 x 1 x 32, Stride: 1 x 1 x 3, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 32 x 32 x 32, Output: 15 x 15 x 32, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 15 x 15 x 32, Output: 15 x 15 x 32, Kernel: 5 x 5 x 32, Map: 1 x 1 x 32, Stride: 1 x 1 x 32, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 15 x 15 x 32, Output: 7 x 7 x 32, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 7 x 7 x 32, Output: 7 x 7 x 64, Kernel: 5 x 5 x 32, Map: 1 x 1 x 64, Stride: 1 x 1 x 32, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 7 x 7 x 64, Output: 3 x 3 x 64, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
13 out of 34 nodes do not share the minibatch layout with the input data.
Post-processing network complete.
05/13/2016 08:17:53: Created model with 34 nodes on GPU 0.
05/13/2016 08:17:53: Training criterion node(s):
05/13/2016 08:17:53: CE = CrossEntropyWithSoftmax
05/13/2016 08:17:53: Evaluation criterion node(s):
05/13/2016 08:17:53: Err = ErrorPrediction
Allocating matrices for forward and/or backward propagation.
Memory Sharing Structure:
0000000000000000: {[Err Gradient[1]] [featOffs Gradient[1 x 1]] [featScaled Gradient[32 x 32 x 3 x *]] [features Gradient[32 x 32 x 3 x *]] [labels Gradient[10 x *]] }
0000004790872560: {[features Value[32 x 32 x 3 x *]] }
00000047A69D0CD0: {[h1.W Value[64 x 3 x 3 x 64]] }
00000047A69D1130: {[conv1_act.b Value[1 x 1 x 32]] }
00000047A69D11D0: {[h1.b Value[64 x 1]] }
00000047A69D1450: {[conv1_act.W Value[32 x 75]] }
00000047A69D14F0: {[conv2_act.b Value[1 x 1 x 32]] }
00000047A69D1810: {[featOffs Value[1 x 1]] }
00000047A69D1950: {[OutputNodes.W Value[10 x 64]] }
00000047A69D1C70: {[conv3_act.W Value[64 x 800]] }
00000047A69D1D10: {[OutputNodes.b Value[10]] }
00000047A69D1F90: {[conv3_act.b Value[1 x 1 x 64]] }
00000047A69D2490: {[labels Value[10 x *]] }
00000047A69D2850: {[conv2_act.W Value[32 x 800]] }
00000047AD12D4B0: {[conv3_act.p Gradient[7 x 7 x 64 x *]] [pool2 Gradient[7 x 7 x 32 x *]] [pool3 Value[3 x 3 x 64 x *]] }
00000047AD12D690: {[conv3_act.W Gradient[64 x 800]] [conv3_act.p Value[7 x 7 x 64 x *]] }
00000047AD12D7D0: {[CE Value[1]] }
00000047AD12D870: {[featScaled Value[32 x 32 x 3 x *]] }
00000047AD12D910: {[conv1_act.c Value[32 x 32 x 32 x *]] }
00000047AD12DC30: {[OutputNodes.t Value[10 x 1 x *]] [h1.b Gradient[64 x 1]] [h1.y Gradient[64 x 1 x *]] }
00000047AD12DEB0: {[conv2_act.p Gradient[15 x 15 x 32 x *]] [pool1 Gradient[15 x 15 x 32 x *]] [pool2 Value[7 x 7 x 32 x *]] }
00000047AD12DF50: {[conv3_act.c Gradient[7 x 7 x 64 x *]] [conv3_act.y Value[7 x 7 x 64 x *]] }
00000047AD12E090: {[OutputNodes.t Gradient[10 x 1 x *]] }
00000047AD12E130: {[conv1_act.p Gradient[32 x 32 x 32 x *]] [pool1 Value[15 x 15 x 32 x *]] }
00000047AD12E270: {[conv2_act.c Gradient[15 x 15 x 32 x *]] [conv2_act.y Value[15 x 15 x 32 x *]] }
00000047AD12E3B0: {[Err Value[1]] }
00000047AD12E450: {[conv2_act.c Value[15 x 15 x 32 x *]] }
00000047AD12E590: {[h1_d Gradient[64 x 1 x *]] }
00000047AD12E6D0: {[conv1_act.W Gradient[32 x 75]] [conv1_act.p Value[32 x 32 x 32 x *]] }
00000047AD12E810: {[h1.t Gradient[64 x *]] [h1.y Value[64 x 1 x *]] }
00000047AD12E8B0: {[h1_d Value[64 x 1 x *]] }
00000047AD12E950: {[conv3_act.c Value[7 x 7 x 64 x *]] }
00000047AD12EA90: {[OutputNodes.W Gradient[10 x 64]] [OutputNodes.z Gradient[10 x 1 x *]] }
00000047AD12EBD0: {[conv2_act.W Gradient[32 x 800]] [conv2_act.p Value[15 x 15 x 32 x *]] }
00000047AD12EC70: {[h1.z Gradient[64 x 1 x *]] [pool3 Gradient[3 x 3 x 64 x *]] }
00000047AD12ED10: {[OutputNodes.b Gradient[10]] }
00000047AD12F2B0: {[CE Gradient[1]] }
00000047AD12F490: {[conv2_act.b Gradient[1 x 1 x 32]] [conv2_act.y Gradient[15 x 15 x 32 x *]] }
00000047AD12F530: {[conv3_act.b Gradient[1 x 1 x 64]] [conv3_act.y Gradient[7 x 7 x 64 x *]] [h1.t Value[64 x *]] }
00000047AD12F5D0: {[h1.W Gradient[64 x 3 x 3 x 64]] [h1.z Value[64 x 1 x *]] }
00000047AD12F670: {[conv1_act.b Gradient[1 x 1 x 32]] [conv1_act.y Gradient[32 x 32 x 32 x *]] }
00000047AD12F710: {[OutputNodes.z Value[10 x 1 x *]] }
00000047AD12F7B0: {[conv1_act.c Gradient[32 x 32 x 32 x *]] [conv1_act.y Value[32 x 32 x 32 x *]] }
05/13/2016 08:17:53: No PreCompute nodes found, skipping PreCompute step.
05/13/2016 08:17:53: Starting Epoch 1: learning rate per sample = 0.000156 effective momentum = 0.900000 momentum as time constant = 607.4 samples
05/13/2016 08:17:53: Starting minibatch loop.
05/13/2016 08:18:02: Finished Epoch[ 1 of 10]: [Training] CE = 2.30266907 * 100; Err = 0.91000000 * 100; totalSamplesSeen = 100; learningRatePerSample = 0.00015625; epochTime=9.23399s
05/13/2016 08:18:02: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution.1'
05/13/2016 08:18:02: Starting Epoch 2: learning rate per sample = 0.000156 effective momentum = 0.900000 momentum as time constant = 607.4 samples
05/13/2016 08:18:02: Starting minibatch loop.
05/13/2016 08:18:02: Finished Epoch[ 2 of 10]: [Training] CE = 2.30141006 * 100; Err = 0.86000000 * 100; totalSamplesSeen = 200; learningRatePerSample = 0.00015625; epochTime=0.0203s
05/13/2016 08:18:02: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution.2'
05/13/2016 08:18:02: Starting Epoch 3: learning rate per sample = 0.000156 effective momentum = 0.900000 momentum as time constant = 607.4 samples
05/13/2016 08:18:02: Starting minibatch loop.
05/13/2016 08:18:02: Finished Epoch[ 3 of 10]: [Training] CE = 2.30164764 * 100; Err = 0.92000000 * 100; totalSamplesSeen = 300; learningRatePerSample = 0.00015625; epochTime=0.020255s
05/13/2016 08:18:02: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution.3'
05/13/2016 08:18:02: Starting Epoch 4: learning rate per sample = 0.000156 effective momentum = 0.900000 momentum as time constant = 607.4 samples
05/13/2016 08:18:02: Starting minibatch loop.
05/13/2016 08:18:02: Finished Epoch[ 4 of 10]: [Training] CE = 2.29509628 * 100; Err = 0.87000000 * 100; totalSamplesSeen = 400; learningRatePerSample = 0.00015625; epochTime=0.020212s
05/13/2016 08:18:02: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution.4'
05/13/2016 08:18:03: Starting Epoch 5: learning rate per sample = 0.000156 effective momentum = 0.900000 momentum as time constant = 607.4 samples
05/13/2016 08:18:03: Starting minibatch loop.
05/13/2016 08:18:03: Finished Epoch[ 5 of 10]: [Training] CE = 2.29264740 * 100; Err = 0.87000000 * 100; totalSamplesSeen = 500; learningRatePerSample = 0.00015625; epochTime=0.02033s
05/13/2016 08:18:03: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution.5'
Setting dropout rate to 0.5.
05/13/2016 08:18:03: Starting Epoch 6: learning rate per sample = 0.000156 effective momentum = 0.900000 momentum as time constant = 607.4 samples
05/13/2016 08:18:03: Starting minibatch loop.
(GPU): creating curand object with seed 5
05/13/2016 08:18:03: Finished Epoch[ 6 of 10]: [Training] CE = 2.30378311 * 100; Err = 0.92000000 * 100; totalSamplesSeen = 600; learningRatePerSample = 0.00015625; epochTime=0.026637s
05/13/2016 08:18:03: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution.6'
05/13/2016 08:18:03: Starting Epoch 7: learning rate per sample = 0.000156 effective momentum = 0.900000 momentum as time constant = 607.4 samples
05/13/2016 08:18:03: Starting minibatch loop.
(GPU): creating curand object with seed 6
05/13/2016 08:18:03: Finished Epoch[ 7 of 10]: [Training] CE = 2.28946518 * 100; Err = 0.90000000 * 100; totalSamplesSeen = 700; learningRatePerSample = 0.00015625; epochTime=0.02618s
05/13/2016 08:18:03: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution.7'
05/13/2016 08:18:03: Starting Epoch 8: learning rate per sample = 0.000156 effective momentum = 0.900000 momentum as time constant = 607.4 samples
05/13/2016 08:18:03: Starting minibatch loop.
(GPU): creating curand object with seed 7
05/13/2016 08:18:03: Finished Epoch[ 8 of 10]: [Training] CE = 2.29619675 * 100; Err = 0.93000000 * 100; totalSamplesSeen = 800; learningRatePerSample = 0.00015625; epochTime=0.026196s
05/13/2016 08:18:03: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution.8'
05/13/2016 08:18:03: Starting Epoch 9: learning rate per sample = 0.000156 effective momentum = 0.900000 momentum as time constant = 607.4 samples
05/13/2016 08:18:03: Starting minibatch loop.
(GPU): creating curand object with seed 8
05/13/2016 08:18:03: Finished Epoch[ 9 of 10]: [Training] CE = 2.27065186 * 100; Err = 0.83000000 * 100; totalSamplesSeen = 900; learningRatePerSample = 0.00015625; epochTime=0.026126s
05/13/2016 08:18:03: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution.9'
05/13/2016 08:18:03: Starting Epoch 10: learning rate per sample = 0.000156 effective momentum = 0.900000 momentum as time constant = 607.4 samples
05/13/2016 08:18:03: Starting minibatch loop.
(GPU): creating curand object with seed 9
05/13/2016 08:18:03: Finished Epoch[10 of 10]: [Training] CE = 2.31216217 * 100; Err = 0.92000000 * 100; totalSamplesSeen = 1000; learningRatePerSample = 0.00015625; epochTime=0.026148s
05/13/2016 08:18:03: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_01_Convolution@release_gpu/Models/01_Convolution'
05/13/2016 08:18:03: CNTKCommandTrainEnd: Train
05/13/2016 08:18:03: Action "train" complete.
05/13/2016 08:18:03: ##############################################################################
05/13/2016 08:18:03: # #
05/13/2016 08:18:03: # Action "test" #
05/13/2016 08:18:03: # #
05/13/2016 08:18:03: ##############################################################################
Post-processing network...
3 roots:
CE = CrossEntropyWithSoftmax()
Err = ErrorPrediction()
OutputNodes.z = Plus()
Validating network. 34 nodes to process in pass 1.
Validating --> labels = InputValue() : -> [10 x *1]
Validating --> OutputNodes.W = LearnableParameter() : -> [10 x 64]
Validating --> h1.W = LearnableParameter() : -> [64 x 3 x 3 x 64]
Validating --> conv3_act.W = LearnableParameter() : -> [64 x 800]
Validating --> conv2_act.W = LearnableParameter() : -> [32 x 800]
Validating --> conv1_act.W = LearnableParameter() : -> [32 x 75]
Validating --> features = InputValue() : -> [32 x 32 x 3 x *1]
Validating --> featOffs = LearnableParameter() : -> [1 x 1]
Validating --> featScaled = Minus (features, featOffs) : [32 x 32 x 3 x *1], [1 x 1] -> [32 x 32 x 3 x *1]
Validating --> conv1_act.c = Convolution (conv1_act.W, featScaled) : [32 x 75], [32 x 32 x 3 x *1] -> [32 x 32 x 32 x *1]
Validating --> conv1_act.b = LearnableParameter() : -> [1 x 1 x 32]
Validating --> conv1_act.p = Plus (conv1_act.c, conv1_act.b) : [32 x 32 x 32 x *1], [1 x 1 x 32] -> [32 x 32 x 32 x *1]
Validating --> conv1_act.y = RectifiedLinear (conv1_act.p) : [32 x 32 x 32 x *1] -> [32 x 32 x 32 x *1]
Validating --> pool1 = MaxPooling (conv1_act.y) : [32 x 32 x 32 x *1] -> [15 x 15 x 32 x *1]
Validating --> conv2_act.c = Convolution (conv2_act.W, pool1) : [32 x 800], [15 x 15 x 32 x *1] -> [15 x 15 x 32 x *1]
Validating --> conv2_act.b = LearnableParameter() : -> [1 x 1 x 32]
Validating --> conv2_act.p = Plus (conv2_act.c, conv2_act.b) : [15 x 15 x 32 x *1], [1 x 1 x 32] -> [15 x 15 x 32 x *1]
Validating --> conv2_act.y = RectifiedLinear (conv2_act.p) : [15 x 15 x 32 x *1] -> [15 x 15 x 32 x *1]
Validating --> pool2 = MaxPooling (conv2_act.y) : [15 x 15 x 32 x *1] -> [7 x 7 x 32 x *1]
Validating --> conv3_act.c = Convolution (conv3_act.W, pool2) : [64 x 800], [7 x 7 x 32 x *1] -> [7 x 7 x 64 x *1]
Validating --> conv3_act.b = LearnableParameter() : -> [1 x 1 x 64]
Validating --> conv3_act.p = Plus (conv3_act.c, conv3_act.b) : [7 x 7 x 64 x *1], [1 x 1 x 64] -> [7 x 7 x 64 x *1]
Validating --> conv3_act.y = RectifiedLinear (conv3_act.p) : [7 x 7 x 64 x *1] -> [7 x 7 x 64 x *1]
Validating --> pool3 = MaxPooling (conv3_act.y) : [7 x 7 x 64 x *1] -> [3 x 3 x 64 x *1]
Validating --> h1.t = Times (h1.W, pool3) : [64 x 3 x 3 x 64], [3 x 3 x 64 x *1] -> [64 x *1]
Validating --> h1.b = LearnableParameter() : -> [64 x 1]
Validating --> h1.z = Plus (h1.t, h1.b) : [64 x *1], [64 x 1] -> [64 x 1 x *1]
Validating --> h1.y = RectifiedLinear (h1.z) : [64 x 1 x *1] -> [64 x 1 x *1]
Validating --> h1_d = Dropout (h1.y) : [64 x 1 x *1] -> [64 x 1 x *1]
Validating --> OutputNodes.t = Times (OutputNodes.W, h1_d) : [10 x 64], [64 x 1 x *1] -> [10 x 1 x *1]
Validating --> OutputNodes.b = LearnableParameter() : -> [10]
Validating --> OutputNodes.z = Plus (OutputNodes.t, OutputNodes.b) : [10 x 1 x *1], [10] -> [10 x 1 x *1]
Validating --> CE = CrossEntropyWithSoftmax (labels, OutputNodes.z) : [10 x *1], [10 x 1 x *1] -> [1]
Validating --> Err = ErrorPrediction (labels, OutputNodes.z) : [10 x *1], [10 x 1 x *1] -> [1]
Validating network. 21 nodes to process in pass 2.
Validating network, final pass.
Using cuDNN convolution engine for geometry: Input: 32 x 32 x 3, Output: 32 x 32 x 32, Kernel: 5 x 5 x 3, Map: 1 x 1 x 32, Stride: 1 x 1 x 3, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 32 x 32 x 32, Output: 15 x 15 x 32, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 15 x 15 x 32, Output: 15 x 15 x 32, Kernel: 5 x 5 x 32, Map: 1 x 1 x 32, Stride: 1 x 1 x 32, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 15 x 15 x 32, Output: 7 x 7 x 32, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 7 x 7 x 32, Output: 7 x 7 x 64, Kernel: 5 x 5 x 32, Map: 1 x 1 x 64, Stride: 1 x 1 x 32, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 7 x 7 x 64, Output: 3 x 3 x 64, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
13 out of 34 nodes do not share the minibatch layout with the input data.
Post-processing network complete.
evalNodeNames are not specified, using all the default evalnodes and training criterion nodes.
Allocating matrices for forward and/or backward propagation.
Memory Sharing Structure:
0000000000000000: {[CE Gradient[1]] [Err Gradient[1]] [OutputNodes.W Gradient[10 x 64]] [OutputNodes.b Gradient[10]] [OutputNodes.t Gradient[10 x 1 x *1]] [OutputNodes.z Gradient[10 x 1 x *1]] [conv1_act.W Gradient[32 x 75]] [conv1_act.b Gradient[1 x 1 x 32]] [conv1_act.c Gradient[32 x 32 x 32 x *1]] [conv1_act.p Gradient[32 x 32 x 32 x *1]] [conv1_act.y Gradient[32 x 32 x 32 x *1]] [conv2_act.W Gradient[32 x 800]] [conv2_act.b Gradient[1 x 1 x 32]] [conv2_act.c Gradient[15 x 15 x 32 x *1]] [conv2_act.p Gradient[15 x 15 x 32 x *1]] [conv2_act.y Gradient[15 x 15 x 32 x *1]] [conv3_act.W Gradient[64 x 800]] [conv3_act.b Gradient[1 x 1 x 64]] [conv3_act.c Gradient[7 x 7 x 64 x *1]] [conv3_act.p Gradient[7 x 7 x 64 x *1]] [conv3_act.y Gradient[7 x 7 x 64 x *1]] [featOffs Gradient[1 x 1]] [featScaled Gradient[32 x 32 x 3 x *1]] [features Gradient[32 x 32 x 3 x *1]] [h1.W Gradient[64 x 3 x 3 x 64]] [h1.b Gradient[64 x 1]] [h1.t Gradient[64 x *1]] [h1.y Gradient[64 x 1 x *1]] [h1.z Gradient[64 x 1 x *1]] [h1_d Gradient[64 x 1 x *1]] [labels Gradient[10 x *1]] [pool1 Gradient[15 x 15 x 32 x *1]] [pool2 Gradient[7 x 7 x 32 x *1]] [pool3 Gradient[3 x 3 x 64 x *1]] }
00000047A69D0FF0: {[h1.t Value[64 x *1]] }
00000047A69D16D0: {[h1_d Value[64 x 1 x *1]] }
00000047A69D1D10: {[OutputNodes.z Value[10 x 1 x *1]] }
00000047A69D1DB0: {[OutputNodes.t Value[10 x 1 x *1]] }
00000047A69D1E50: {[h1.y Value[64 x 1 x *1]] }
00000047A69D2530: {[h1.z Value[64 x 1 x *1]] }
00000047AD12D7D0: {[featOffs Value[1 x 1]] }
00000047AD12D910: {[OutputNodes.W Value[10 x 64]] }
00000047AD12DC30: {[conv1_act.b Value[1 x 1 x 32]] }
00000047AD12DF50: {[conv2_act.W Value[32 x 800]] }
00000047AD12E090: {[conv3_act.W Value[64 x 800]] }
00000047AD12E3B0: {[features Value[32 x 32 x 3 x *1]] }
00000047AD12E9F0: {[h1.W Value[64 x 3 x 3 x 64]] }
00000047AD12EA90: {[conv1_act.W Value[32 x 75]] }
00000047AD12EBD0: {[conv2_act.b Value[1 x 1 x 32]] }
00000047AD12ED10: {[labels Value[10 x *1]] }
00000047AD12F210: {[h1.b Value[64 x 1]] }
00000047AD12F670: {[conv3_act.b Value[1 x 1 x 64]] }
00000047AD12FB70: {[OutputNodes.b Value[10]] }
00000047AD12FCB0: {[conv2_act.y Value[15 x 15 x 32 x *1]] }
00000047AD12FD50: {[pool2 Value[7 x 7 x 32 x *1]] }
00000047AD12FFD0: {[conv2_act.p Value[15 x 15 x 32 x *1]] }
00000047AD130110: {[featScaled Value[32 x 32 x 3 x *1]] }
00000047AD130250: {[conv3_act.c Value[7 x 7 x 64 x *1]] }
00000047AD130390: {[conv3_act.p Value[7 x 7 x 64 x *1]] }
00000047AD130430: {[conv3_act.y Value[7 x 7 x 64 x *1]] }
00000047AD1304D0: {[conv1_act.p Value[32 x 32 x 32 x *1]] }
00000047AD130570: {[pool3 Value[3 x 3 x 64 x *1]] }
00000047AD1307F0: {[conv1_act.y Value[32 x 32 x 32 x *1]] }
00000047AD130BB0: {[conv1_act.c Value[32 x 32 x 32 x *1]] }
00000047AD1310B0: {[Err Value[1]] }
00000047AD131150: {[pool1 Value[15 x 15 x 32 x *1]] }
00000047AD1311F0: {[CE Value[1]] }
00000047AD131330: {[conv2_act.c Value[15 x 15 x 32 x *1]] }
05/13/2016 08:18:17: Final Results: Minibatch[1-625]: Err = 0.84020000 * 10000; CE = 2.27465317 * 10000; perplexity = 9.72454569
05/13/2016 08:18:17: Action "test" complete.
05/13/2016 08:18:17: __COMPLETED__

12
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution/run-test Executable file
Просмотреть файл

@ -0,0 +1,12 @@
#!/bin/bash
. $TEST_DIR/../run-test-common
# cntkrun <CNTK config file name> <additional CNTK args>
cntkrun $CNTKTextConfigDir/01_Conv.cntk "Train=[SGD=[maxEpochs=10]] Train=[SGD=[epochSize=100]] stderr=-"
ExitCode=$?
# Delete the test data if copied
[[ "$Copied" -eq "1" ]] && rm -rf "$DataDir"
exit $ExitCode

34
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/01_Convolution/testcases.yml Normal file
Просмотреть файл

@ -0,0 +1,34 @@
dataDir: ../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10
tags:
# Debug version asserts on Windows.
# In BVT, run Release GPU
- bvt-e (build_sku=='gpu') & (device=='gpu') & (flavor=='release')
# In Nightly on Linux, additionally run Debug GPU and Release CPU
- nightly-e (build_sku=='gpu') & (((device=='gpu') & (flavor=='release')) | ((os=='linux') & ((flavor=='debug') ^ (device=='cpu'))))
testCases:
CNTK Run must be completed:
patterns:
- __COMPLETED__
Must train epochs in exactly same order and parameters:
patterns:
- Starting Epoch {{integer}}
- learning rate per sample = {{float}}
- momentum = {{float}}
Epochs must be finished with expected results:
patterns:
- Finished Epoch
- CE = {{float,tolerance=2.0%}} * {{integer}}
# this part is flaky, needs to be investigated
#- Err = {{float,tolerance=2.0%}} * {{integer}}
- totalSamplesSeen = {{integer}}
- learningRatePerSample = {{float,tolerance=0.001%}}
Final test results must match:
patterns:
- "Final Results: Minibatch[{{integer}}-{{integer}}]"
- Err = {{float,tolerance=2.0%}} * {{integer}}
- CE = {{float,tolerance=2.0%}} * {{integer}}

9
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv/README.txt Normal file
Просмотреть файл

@ -0,0 +1,9 @@
Test runtimes
Windows:
Running test CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv (debug gpu) - [OK] 150.11 sec
Running test CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv (release gpu) - [OK] 25.94 sec
Linux:
Running test CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv (debug gpu) - [OK] 30.01 sec
Running test CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv (release gpu) - [OK] 10.32 sec

634
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv/baseline.linux.txt Normal file
Просмотреть файл

@ -0,0 +1,634 @@
=== Running /home/philly/jenkins/workspace/CNTK-Test-Linux-W1/build/gpu/release/bin/cntk configFile=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10/../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/Config/02_BatchNormConv.cntk currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/TestData RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/TestData ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10 OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu DeviceId=0 timestamping=true Train=[SGD=[maxEpochs=5]] Train=[SGD=[epochSize=100]] stderr=-
-------------------------------------------------------------------
Build info:
Built time: May 13 2016 14:50:25
Last modified date: Thu May 12 14:00:37 2016
Build type: release
Build target: GPU
With 1bit-SGD: no
Math lib: acml
CUDA_PATH: /usr/local/cuda-7.5
CUB_PATH: /usr/local/cub-1.4.1
CUDNN_PATH: /usr/local/cudnn-4.0
Build Branch: HEAD
Build SHA1: 35fadc316f045d843bbd9b85061250a959268787
Built by philly on d8dc82703b0f
Build Path: /home/philly/jenkins/workspace/CNTK-Build-Linux
-------------------------------------------------------------------
Changed current directory to /tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/TestData
05/13/2016 15:10:59: Redirecting stderr to file -_Train_Test.log
05/13/2016 15:10:59: -------------------------------------------------------------------
05/13/2016 15:10:59: Build info:
05/13/2016 15:10:59: Built time: May 13 2016 14:50:25
05/13/2016 15:10:59: Last modified date: Thu May 12 14:00:37 2016
05/13/2016 15:10:59: Build type: release
05/13/2016 15:10:59: Build target: GPU
05/13/2016 15:10:59: With 1bit-SGD: no
05/13/2016 15:10:59: Math lib: acml
05/13/2016 15:10:59: CUDA_PATH: /usr/local/cuda-7.5
05/13/2016 15:10:59: CUB_PATH: /usr/local/cub-1.4.1
05/13/2016 15:10:59: CUDNN_PATH: /usr/local/cudnn-4.0
05/13/2016 15:10:59: Build Branch: HEAD
05/13/2016 15:10:59: Build SHA1: 35fadc316f045d843bbd9b85061250a959268787
05/13/2016 15:10:59: Built by philly on d8dc82703b0f
05/13/2016 15:10:59: Build Path: /home/philly/jenkins/workspace/CNTK-Build-Linux
05/13/2016 15:10:59: -------------------------------------------------------------------
05/13/2016 15:10:59: Running on localhost at 2016/05/13 15:10:59
05/13/2016 15:10:59: Command line:
/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/build/gpu/release/bin/cntk configFile=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10/../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/Config/02_BatchNormConv.cntk currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/TestData RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/TestData ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10 OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu DeviceId=0 timestamping=true Train=[SGD=[maxEpochs=5]] Train=[SGD=[epochSize=100]] stderr=-
05/13/2016 15:10:59: >>>>>>>>>>>>>>>>>>>> RAW CONFIG (VARIABLES NOT RESOLVED) >>>>>>>>>>>>>>>>>>>>
05/13/2016 15:10:59: RootDir = "."
ConfigDir = "$RootDir$"
DataDir = "$RootDir$"
OutputDir = "$RootDir$/Output"
ModelDir = "$OutputDir$/Models"
ndlMacros = "$ConfigDir$/Macros.ndl"
precision = "float"
deviceId = 0
imageLayout = "cudnn"
initOnCPUOnly=true
prefetch = "true"
command = Train:Test
stderr = "$OutputDir$/02_BatchNormConv"
traceLevel = 1
numMBsToShowResult = 500
Train = [
action = "train"
modelPath = "$ModelDir$/02_BatchNormConv"
NDLNetworkBuilder = [
networkDescription = "$ConfigDir$/02_BatchNormConv.ndl"
]
SGD = [
epochSize = 49984
minibatchSize = 64
learningRatesPerMB = 0.03*7:0.01
momentumPerMB = 0
maxEpochs = 10
L2RegWeight = 0
dropoutRate = 0
]
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Train_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
Test = [
action = "test"
modelPath = "$ModelDir$/02_BatchNormConv"
minibatchSize = 16
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Test_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/TestData
RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu
DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/TestData
ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10
OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu
DeviceId=0
timestamping=true
Train=[SGD=[maxEpochs=5]]
Train=[SGD=[epochSize=100]]
stderr=-
05/13/2016 15:10:59: <<<<<<<<<<<<<<<<<<<< RAW CONFIG (VARIABLES NOT RESOLVED) <<<<<<<<<<<<<<<<<<<<
05/13/2016 15:10:59: >>>>>>>>>>>>>>>>>>>> RAW CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>
05/13/2016 15:10:59: RootDir = "."
ConfigDir = "."
DataDir = "."
OutputDir = "./Output"
ModelDir = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/Models"
ndlMacros = "/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10/Macros.ndl"
precision = "float"
deviceId = 0
imageLayout = "cudnn"
initOnCPUOnly=true
prefetch = "true"
command = Train:Test
stderr = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/02_BatchNormConv"
traceLevel = 1
numMBsToShowResult = 500
Train = [
action = "train"
modelPath = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/Models/02_BatchNormConv"
NDLNetworkBuilder = [
networkDescription = "/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv.ndl"
]
SGD = [
epochSize = 49984
minibatchSize = 64
learningRatesPerMB = 0.03*7:0.01
momentumPerMB = 0
maxEpochs = 10
L2RegWeight = 0
dropoutRate = 0
]
reader = [
readerType = "CNTKTextFormatReader"
file = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/TestData/Train_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
Test = [
action = "test"
modelPath = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/Models/02_BatchNormConv"
minibatchSize = 16
reader = [
readerType = "CNTKTextFormatReader"
file = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/TestData/Test_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/TestData
RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu
DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/TestData
ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10
OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu
DeviceId=0
timestamping=true
Train=[SGD=[maxEpochs=5]]
Train=[SGD=[epochSize=100]]
stderr=-
05/13/2016 15:10:59: <<<<<<<<<<<<<<<<<<<< RAW CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<
05/13/2016 15:10:59: >>>>>>>>>>>>>>>>>>>> PROCESSED CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>
configparameters: 02_BatchNormConv.cntk:command=Train:Test
configparameters: 02_BatchNormConv.cntk:ConfigDir=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10
configparameters: 02_BatchNormConv.cntk:currentDirectory=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/TestData
configparameters: 02_BatchNormConv.cntk:DataDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/TestData
configparameters: 02_BatchNormConv.cntk:deviceId=0
configparameters: 02_BatchNormConv.cntk:imageLayout=cudnn
configparameters: 02_BatchNormConv.cntk:initOnCPUOnly=true
configparameters: 02_BatchNormConv.cntk:ModelDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/Models
configparameters: 02_BatchNormConv.cntk:ndlMacros=/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10/Macros.ndl
configparameters: 02_BatchNormConv.cntk:numMBsToShowResult=500
configparameters: 02_BatchNormConv.cntk:OutputDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu
configparameters: 02_BatchNormConv.cntk:precision=float
configparameters: 02_BatchNormConv.cntk:prefetch=true
configparameters: 02_BatchNormConv.cntk:RootDir=.
configparameters: 02_BatchNormConv.cntk:RunDir=/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu
configparameters: 02_BatchNormConv.cntk:stderr=-
configparameters: 02_BatchNormConv.cntk:Test=[
action = "test"
modelPath = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/Models/02_BatchNormConv"
minibatchSize = 16
reader = [
readerType = "CNTKTextFormatReader"
file = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/TestData/Test_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
configparameters: 02_BatchNormConv.cntk:timestamping=true
configparameters: 02_BatchNormConv.cntk:traceLevel=1
configparameters: 02_BatchNormConv.cntk:Train=[
action = "train"
modelPath = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/Models/02_BatchNormConv"
NDLNetworkBuilder = [
networkDescription = "/home/philly/jenkins/workspace/CNTK-Test-Linux-W1/Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv/../../../../../../../../Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv.ndl"
]
SGD = [
epochSize = 49984
minibatchSize = 64
learningRatesPerMB = 0.03*7:0.01
momentumPerMB = 0
maxEpochs = 10
L2RegWeight = 0
dropoutRate = 0
]
reader = [
readerType = "CNTKTextFormatReader"
file = "/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/TestData/Train_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
] [SGD=[maxEpochs=5]] [SGD=[epochSize=100]]
05/13/2016 15:10:59: <<<<<<<<<<<<<<<<<<<< PROCESSED CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<
05/13/2016 15:10:59: Commands: Train Test
05/13/2016 15:10:59: Precision = "float"
05/13/2016 15:10:59: CNTKModelPath: /tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/Models/02_BatchNormConv
05/13/2016 15:10:59: CNTKCommandTrainInfo: Train : 5
05/13/2016 15:10:59: CNTKCommandTrainInfo: CNTKNoMoreCommands_Total : 5
05/13/2016 15:10:59: ##############################################################################
05/13/2016 15:10:59: # #
05/13/2016 15:10:59: # Action "train" #
05/13/2016 15:10:59: # #
05/13/2016 15:10:59: ##############################################################################
05/13/2016 15:10:59: CNTKCommandTrainBegin: Train
NDLBuilder Using GPU 0
05/13/2016 15:10:59: Creating virgin network.
SetGaussianRandomValue (GPU): creating curand object with seed 1, sizeof(ElemType)==4
Post-processing network...
3 roots:
CE = CrossEntropyWithSoftmax()
Err = ErrorPrediction()
OutputNodes.z = Plus()
Validating network. 45 nodes to process in pass 1.
Validating --> labels = InputValue() : -> [10 x *]
Validating --> OutputNodes.W = LearnableParameter() : -> [10 x 64]
Validating --> h1.W = LearnableParameter() : -> [64 x 3 x 3 x 64]
Validating --> conv3.c.W = LearnableParameter() : -> [64 x 800]
Validating --> conv2.c.W = LearnableParameter() : -> [32 x 800]
Validating --> conv1.c.W = LearnableParameter() : -> [32 x 75]
Validating --> features = InputValue() : -> [32 x 32 x 3 x *]
Validating --> featOffs = LearnableParameter() : -> [1 x 1]
Validating --> featScaled = Minus (features, featOffs) : [32 x 32 x 3 x *], [1 x 1] -> [32 x 32 x 3 x *]
Validating --> conv1.c.c.c = Convolution (conv1.c.W, featScaled) : [32 x 75], [32 x 32 x 3 x *] -> [32 x 32 x 32 x *]
Validating --> conv1.c.c.sc = LearnableParameter() : -> [32 x 1]
Validating --> conv1.c.c.b = LearnableParameter() : -> [32 x 1]
Validating --> conv1.c.c.m = LearnableParameter() : -> [32 x 1]
Validating --> conv1.c.c.isd = LearnableParameter() : -> [32 x 1]
Validating --> conv1.c.c.y = BatchNormalization (conv1.c.c.c, conv1.c.c.sc, conv1.c.c.b, conv1.c.c.m, conv1.c.c.isd) : [32 x 32 x 32 x *], [32 x 1], [32 x 1], [32 x 1], [32 x 1] -> [32 x 32 x 32 x *]
Validating --> conv1.y = RectifiedLinear (conv1.c.c.y) : [32 x 32 x 32 x *] -> [32 x 32 x 32 x *]
Validating --> pool1 = MaxPooling (conv1.y) : [32 x 32 x 32 x *] -> [15 x 15 x 32 x *]
Validating --> conv2.c.c.c = Convolution (conv2.c.W, pool1) : [32 x 800], [15 x 15 x 32 x *] -> [15 x 15 x 32 x *]
Validating --> conv2.c.c.sc = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.b = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.m = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.isd = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.y = BatchNormalization (conv2.c.c.c, conv2.c.c.sc, conv2.c.c.b, conv2.c.c.m, conv2.c.c.isd) : [15 x 15 x 32 x *], [32 x 1], [32 x 1], [32 x 1], [32 x 1] -> [15 x 15 x 32 x *]
Validating --> conv2.y = RectifiedLinear (conv2.c.c.y) : [15 x 15 x 32 x *] -> [15 x 15 x 32 x *]
Validating --> pool2 = MaxPooling (conv2.y) : [15 x 15 x 32 x *] -> [7 x 7 x 32 x *]
Validating --> conv3.c.c.c = Convolution (conv3.c.W, pool2) : [64 x 800], [7 x 7 x 32 x *] -> [7 x 7 x 64 x *]
Validating --> conv3.c.c.sc = LearnableParameter() : -> [64 x 1]
Validating --> conv3.c.c.b = LearnableParameter() : -> [64 x 1]
Validating --> conv3.c.c.m = LearnableParameter() : -> [64 x 1]
Validating --> conv3.c.c.isd = LearnableParameter() : -> [64 x 1]
Validating --> conv3.c.c.y = BatchNormalization (conv3.c.c.c, conv3.c.c.sc, conv3.c.c.b, conv3.c.c.m, conv3.c.c.isd) : [7 x 7 x 64 x *], [64 x 1], [64 x 1], [64 x 1], [64 x 1] -> [7 x 7 x 64 x *]
Validating --> conv3.y = RectifiedLinear (conv3.c.c.y) : [7 x 7 x 64 x *] -> [7 x 7 x 64 x *]
Validating --> pool3 = MaxPooling (conv3.y) : [7 x 7 x 64 x *] -> [3 x 3 x 64 x *]
Validating --> h1.t = Times (h1.W, pool3) : [64 x 3 x 3 x 64], [3 x 3 x 64 x *] -> [64 x *]
Validating --> h1.sc = LearnableParameter() : -> [64 x 1]
Validating --> h1.b = LearnableParameter() : -> [64 x 1]
Validating --> h1.m = LearnableParameter() : -> [64 x 1]
Validating --> h1.isd = LearnableParameter() : -> [64 x 1]
Validating --> h1.bn = BatchNormalization (h1.t, h1.sc, h1.b, h1.m, h1.isd) : [64 x *], [64 x 1], [64 x 1], [64 x 1], [64 x 1] -> [64 x *]
Validating --> h1.y = RectifiedLinear (h1.bn) : [64 x *] -> [64 x *]
Validating --> OutputNodes.t = Times (OutputNodes.W, h1.y) : [10 x 64], [64 x *] -> [10 x *]
Validating --> OutputNodes.b = LearnableParameter() : -> [10]
Validating --> OutputNodes.z = Plus (OutputNodes.t, OutputNodes.b) : [10 x *], [10] -> [10 x *]
Validating --> CE = CrossEntropyWithSoftmax (labels, OutputNodes.z) : [10 x *], [10 x *] -> [1]
Validating --> Err = ErrorPrediction (labels, OutputNodes.z) : [10 x *], [10 x *] -> [1]
Validating network. 20 nodes to process in pass 2.
Validating network, final pass.
Using cuDNN convolution engine for geometry: Input: 32 x 32 x 3, Output: 32 x 32 x 32, Kernel: 5 x 5 x 3, Map: 1 x 1 x 32, Stride: 1 x 1 x 3, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
Using cuDNN convolution engine for geometry: Input: 32 x 32 x 32, Output: 15 x 15 x 32, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 15 x 15 x 32, Output: 15 x 15 x 32, Kernel: 5 x 5 x 32, Map: 1 x 1 x 32, Stride: 1 x 1 x 32, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
Using cuDNN convolution engine for geometry: Input: 15 x 15 x 32, Output: 7 x 7 x 32, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 7 x 7 x 32, Output: 7 x 7 x 64, Kernel: 5 x 5 x 32, Map: 1 x 1 x 64, Stride: 1 x 1 x 32, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
Using cuDNN convolution engine for geometry: Input: 7 x 7 x 64, Output: 3 x 3 x 64, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
25 out of 45 nodes do not share the minibatch layout with the input data.
Post-processing network complete.
05/13/2016 15:10:59: Created model with 45 nodes on GPU 0.
05/13/2016 15:10:59: Training criterion node(s):
05/13/2016 15:10:59: CE = CrossEntropyWithSoftmax
05/13/2016 15:10:59: Evaluation criterion node(s):
05/13/2016 15:10:59: Err = ErrorPrediction
Allocating matrices for forward and/or backward propagation.
Memory Sharing Structure:
(nil): {[Err Gradient[1]] [conv1.c.c.isd Gradient[32 x 1]] [conv1.c.c.m Gradient[32 x 1]] [conv2.c.c.isd Gradient[32 x 1]] [conv2.c.c.m Gradient[32 x 1]] [conv3.c.c.isd Gradient[64 x 1]] [conv3.c.c.m Gradient[64 x 1]] [featOffs Gradient[1 x 1]] [featScaled Gradient[32 x 32 x 3 x *]] [features Gradient[32 x 32 x 3 x *]] [h1.isd Gradient[64 x 1]] [h1.m Gradient[64 x 1]] [labels Gradient[10 x *]] }
0x28d4ae8: {[features Value[32 x 32 x 3 x *]] }
0x31e4be8: {[featOffs Value[1 x 1]] }
0x31e6988: {[labels Value[10 x *]] }
0x31e74c8: {[conv1.c.W Value[32 x 75]] }
0x31e7a08: {[conv1.c.c.b Value[32 x 1]] }
0x31e86c8: {[conv1.c.c.sc Value[32 x 1]] }
0x31e9228: {[conv1.c.c.m Value[32 x 1]] }
0x31e9d78: {[conv1.c.c.isd Value[32 x 1]] }
0x31ead98: {[conv2.c.W Value[32 x 800]] }
0x31ec628: {[conv2.c.c.b Value[32 x 1]] }
0x31ed2e8: {[conv2.c.c.sc Value[32 x 1]] }
0x31ed9b8: {[h1.b Value[64 x 1]] }
0x31edfc8: {[conv2.c.c.m Value[32 x 1]] }
0x31eeb38: {[conv2.c.c.isd Value[32 x 1]] }
0x31efbe8: {[conv3.c.W Value[64 x 800]] }
0x31f0ca8: {[conv3.c.c.b Value[64 x 1]] }
0x31f16a8: {[conv3.c.c.sc Value[64 x 1]] }
0x31f2548: {[conv3.c.c.m Value[64 x 1]] }
0x31f30b8: {[conv3.c.c.isd Value[64 x 1]] }
0x31f4a48: {[h1.W Value[64 x 3 x 3 x 64]] }
0x31f6098: {[h1.sc Value[64 x 1]] }
0x31f6b48: {[h1.m Value[64 x 1]] }
0x3629fb8: {[h1.isd Value[64 x 1]] }
0x362b3f8: {[OutputNodes.W Value[10 x 64]] }
0x362c2e8: {[OutputNodes.b Value[10]] }
0x7f51399894d8: {[Err Value[1]] }
0x7f513998b9d8: {[conv1.c.c.y Value[32 x 32 x 32 x *]] }
0x7f51399ab038: {[featScaled Value[32 x 32 x 3 x *]] }
0x7f51399ab378: {[conv1.c.c.c Value[32 x 32 x 32 x *]] }
0x7f51399ac5f8: {[conv1.c.c.c Gradient[32 x 32 x 32 x *]] [conv1.y Value[32 x 32 x 32 x *]] }
0x7f51399ac7b8: {[conv1.c.c.y Gradient[32 x 32 x 32 x *]] [pool1 Value[15 x 15 x 32 x *]] }
0x7f51399ac978: {[conv1.c.W Gradient[32 x 75]] [conv2.c.c.c Value[15 x 15 x 32 x *]] }
0x7f51399acb38: {[conv1.c.c.sc Gradient[32 x 1]] [conv1.y Gradient[32 x 32 x 32 x *]] }
0x7f51399accf8: {[conv2.c.c.y Value[15 x 15 x 32 x *]] }
0x7f51399ad238: {[conv1.c.c.b Gradient[32 x 1]] [conv2.c.c.c Gradient[15 x 15 x 32 x *]] [conv2.y Value[15 x 15 x 32 x *]] }
0x7f51399ad3f8: {[conv2.c.c.y Gradient[15 x 15 x 32 x *]] [pool2 Value[7 x 7 x 32 x *]] }
0x7f51399ad5b8: {[conv2.c.W Gradient[32 x 800]] [conv3.c.c.c Value[7 x 7 x 64 x *]] }
0x7f51399ad778: {[conv2.c.c.sc Gradient[32 x 1]] [conv2.y Gradient[15 x 15 x 32 x *]] }
0x7f51399ad938: {[conv3.c.c.y Value[7 x 7 x 64 x *]] }
0x7f51399ade78: {[conv2.c.c.b Gradient[32 x 1]] [conv3.c.c.c Gradient[7 x 7 x 64 x *]] [conv3.y Value[7 x 7 x 64 x *]] }
0x7f51399ae038: {[conv3.c.c.y Gradient[7 x 7 x 64 x *]] [pool3 Value[3 x 3 x 64 x *]] }
0x7f51399ae1f8: {[conv3.c.c.sc Gradient[64 x 1]] [conv3.y Gradient[7 x 7 x 64 x *]] [h1.t Value[64 x *]] }
0x7f51399ae3b8: {[h1.bn Value[64 x *]] }
0x7f51399ae738: {[conv3.c.c.b Gradient[64 x 1]] }
0x7f51399ae8f8: {[conv3.c.W Gradient[64 x 800]] [h1.t Gradient[64 x *]] [h1.y Value[64 x *]] }
0x7f51399aeab8: {[OutputNodes.t Value[10 x *]] [h1.bn Gradient[64 x *]] }
0x7f51399af598: {[CE Gradient[1]] }
0x7f51399af758: {[OutputNodes.W Gradient[10 x 64]] [OutputNodes.z Gradient[10 x *]] }
0x7f51399af918: {[OutputNodes.t Gradient[10 x *]] [pool1 Gradient[15 x 15 x 32 x *]] [pool2 Gradient[7 x 7 x 32 x *]] [pool3 Gradient[3 x 3 x 64 x *]] }
0x7f51399afad8: {[OutputNodes.b Gradient[10]] }
0x7f51399afc98: {[h1.sc Gradient[64 x 1]] [h1.y Gradient[64 x *]] }
0x7f51399afe88: {[h1.W Gradient[64 x 3 x 3 x 64]] }
0x7f51399b0048: {[h1.b Gradient[64 x 1]] }
0x7f51399b6728: {[OutputNodes.z Value[10 x *]] }
0x7f51399b68e8: {[CE Value[1]] }
05/13/2016 15:10:59: No PreCompute nodes found, skipping PreCompute step.
05/13/2016 15:10:59: Starting Epoch 1: learning rate per sample = 0.000469 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 15:10:59: Starting minibatch loop.
05/13/2016 15:11:03: Finished Epoch[ 1 of 5]: [Training] CE = 2.29343704 * 100; Err = 0.87000000 * 100; totalSamplesSeen = 100; learningRatePerSample = 0.00046874999; epochTime=3.58144s
05/13/2016 15:11:03: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/Models/02_BatchNormConv.1'
05/13/2016 15:11:03: Starting Epoch 2: learning rate per sample = 0.000469 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 15:11:03: Starting minibatch loop.
05/13/2016 15:11:03: Finished Epoch[ 2 of 5]: [Training] CE = 2.22764633 * 100; Err = 0.88000000 * 100; totalSamplesSeen = 200; learningRatePerSample = 0.00046874999; epochTime=0.01264s
05/13/2016 15:11:03: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/Models/02_BatchNormConv.2'
05/13/2016 15:11:03: Starting Epoch 3: learning rate per sample = 0.000469 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 15:11:03: Starting minibatch loop.
05/13/2016 15:11:03: Finished Epoch[ 3 of 5]: [Training] CE = 2.20062683 * 100; Err = 0.77000000 * 100; totalSamplesSeen = 300; learningRatePerSample = 0.00046874999; epochTime=0.01151s
05/13/2016 15:11:03: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/Models/02_BatchNormConv.3'
05/13/2016 15:11:03: Starting Epoch 4: learning rate per sample = 0.000469 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 15:11:03: Starting minibatch loop.
05/13/2016 15:11:03: Finished Epoch[ 4 of 5]: [Training] CE = 2.19534531 * 100; Err = 0.81000000 * 100; totalSamplesSeen = 400; learningRatePerSample = 0.00046874999; epochTime=0.012353s
05/13/2016 15:11:03: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/Models/02_BatchNormConv.4'
05/13/2016 15:11:03: Starting Epoch 5: learning rate per sample = 0.000469 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 15:11:03: Starting minibatch loop.
05/13/2016 15:11:03: Finished Epoch[ 5 of 5]: [Training] CE = 2.16844864 * 100; Err = 0.79000000 * 100; totalSamplesSeen = 500; learningRatePerSample = 0.00046874999; epochTime=0.01142s
05/13/2016 15:11:03: SGD: Saving checkpoint model '/tmp/cntk-test-20160513145544.775982/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10_02_BatchNormConv@release_gpu/Models/02_BatchNormConv'
05/13/2016 15:11:03: CNTKCommandTrainEnd: Train
05/13/2016 15:11:03: Action "train" complete.
05/13/2016 15:11:03: ##############################################################################
05/13/2016 15:11:03: # #
05/13/2016 15:11:03: # Action "test" #
05/13/2016 15:11:03: # #
05/13/2016 15:11:03: ##############################################################################
Post-processing network...
3 roots:
CE = CrossEntropyWithSoftmax()
Err = ErrorPrediction()
OutputNodes.z = Plus()
Validating network. 45 nodes to process in pass 1.
Validating --> labels = InputValue() : -> [10 x *1]
Validating --> OutputNodes.W = LearnableParameter() : -> [10 x 64]
Validating --> h1.W = LearnableParameter() : -> [64 x 3 x 3 x 64]
Validating --> conv3.c.W = LearnableParameter() : -> [64 x 800]
Validating --> conv2.c.W = LearnableParameter() : -> [32 x 800]
Validating --> conv1.c.W = LearnableParameter() : -> [32 x 75]
Validating --> features = InputValue() : -> [32 x 32 x 3 x *1]
Validating --> featOffs = LearnableParameter() : -> [1 x 1]
Validating --> featScaled = Minus (features, featOffs) : [32 x 32 x 3 x *1], [1 x 1] -> [32 x 32 x 3 x *1]
Validating --> conv1.c.c.c = Convolution (conv1.c.W, featScaled) : [32 x 75], [32 x 32 x 3 x *1] -> [32 x 32 x 32 x *1]
Validating --> conv1.c.c.sc = LearnableParameter() : -> [32 x 1]
Validating --> conv1.c.c.b = LearnableParameter() : -> [32 x 1]
Validating --> conv1.c.c.m = LearnableParameter() : -> [32 x 1]
Validating --> conv1.c.c.isd = LearnableParameter() : -> [32 x 1]
Validating --> conv1.c.c.y = BatchNormalization (conv1.c.c.c, conv1.c.c.sc, conv1.c.c.b, conv1.c.c.m, conv1.c.c.isd) : [32 x 32 x 32 x *1], [32 x 1], [32 x 1], [32 x 1], [32 x 1] -> [32 x 32 x 32 x *1]
Validating --> conv1.y = RectifiedLinear (conv1.c.c.y) : [32 x 32 x 32 x *1] -> [32 x 32 x 32 x *1]
Validating --> pool1 = MaxPooling (conv1.y) : [32 x 32 x 32 x *1] -> [15 x 15 x 32 x *1]
Validating --> conv2.c.c.c = Convolution (conv2.c.W, pool1) : [32 x 800], [15 x 15 x 32 x *1] -> [15 x 15 x 32 x *1]
Validating --> conv2.c.c.sc = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.b = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.m = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.isd = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.y = BatchNormalization (conv2.c.c.c, conv2.c.c.sc, conv2.c.c.b, conv2.c.c.m, conv2.c.c.isd) : [15 x 15 x 32 x *1], [32 x 1], [32 x 1], [32 x 1], [32 x 1] -> [15 x 15 x 32 x *1]
Validating --> conv2.y = RectifiedLinear (conv2.c.c.y) : [15 x 15 x 32 x *1] -> [15 x 15 x 32 x *1]
Validating --> pool2 = MaxPooling (conv2.y) : [15 x 15 x 32 x *1] -> [7 x 7 x 32 x *1]
Validating --> conv3.c.c.c = Convolution (conv3.c.W, pool2) : [64 x 800], [7 x 7 x 32 x *1] -> [7 x 7 x 64 x *1]
Validating --> conv3.c.c.sc = LearnableParameter() : -> [64 x 1]
Validating --> conv3.c.c.b = LearnableParameter() : -> [64 x 1]
Validating --> conv3.c.c.m = LearnableParameter() : -> [64 x 1]
Validating --> conv3.c.c.isd = LearnableParameter() : -> [64 x 1]
Validating --> conv3.c.c.y = BatchNormalization (conv3.c.c.c, conv3.c.c.sc, conv3.c.c.b, conv3.c.c.m, conv3.c.c.isd) : [7 x 7 x 64 x *1], [64 x 1], [64 x 1], [64 x 1], [64 x 1] -> [7 x 7 x 64 x *1]
Validating --> conv3.y = RectifiedLinear (conv3.c.c.y) : [7 x 7 x 64 x *1] -> [7 x 7 x 64 x *1]
Validating --> pool3 = MaxPooling (conv3.y) : [7 x 7 x 64 x *1] -> [3 x 3 x 64 x *1]
Validating --> h1.t = Times (h1.W, pool3) : [64 x 3 x 3 x 64], [3 x 3 x 64 x *1] -> [64 x *1]
Validating --> h1.sc = LearnableParameter() : -> [64 x 1]
Validating --> h1.b = LearnableParameter() : -> [64 x 1]
Validating --> h1.m = LearnableParameter() : -> [64 x 1]
Validating --> h1.isd = LearnableParameter() : -> [64 x 1]
Validating --> h1.bn = BatchNormalization (h1.t, h1.sc, h1.b, h1.m, h1.isd) : [64 x *1], [64 x 1], [64 x 1], [64 x 1], [64 x 1] -> [64 x *1]
Validating --> h1.y = RectifiedLinear (h1.bn) : [64 x *1] -> [64 x *1]
Validating --> OutputNodes.t = Times (OutputNodes.W, h1.y) : [10 x 64], [64 x *1] -> [10 x *1]
Validating --> OutputNodes.b = LearnableParameter() : -> [10]
Validating --> OutputNodes.z = Plus (OutputNodes.t, OutputNodes.b) : [10 x *1], [10] -> [10 x *1]
Validating --> CE = CrossEntropyWithSoftmax (labels, OutputNodes.z) : [10 x *1], [10 x *1] -> [1]
Validating --> Err = ErrorPrediction (labels, OutputNodes.z) : [10 x *1], [10 x *1] -> [1]
Validating network. 20 nodes to process in pass 2.
Validating network, final pass.
Using cuDNN convolution engine for geometry: Input: 32 x 32 x 3, Output: 32 x 32 x 32, Kernel: 5 x 5 x 3, Map: 1 x 1 x 32, Stride: 1 x 1 x 3, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
Using cuDNN convolution engine for geometry: Input: 32 x 32 x 32, Output: 15 x 15 x 32, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 15 x 15 x 32, Output: 15 x 15 x 32, Kernel: 5 x 5 x 32, Map: 1 x 1 x 32, Stride: 1 x 1 x 32, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
Using cuDNN convolution engine for geometry: Input: 15 x 15 x 32, Output: 7 x 7 x 32, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 7 x 7 x 32, Output: 7 x 7 x 64, Kernel: 5 x 5 x 32, Map: 1 x 1 x 64, Stride: 1 x 1 x 32, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
Using cuDNN convolution engine for geometry: Input: 7 x 7 x 64, Output: 3 x 3 x 64, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
25 out of 45 nodes do not share the minibatch layout with the input data.
Post-processing network complete.
evalNodeNames are not specified, using all the default evalnodes and training criterion nodes.
Allocating matrices for forward and/or backward propagation.
Memory Sharing Structure:
(nil): {[CE Gradient[1]] [Err Gradient[1]] [OutputNodes.W Gradient[10 x 64]] [OutputNodes.b Gradient[10]] [OutputNodes.t Gradient[10 x *1]] [OutputNodes.z Gradient[10 x *1]] [conv1.c.W Gradient[32 x 75]] [conv1.c.c.b Gradient[32 x 1]] [conv1.c.c.c Gradient[32 x 32 x 32 x *1]] [conv1.c.c.isd Gradient[32 x 1]] [conv1.c.c.m Gradient[32 x 1]] [conv1.c.c.sc Gradient[32 x 1]] [conv1.c.c.y Gradient[32 x 32 x 32 x *1]] [conv1.y Gradient[32 x 32 x 32 x *1]] [conv2.c.W Gradient[32 x 800]] [conv2.c.c.b Gradient[32 x 1]] [conv2.c.c.c Gradient[15 x 15 x 32 x *1]] [conv2.c.c.isd Gradient[32 x 1]] [conv2.c.c.m Gradient[32 x 1]] [conv2.c.c.sc Gradient[32 x 1]] [conv2.c.c.y Gradient[15 x 15 x 32 x *1]] [conv2.y Gradient[15 x 15 x 32 x *1]] [conv3.c.W Gradient[64 x 800]] [conv3.c.c.b Gradient[64 x 1]] [conv3.c.c.c Gradient[7 x 7 x 64 x *1]] [conv3.c.c.isd Gradient[64 x 1]] [conv3.c.c.m Gradient[64 x 1]] [conv3.c.c.sc Gradient[64 x 1]] [conv3.c.c.y Gradient[7 x 7 x 64 x *1]] [conv3.y Gradient[7 x 7 x 64 x *1]] [featOffs Gradient[1 x 1]] [featScaled Gradient[32 x 32 x 3 x *1]] [features Gradient[32 x 32 x 3 x *1]] [h1.W Gradient[64 x 3 x 3 x 64]] [h1.b Gradient[64 x 1]] [h1.bn Gradient[64 x *1]] [h1.isd Gradient[64 x 1]] [h1.m Gradient[64 x 1]] [h1.sc Gradient[64 x 1]] [h1.t Gradient[64 x *1]] [h1.y Gradient[64 x *1]] [labels Gradient[10 x *1]] [pool1 Gradient[15 x 15 x 32 x *1]] [pool2 Gradient[7 x 7 x 32 x *1]] [pool3 Gradient[3 x 3 x 64 x *1]] }
0x7f5132919b38: {[conv2.c.c.isd Value[32 x 1]] }
0x7f5132919cf8: {[conv2.c.c.m Value[32 x 1]] }
0x7f513291a7b8: {[conv2.c.c.sc Value[32 x 1]] }
0x7f5132974258: {[conv1.c.W Value[32 x 75]] }
0x7f51329744c8: {[conv2.c.c.b Value[32 x 1]] }
0x7f5132974fa8: {[conv1.c.c.isd Value[32 x 1]] }
0x7f5132975168: {[conv1.c.c.m Value[32 x 1]] }
0x7f513298b078: {[conv1.y Value[32 x 32 x 32 x *1]] }
0x7f513298b238: {[pool1 Value[15 x 15 x 32 x *1]] }
0x7f513298b3f8: {[conv2.c.c.c Value[15 x 15 x 32 x *1]] }
0x7f513298b778: {[conv2.c.c.y Value[15 x 15 x 32 x *1]] }
0x7f513298bcb8: {[conv2.y Value[15 x 15 x 32 x *1]] }
0x7f513298be78: {[pool2 Value[7 x 7 x 32 x *1]] }
0x7f513298c038: {[conv3.c.c.c Value[7 x 7 x 64 x *1]] }
0x7f513298c3b8: {[conv3.c.c.y Value[7 x 7 x 64 x *1]] }
0x7f513298c8f8: {[conv3.y Value[7 x 7 x 64 x *1]] }
0x7f513298cab8: {[pool3 Value[3 x 3 x 64 x *1]] }
0x7f513298cc78: {[h1.t Value[64 x *1]] }
0x7f513298ce38: {[h1.bn Value[64 x *1]] }
0x7f513298d378: {[h1.y Value[64 x *1]] }
0x7f513298d538: {[OutputNodes.t Value[10 x *1]] }
0x7f513298d6f8: {[OutputNodes.z Value[10 x *1]] }
0x7f51397eafd8: {[featScaled Value[32 x 32 x 3 x *1]] }
0x7f51397eb318: {[conv1.c.c.c Value[32 x 32 x 32 x *1]] }
0x7f5139989708: {[labels Value[10 x *1]] }
0x7f513998a198: {[OutputNodes.b Value[10]] }
0x7f513998aea8: {[OutputNodes.W Value[10 x 64]] }
0x7f513998d7f8: {[conv1.c.c.sc Value[32 x 1]] }
0x7f51399aaca8: {[CE Value[1]] }
0x7f51399ab038: {[Err Value[1]] }
0x7f51399cc578: {[conv1.c.c.y Value[32 x 32 x 32 x *1]] }
0x7f51399ce9c8: {[conv3.c.c.sc Value[64 x 1]] }
0x7f51399cebd8: {[conv3.c.W Value[64 x 800]] }
0x7f51399d04a8: {[featOffs Value[1 x 1]] }
0x7f51399d1148: {[features Value[32 x 32 x 3 x *1]] }
0x7f51399d1b98: {[h1.b Value[64 x 1]] }
0x7f51399d1e78: {[h1.isd Value[64 x 1]] }
0x7f51399d2ab8: {[h1.m Value[64 x 1]] }
0x7f51399d3518: {[h1.sc Value[64 x 1]] }
0x7f51399d4f88: {[conv2.c.W Value[32 x 800]] }
0x7f51399d5a88: {[conv3.c.c.b Value[64 x 1]] }
0x7f51399d6bf8: {[conv3.c.c.isd Value[64 x 1]] }
0x7f51399d6df8: {[conv3.c.c.m Value[64 x 1]] }
0x7f51399f43a8: {[h1.W Value[64 x 3 x 3 x 64]] }
0x7f51399fb388: {[conv1.c.c.b Value[32 x 1]] }
05/13/2016 15:11:09: Final Results: Minibatch[1-625]: Err = 0.86810000 * 10000; CE = 2.32970283 * 10000; perplexity = 10.27488769
05/13/2016 15:11:09: Action "test" complete.
05/13/2016 15:11:09: __COMPLETED__

632
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv/baseline.windows.txt Normal file
Просмотреть файл

@ -0,0 +1,632 @@
=== Running /cygdrive/c/jenkins/workspace/CNTK-Test-Windows-W1/x64/release/cntk.exe configFile=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10/../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/Config/02_BatchNormConv.cntk currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu\TestData RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu\TestData ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10 OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu DeviceId=0 timestamping=true Train=[SGD=[maxEpochs=5]] Train=[SGD=[epochSize=100]] stderr=-
-------------------------------------------------------------------
Build info:
Built time: May 13 2016 08:06:01
Last modified date: Thu May 12 07:31:50 2016
Build type: Release
Build target: GPU
With 1bit-SGD: no
CUDA_PATH: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5
CUB_PATH: c:\src\cub-1.4.1
CUDNN_PATH: c:\NVIDIA\cudnn-4.0\cuda
Build Branch: HEAD
Build SHA1: 35fadc316f045d843bbd9b85061250a959268787
Built by svcphil on Philly-Pool3
Build Path: c:\Jenkins\workspace\CNTK-Build-Windows\Source\CNTK\
-------------------------------------------------------------------
Changed current directory to C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu\TestData
05/13/2016 08:18:23: Redirecting stderr to file -_Train_Test.log
05/13/2016 08:18:23: -------------------------------------------------------------------
05/13/2016 08:18:23: Build info:
05/13/2016 08:18:23: Built time: May 13 2016 08:06:01
05/13/2016 08:18:23: Last modified date: Thu May 12 07:31:50 2016
05/13/2016 08:18:23: Build type: Release
05/13/2016 08:18:23: Build target: GPU
05/13/2016 08:18:23: With 1bit-SGD: no
05/13/2016 08:18:23: CUDA_PATH: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v7.5
05/13/2016 08:18:23: CUB_PATH: c:\src\cub-1.4.1
05/13/2016 08:18:23: CUDNN_PATH: c:\NVIDIA\cudnn-4.0\cuda
05/13/2016 08:18:23: Build Branch: HEAD
05/13/2016 08:18:23: Build SHA1: 35fadc316f045d843bbd9b85061250a959268787
05/13/2016 08:18:23: Built by svcphil on Philly-Pool3
05/13/2016 08:18:23: Build Path: c:\Jenkins\workspace\CNTK-Build-Windows\Source\CNTK\
05/13/2016 08:18:23: -------------------------------------------------------------------
05/13/2016 08:18:23: Running on Philly-Pool2 at 2016/05/13 08:18:23
05/13/2016 08:18:23: Command line:
C:\jenkins\workspace\CNTK-Test-Windows-W1\x64\release\cntk.exe configFile=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10/../../../../Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/Config/02_BatchNormConv.cntk currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu\TestData RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu\TestData ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10 OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu DeviceId=0 timestamping=true Train=[SGD=[maxEpochs=5]] Train=[SGD=[epochSize=100]] stderr=-
05/13/2016 08:18:23: >>>>>>>>>>>>>>>>>>>> RAW CONFIG (VARIABLES NOT RESOLVED) >>>>>>>>>>>>>>>>>>>>
05/13/2016 08:18:23: RootDir = "."
ConfigDir = "$RootDir$"
DataDir = "$RootDir$"
OutputDir = "$RootDir$/Output"
ModelDir = "$OutputDir$/Models"
ndlMacros = "$ConfigDir$/Macros.ndl"
precision = "float"
deviceId = 0
imageLayout = "cudnn"
initOnCPUOnly=true
prefetch = "true"
command = Train:Test
stderr = "$OutputDir$/02_BatchNormConv"
traceLevel = 1
numMBsToShowResult = 500
Train = [
action = "train"
modelPath = "$ModelDir$/02_BatchNormConv"
NDLNetworkBuilder = [
networkDescription = "$ConfigDir$/02_BatchNormConv.ndl"
]
SGD = [
epochSize = 49984
minibatchSize = 64
learningRatesPerMB = 0.03*7:0.01
momentumPerMB = 0
maxEpochs = 10
L2RegWeight = 0
dropoutRate = 0
]
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Train_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
Test = [
action = "test"
modelPath = "$ModelDir$/02_BatchNormConv"
minibatchSize = 16
reader = [
readerType = "CNTKTextFormatReader"
file = "$DataDir$/Test_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu\TestData
RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu
DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu\TestData
ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10
OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu
DeviceId=0
timestamping=true
Train=[SGD=[maxEpochs=5]]
Train=[SGD=[epochSize=100]]
stderr=-
05/13/2016 08:18:23: <<<<<<<<<<<<<<<<<<<< RAW CONFIG (VARIABLES NOT RESOLVED) <<<<<<<<<<<<<<<<<<<<
05/13/2016 08:18:23: >>>>>>>>>>>>>>>>>>>> RAW CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>
05/13/2016 08:18:23: RootDir = "."
ConfigDir = "."
DataDir = "."
OutputDir = "./Output"
ModelDir = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu/Models"
ndlMacros = "C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10/Macros.ndl"
precision = "float"
deviceId = 0
imageLayout = "cudnn"
initOnCPUOnly=true
prefetch = "true"
command = Train:Test
stderr = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu/02_BatchNormConv"
traceLevel = 1
numMBsToShowResult = 500
Train = [
action = "train"
modelPath = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu/Models/02_BatchNormConv"
NDLNetworkBuilder = [
networkDescription = "C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10/02_BatchNormConv.ndl"
]
SGD = [
epochSize = 49984
minibatchSize = 64
learningRatesPerMB = 0.03*7:0.01
momentumPerMB = 0
maxEpochs = 10
L2RegWeight = 0
dropoutRate = 0
]
reader = [
readerType = "CNTKTextFormatReader"
file = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu\TestData/Train_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
Test = [
action = "test"
modelPath = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu/Models/02_BatchNormConv"
minibatchSize = 16
reader = [
readerType = "CNTKTextFormatReader"
file = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu\TestData/Test_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu\TestData
RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu
DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu\TestData
ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10
OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu
DeviceId=0
timestamping=true
Train=[SGD=[maxEpochs=5]]
Train=[SGD=[epochSize=100]]
stderr=-
05/13/2016 08:18:23: <<<<<<<<<<<<<<<<<<<< RAW CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<
05/13/2016 08:18:23: >>>>>>>>>>>>>>>>>>>> PROCESSED CONFIG WITH ALL VARIABLES RESOLVED >>>>>>>>>>>>>>>>>>>>
configparameters: 02_BatchNormConv.cntk:command=Train:Test
configparameters: 02_BatchNormConv.cntk:ConfigDir=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10
configparameters: 02_BatchNormConv.cntk:currentDirectory=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu\TestData
configparameters: 02_BatchNormConv.cntk:DataDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu\TestData
configparameters: 02_BatchNormConv.cntk:deviceId=0
configparameters: 02_BatchNormConv.cntk:imageLayout=cudnn
configparameters: 02_BatchNormConv.cntk:initOnCPUOnly=true
configparameters: 02_BatchNormConv.cntk:ModelDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu/Models
configparameters: 02_BatchNormConv.cntk:ndlMacros=C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10/Macros.ndl
configparameters: 02_BatchNormConv.cntk:numMBsToShowResult=500
configparameters: 02_BatchNormConv.cntk:OutputDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu
configparameters: 02_BatchNormConv.cntk:precision=float
configparameters: 02_BatchNormConv.cntk:prefetch=true
configparameters: 02_BatchNormConv.cntk:RootDir=.
configparameters: 02_BatchNormConv.cntk:RunDir=C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu
configparameters: 02_BatchNormConv.cntk:stderr=-
configparameters: 02_BatchNormConv.cntk:Test=[
action = "test"
modelPath = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu/Models/02_BatchNormConv"
minibatchSize = 16
reader = [
readerType = "CNTKTextFormatReader"
file = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu\TestData/Test_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
]
configparameters: 02_BatchNormConv.cntk:timestamping=true
configparameters: 02_BatchNormConv.cntk:traceLevel=1
configparameters: 02_BatchNormConv.cntk:Train=[
action = "train"
modelPath = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu/Models/02_BatchNormConv"
NDLNetworkBuilder = [
networkDescription = "C:\jenkins\workspace\CNTK-Test-Windows-W1\Examples\Image\Miscellaneous\CIFAR-10/02_BatchNormConv.ndl"
]
SGD = [
epochSize = 49984
minibatchSize = 64
learningRatesPerMB = 0.03*7:0.01
momentumPerMB = 0
maxEpochs = 10
L2RegWeight = 0
dropoutRate = 0
]
reader = [
readerType = "CNTKTextFormatReader"
file = "C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu\TestData/Train_cntk_text.txt"
input = [
features = [
dim = 3072
format = "dense"
]
labels = [
dim = 10
format = "dense"
]
]
]
] [SGD=[maxEpochs=5]] [SGD=[epochSize=100]]
05/13/2016 08:18:23: <<<<<<<<<<<<<<<<<<<< PROCESSED CONFIG WITH ALL VARIABLES RESOLVED <<<<<<<<<<<<<<<<<<<<
05/13/2016 08:18:23: Commands: Train Test
05/13/2016 08:18:23: Precision = "float"
05/13/2016 08:18:23: CNTKModelPath: C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu/Models/02_BatchNormConv
05/13/2016 08:18:23: CNTKCommandTrainInfo: Train : 5
05/13/2016 08:18:23: CNTKCommandTrainInfo: CNTKNoMoreCommands_Total : 5
05/13/2016 08:18:23: ##############################################################################
05/13/2016 08:18:23: # #
05/13/2016 08:18:23: # Action "train" #
05/13/2016 08:18:23: # #
05/13/2016 08:18:23: ##############################################################################
05/13/2016 08:18:23: CNTKCommandTrainBegin: Train
NDLBuilder Using GPU 0
05/13/2016 08:18:24: Creating virgin network.
Microsoft::MSR::CNTK::GPUMatrix<ElemType>::SetGaussianRandomValue (GPU): creating curand object with seed 1, sizeof(ElemType)==4
Post-processing network...
3 roots:
CE = CrossEntropyWithSoftmax()
Err = ErrorPrediction()
OutputNodes.z = Plus()
Validating network. 45 nodes to process in pass 1.
Validating --> labels = InputValue() : -> [10 x *]
Validating --> OutputNodes.W = LearnableParameter() : -> [10 x 64]
Validating --> h1.W = LearnableParameter() : -> [64 x 3 x 3 x 64]
Validating --> conv3.c.W = LearnableParameter() : -> [64 x 800]
Validating --> conv2.c.W = LearnableParameter() : -> [32 x 800]
Validating --> conv1.c.W = LearnableParameter() : -> [32 x 75]
Validating --> features = InputValue() : -> [32 x 32 x 3 x *]
Validating --> featOffs = LearnableParameter() : -> [1 x 1]
Validating --> featScaled = Minus (features, featOffs) : [32 x 32 x 3 x *], [1 x 1] -> [32 x 32 x 3 x *]
Validating --> conv1.c.c.c = Convolution (conv1.c.W, featScaled) : [32 x 75], [32 x 32 x 3 x *] -> [32 x 32 x 32 x *]
Validating --> conv1.c.c.sc = LearnableParameter() : -> [32 x 1]
Validating --> conv1.c.c.b = LearnableParameter() : -> [32 x 1]
Validating --> conv1.c.c.m = LearnableParameter() : -> [32 x 1]
Validating --> conv1.c.c.isd = LearnableParameter() : -> [32 x 1]
Validating --> conv1.c.c.y = BatchNormalization (conv1.c.c.c, conv1.c.c.sc, conv1.c.c.b, conv1.c.c.m, conv1.c.c.isd) : [32 x 32 x 32 x *], [32 x 1], [32 x 1], [32 x 1], [32 x 1] -> [32 x 32 x 32 x *]
Validating --> conv1.y = RectifiedLinear (conv1.c.c.y) : [32 x 32 x 32 x *] -> [32 x 32 x 32 x *]
Validating --> pool1 = MaxPooling (conv1.y) : [32 x 32 x 32 x *] -> [15 x 15 x 32 x *]
Validating --> conv2.c.c.c = Convolution (conv2.c.W, pool1) : [32 x 800], [15 x 15 x 32 x *] -> [15 x 15 x 32 x *]
Validating --> conv2.c.c.sc = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.b = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.m = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.isd = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.y = BatchNormalization (conv2.c.c.c, conv2.c.c.sc, conv2.c.c.b, conv2.c.c.m, conv2.c.c.isd) : [15 x 15 x 32 x *], [32 x 1], [32 x 1], [32 x 1], [32 x 1] -> [15 x 15 x 32 x *]
Validating --> conv2.y = RectifiedLinear (conv2.c.c.y) : [15 x 15 x 32 x *] -> [15 x 15 x 32 x *]
Validating --> pool2 = MaxPooling (conv2.y) : [15 x 15 x 32 x *] -> [7 x 7 x 32 x *]
Validating --> conv3.c.c.c = Convolution (conv3.c.W, pool2) : [64 x 800], [7 x 7 x 32 x *] -> [7 x 7 x 64 x *]
Validating --> conv3.c.c.sc = LearnableParameter() : -> [64 x 1]
Validating --> conv3.c.c.b = LearnableParameter() : -> [64 x 1]
Validating --> conv3.c.c.m = LearnableParameter() : -> [64 x 1]
Validating --> conv3.c.c.isd = LearnableParameter() : -> [64 x 1]
Validating --> conv3.c.c.y = BatchNormalization (conv3.c.c.c, conv3.c.c.sc, conv3.c.c.b, conv3.c.c.m, conv3.c.c.isd) : [7 x 7 x 64 x *], [64 x 1], [64 x 1], [64 x 1], [64 x 1] -> [7 x 7 x 64 x *]
Validating --> conv3.y = RectifiedLinear (conv3.c.c.y) : [7 x 7 x 64 x *] -> [7 x 7 x 64 x *]
Validating --> pool3 = MaxPooling (conv3.y) : [7 x 7 x 64 x *] -> [3 x 3 x 64 x *]
Validating --> h1.t = Times (h1.W, pool3) : [64 x 3 x 3 x 64], [3 x 3 x 64 x *] -> [64 x *]
Validating --> h1.sc = LearnableParameter() : -> [64 x 1]
Validating --> h1.b = LearnableParameter() : -> [64 x 1]
Validating --> h1.m = LearnableParameter() : -> [64 x 1]
Validating --> h1.isd = LearnableParameter() : -> [64 x 1]
Validating --> h1.bn = BatchNormalization (h1.t, h1.sc, h1.b, h1.m, h1.isd) : [64 x *], [64 x 1], [64 x 1], [64 x 1], [64 x 1] -> [64 x *]
Validating --> h1.y = RectifiedLinear (h1.bn) : [64 x *] -> [64 x *]
Validating --> OutputNodes.t = Times (OutputNodes.W, h1.y) : [10 x 64], [64 x *] -> [10 x *]
Validating --> OutputNodes.b = LearnableParameter() : -> [10]
Validating --> OutputNodes.z = Plus (OutputNodes.t, OutputNodes.b) : [10 x *], [10] -> [10 x *]
Validating --> CE = CrossEntropyWithSoftmax (labels, OutputNodes.z) : [10 x *], [10 x *] -> [1]
Validating --> Err = ErrorPrediction (labels, OutputNodes.z) : [10 x *], [10 x *] -> [1]
Validating network. 20 nodes to process in pass 2.
Validating network, final pass.
Using cuDNN convolution engine for geometry: Input: 32 x 32 x 3, Output: 32 x 32 x 32, Kernel: 5 x 5 x 3, Map: 1 x 1 x 32, Stride: 1 x 1 x 3, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
Using cuDNN convolution engine for geometry: Input: 32 x 32 x 32, Output: 15 x 15 x 32, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 15 x 15 x 32, Output: 15 x 15 x 32, Kernel: 5 x 5 x 32, Map: 1 x 1 x 32, Stride: 1 x 1 x 32, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
Using cuDNN convolution engine for geometry: Input: 15 x 15 x 32, Output: 7 x 7 x 32, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 7 x 7 x 32, Output: 7 x 7 x 64, Kernel: 5 x 5 x 32, Map: 1 x 1 x 64, Stride: 1 x 1 x 32, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
Using cuDNN convolution engine for geometry: Input: 7 x 7 x 64, Output: 3 x 3 x 64, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
25 out of 45 nodes do not share the minibatch layout with the input data.
Post-processing network complete.
05/13/2016 08:18:26: Created model with 45 nodes on GPU 0.
05/13/2016 08:18:26: Training criterion node(s):
05/13/2016 08:18:26: CE = CrossEntropyWithSoftmax
05/13/2016 08:18:26: Evaluation criterion node(s):
05/13/2016 08:18:26: Err = ErrorPrediction
Allocating matrices for forward and/or backward propagation.
Memory Sharing Structure:
0000000000000000: {[Err Gradient[1]] [conv1.c.c.isd Gradient[32 x 1]] [conv1.c.c.m Gradient[32 x 1]] [conv2.c.c.isd Gradient[32 x 1]] [conv2.c.c.m Gradient[32 x 1]] [conv3.c.c.isd Gradient[64 x 1]] [conv3.c.c.m Gradient[64 x 1]] [featOffs Gradient[1 x 1]] [featScaled Gradient[32 x 32 x 3 x *]] [features Gradient[32 x 32 x 3 x *]] [h1.isd Gradient[64 x 1]] [h1.m Gradient[64 x 1]] [labels Gradient[10 x *]] }
000000E89AC81140: {[conv3.c.c.sc Value[64 x 1]] }
000000E89AC813C0: {[conv2.c.c.sc Value[32 x 1]] }
000000E89AC815A0: {[conv2.c.c.b Value[32 x 1]] }
000000E89AC81820: {[h1.isd Value[64 x 1]] }
000000E89AC81A00: {[OutputNodes.W Value[10 x 64]] }
000000E89AC81BE0: {[OutputNodes.b Value[10]] }
000000E89AC81F00: {[h1.W Value[64 x 3 x 3 x 64]] }
000000E89AC81FA0: {[h1.m Value[64 x 1]] }
000000E89AC82180: {[conv2.c.c.m Value[32 x 1]] }
000000E89AC822C0: {[conv3.c.c.isd Value[64 x 1]] }
000000E89AC82540: {[h1.sc Value[64 x 1]] }
000000E89AC825E0: {[conv2.c.c.isd Value[32 x 1]] }
000000E89AC82680: {[conv3.c.c.m Value[64 x 1]] }
000000E89AC82720: {[h1.b Value[64 x 1]] }
000000E89AC82860: {[conv3.c.W Value[64 x 800]] }
000000E89AC82AE0: {[conv2.c.W Value[32 x 800]] }
000000E89AC82B80: {[conv3.c.c.b Value[64 x 1]] }
000000E8A0568140: {[featScaled Value[32 x 32 x 3 x *]] }
000000E8A05681E0: {[conv2.c.c.y Gradient[15 x 15 x 32 x *]] [pool2 Value[7 x 7 x 32 x *]] }
000000E8A0568280: {[conv2.c.c.sc Gradient[32 x 1]] [conv2.y Gradient[15 x 15 x 32 x *]] }
000000E8A0568320: {[conv3.c.c.y Value[7 x 7 x 64 x *]] }
000000E8A0568460: {[conv2.c.c.b Gradient[32 x 1]] [conv3.c.c.c Gradient[7 x 7 x 64 x *]] [conv3.y Value[7 x 7 x 64 x *]] }
000000E8A05685A0: {[OutputNodes.t Value[10 x *]] [h1.bn Gradient[64 x *]] }
000000E8A0568A00: {[Err Value[1]] }
000000E8A0568AA0: {[conv2.c.c.y Value[15 x 15 x 32 x *]] }
000000E8A0568BE0: {[conv1.c.c.b Gradient[32 x 1]] [conv2.c.c.c Gradient[15 x 15 x 32 x *]] [conv2.y Value[15 x 15 x 32 x *]] }
000000E8A0568D20: {[conv3.c.c.b Gradient[64 x 1]] }
000000E8A0568DC0: {[conv3.c.c.sc Gradient[64 x 1]] [conv3.y Gradient[7 x 7 x 64 x *]] [h1.t Value[64 x *]] }
000000E8A0568E60: {[conv3.c.W Gradient[64 x 800]] [h1.t Gradient[64 x *]] [h1.y Value[64 x *]] }
000000E8A0569040: {[conv1.c.c.y Gradient[32 x 32 x 32 x *]] [pool1 Value[15 x 15 x 32 x *]] }
000000E8A0569400: {[conv1.c.c.y Value[32 x 32 x 32 x *]] }
000000E8A05694A0: {[conv2.c.W Gradient[32 x 800]] [conv3.c.c.c Value[7 x 7 x 64 x *]] }
000000E8A0569540: {[OutputNodes.W Gradient[10 x 64]] [OutputNodes.z Gradient[10 x *]] }
000000E8A0569680: {[OutputNodes.t Gradient[10 x *]] [pool1 Gradient[15 x 15 x 32 x *]] [pool2 Gradient[7 x 7 x 32 x *]] [pool3 Gradient[3 x 3 x 64 x *]] }
000000E8A0569720: {[OutputNodes.b Gradient[10]] }
000000E8A05697C0: {[h1.sc Gradient[64 x 1]] [h1.y Gradient[64 x *]] }
000000E8A0569860: {[conv1.c.W Gradient[32 x 75]] [conv2.c.c.c Value[15 x 15 x 32 x *]] }
000000E8A0569900: {[conv1.c.c.c Gradient[32 x 32 x 32 x *]] [conv1.y Value[32 x 32 x 32 x *]] }
000000E8A05699A0: {[CE Gradient[1]] }
000000E8A0569A40: {[h1.W Gradient[64 x 3 x 3 x 64]] }
000000E8A0569B80: {[conv3.c.c.y Gradient[7 x 7 x 64 x *]] [pool3 Value[3 x 3 x 64 x *]] }
000000E8A0569E00: {[h1.bn Value[64 x *]] }
000000E8A0569FE0: {[h1.b Gradient[64 x 1]] }
000000E8A056A120: {[conv1.c.c.sc Gradient[32 x 1]] [conv1.y Gradient[32 x 32 x 32 x *]] }
000000E8A056A3A0: {[CE Value[1]] }
000000E8A056A620: {[OutputNodes.z Value[10 x *]] }
000000E8A056A760: {[conv1.c.c.c Value[32 x 32 x 32 x *]] }
000000E8FC080980: {[featOffs Value[1 x 1]] }
000000E8FC0811A0: {[conv1.c.W Value[32 x 75]] }
000000E8FC081240: {[conv1.c.c.b Value[32 x 1]] }
000000E8FC081740: {[conv1.c.c.sc Value[32 x 1]] }
000000E8FC081920: {[labels Value[10 x *]] }
000000E8FC081D80: {[features Value[32 x 32 x 3 x *]] }
000000E8FC081EC0: {[conv1.c.c.m Value[32 x 1]] }
000000E8FC081F60: {[conv1.c.c.isd Value[32 x 1]] }
05/13/2016 08:18:26: No PreCompute nodes found, skipping PreCompute step.
05/13/2016 08:18:26: Starting Epoch 1: learning rate per sample = 0.000469 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 08:18:26: Starting minibatch loop.
05/13/2016 08:18:35: Finished Epoch[ 1 of 5]: [Training] CE = 2.31451355 * 100; Err = 0.87000000 * 100; totalSamplesSeen = 100; learningRatePerSample = 0.00046874999; epochTime=9.33323s
05/13/2016 08:18:36: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu/Models/02_BatchNormConv.1'
05/13/2016 08:18:37: Starting Epoch 2: learning rate per sample = 0.000469 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 08:18:37: Starting minibatch loop.
05/13/2016 08:18:37: Finished Epoch[ 2 of 5]: [Training] CE = 2.27380722 * 100; Err = 0.82000000 * 100; totalSamplesSeen = 200; learningRatePerSample = 0.00046874999; epochTime=0.020597s
05/13/2016 08:18:37: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu/Models/02_BatchNormConv.2'
05/13/2016 08:18:37: Starting Epoch 3: learning rate per sample = 0.000469 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 08:18:37: Starting minibatch loop.
05/13/2016 08:18:37: Finished Epoch[ 3 of 5]: [Training] CE = 2.25248398 * 100; Err = 0.83000000 * 100; totalSamplesSeen = 300; learningRatePerSample = 0.00046874999; epochTime=0.020236s
05/13/2016 08:18:37: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu/Models/02_BatchNormConv.3'
05/13/2016 08:18:37: Starting Epoch 4: learning rate per sample = 0.000469 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 08:18:37: Starting minibatch loop.
05/13/2016 08:18:37: Finished Epoch[ 4 of 5]: [Training] CE = 2.15781601 * 100; Err = 0.77000000 * 100; totalSamplesSeen = 400; learningRatePerSample = 0.00046874999; epochTime=0.020351s
05/13/2016 08:18:37: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu/Models/02_BatchNormConv.4'
05/13/2016 08:18:37: Starting Epoch 5: learning rate per sample = 0.000469 effective momentum = 0.000000 momentum as time constant = 0.0 samples
05/13/2016 08:18:37: Starting minibatch loop.
05/13/2016 08:18:37: Finished Epoch[ 5 of 5]: [Training] CE = 2.12939789 * 100; Err = 0.71000000 * 100; totalSamplesSeen = 500; learningRatePerSample = 0.00046874999; epochTime=0.02018s
05/13/2016 08:18:37: SGD: Saving checkpoint model 'C:\Users\svcphil\AppData\Local\Temp\cntk-test-20160513081543.861015\CNTKTextFormatReader\Examples\Image\Miscellaneous\CIFAR-10_02_BatchNormConv@release_gpu/Models/02_BatchNormConv'
05/13/2016 08:18:37: CNTKCommandTrainEnd: Train
05/13/2016 08:18:37: Action "train" complete.
05/13/2016 08:18:37: ##############################################################################
05/13/2016 08:18:37: # #
05/13/2016 08:18:37: # Action "test" #
05/13/2016 08:18:37: # #
05/13/2016 08:18:37: ##############################################################################
Post-processing network...
3 roots:
CE = CrossEntropyWithSoftmax()
Err = ErrorPrediction()
OutputNodes.z = Plus()
Validating network. 45 nodes to process in pass 1.
Validating --> labels = InputValue() : -> [10 x *1]
Validating --> OutputNodes.W = LearnableParameter() : -> [10 x 64]
Validating --> h1.W = LearnableParameter() : -> [64 x 3 x 3 x 64]
Validating --> conv3.c.W = LearnableParameter() : -> [64 x 800]
Validating --> conv2.c.W = LearnableParameter() : -> [32 x 800]
Validating --> conv1.c.W = LearnableParameter() : -> [32 x 75]
Validating --> features = InputValue() : -> [32 x 32 x 3 x *1]
Validating --> featOffs = LearnableParameter() : -> [1 x 1]
Validating --> featScaled = Minus (features, featOffs) : [32 x 32 x 3 x *1], [1 x 1] -> [32 x 32 x 3 x *1]
Validating --> conv1.c.c.c = Convolution (conv1.c.W, featScaled) : [32 x 75], [32 x 32 x 3 x *1] -> [32 x 32 x 32 x *1]
Validating --> conv1.c.c.sc = LearnableParameter() : -> [32 x 1]
Validating --> conv1.c.c.b = LearnableParameter() : -> [32 x 1]
Validating --> conv1.c.c.m = LearnableParameter() : -> [32 x 1]
Validating --> conv1.c.c.isd = LearnableParameter() : -> [32 x 1]
Validating --> conv1.c.c.y = BatchNormalization (conv1.c.c.c, conv1.c.c.sc, conv1.c.c.b, conv1.c.c.m, conv1.c.c.isd) : [32 x 32 x 32 x *1], [32 x 1], [32 x 1], [32 x 1], [32 x 1] -> [32 x 32 x 32 x *1]
Validating --> conv1.y = RectifiedLinear (conv1.c.c.y) : [32 x 32 x 32 x *1] -> [32 x 32 x 32 x *1]
Validating --> pool1 = MaxPooling (conv1.y) : [32 x 32 x 32 x *1] -> [15 x 15 x 32 x *1]
Validating --> conv2.c.c.c = Convolution (conv2.c.W, pool1) : [32 x 800], [15 x 15 x 32 x *1] -> [15 x 15 x 32 x *1]
Validating --> conv2.c.c.sc = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.b = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.m = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.isd = LearnableParameter() : -> [32 x 1]
Validating --> conv2.c.c.y = BatchNormalization (conv2.c.c.c, conv2.c.c.sc, conv2.c.c.b, conv2.c.c.m, conv2.c.c.isd) : [15 x 15 x 32 x *1], [32 x 1], [32 x 1], [32 x 1], [32 x 1] -> [15 x 15 x 32 x *1]
Validating --> conv2.y = RectifiedLinear (conv2.c.c.y) : [15 x 15 x 32 x *1] -> [15 x 15 x 32 x *1]
Validating --> pool2 = MaxPooling (conv2.y) : [15 x 15 x 32 x *1] -> [7 x 7 x 32 x *1]
Validating --> conv3.c.c.c = Convolution (conv3.c.W, pool2) : [64 x 800], [7 x 7 x 32 x *1] -> [7 x 7 x 64 x *1]
Validating --> conv3.c.c.sc = LearnableParameter() : -> [64 x 1]
Validating --> conv3.c.c.b = LearnableParameter() : -> [64 x 1]
Validating --> conv3.c.c.m = LearnableParameter() : -> [64 x 1]
Validating --> conv3.c.c.isd = LearnableParameter() : -> [64 x 1]
Validating --> conv3.c.c.y = BatchNormalization (conv3.c.c.c, conv3.c.c.sc, conv3.c.c.b, conv3.c.c.m, conv3.c.c.isd) : [7 x 7 x 64 x *1], [64 x 1], [64 x 1], [64 x 1], [64 x 1] -> [7 x 7 x 64 x *1]
Validating --> conv3.y = RectifiedLinear (conv3.c.c.y) : [7 x 7 x 64 x *1] -> [7 x 7 x 64 x *1]
Validating --> pool3 = MaxPooling (conv3.y) : [7 x 7 x 64 x *1] -> [3 x 3 x 64 x *1]
Validating --> h1.t = Times (h1.W, pool3) : [64 x 3 x 3 x 64], [3 x 3 x 64 x *1] -> [64 x *1]
Validating --> h1.sc = LearnableParameter() : -> [64 x 1]
Validating --> h1.b = LearnableParameter() : -> [64 x 1]
Validating --> h1.m = LearnableParameter() : -> [64 x 1]
Validating --> h1.isd = LearnableParameter() : -> [64 x 1]
Validating --> h1.bn = BatchNormalization (h1.t, h1.sc, h1.b, h1.m, h1.isd) : [64 x *1], [64 x 1], [64 x 1], [64 x 1], [64 x 1] -> [64 x *1]
Validating --> h1.y = RectifiedLinear (h1.bn) : [64 x *1] -> [64 x *1]
Validating --> OutputNodes.t = Times (OutputNodes.W, h1.y) : [10 x 64], [64 x *1] -> [10 x *1]
Validating --> OutputNodes.b = LearnableParameter() : -> [10]
Validating --> OutputNodes.z = Plus (OutputNodes.t, OutputNodes.b) : [10 x *1], [10] -> [10 x *1]
Validating --> CE = CrossEntropyWithSoftmax (labels, OutputNodes.z) : [10 x *1], [10 x *1] -> [1]
Validating --> Err = ErrorPrediction (labels, OutputNodes.z) : [10 x *1], [10 x *1] -> [1]
Validating network. 20 nodes to process in pass 2.
Validating network, final pass.
Using cuDNN convolution engine for geometry: Input: 32 x 32 x 3, Output: 32 x 32 x 32, Kernel: 5 x 5 x 3, Map: 1 x 1 x 32, Stride: 1 x 1 x 3, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
Using cuDNN convolution engine for geometry: Input: 32 x 32 x 32, Output: 15 x 15 x 32, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 15 x 15 x 32, Output: 15 x 15 x 32, Kernel: 5 x 5 x 32, Map: 1 x 1 x 32, Stride: 1 x 1 x 32, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
Using cuDNN convolution engine for geometry: Input: 15 x 15 x 32, Output: 7 x 7 x 32, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using cuDNN convolution engine for geometry: Input: 7 x 7 x 32, Output: 7 x 7 x 64, Kernel: 5 x 5 x 32, Map: 1 x 1 x 64, Stride: 1 x 1 x 32, Sharing: (1), AutoPad: (1), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
Using cuDNN convolution engine for geometry: Input: 7 x 7 x 64, Output: 3 x 3 x 64, Kernel: 3 x 3 x 1, Map: 1, Stride: 2 x 2 x 1, Sharing: (1), AutoPad: (0), LowerPad: 0, UpperPad: 0.
Using CNTK batch normalization engine.
25 out of 45 nodes do not share the minibatch layout with the input data.
Post-processing network complete.
evalNodeNames are not specified, using all the default evalnodes and training criterion nodes.
Allocating matrices for forward and/or backward propagation.
Memory Sharing Structure:
0000000000000000: {[CE Gradient[1]] [Err Gradient[1]] [OutputNodes.W Gradient[10 x 64]] [OutputNodes.b Gradient[10]] [OutputNodes.t Gradient[10 x *1]] [OutputNodes.z Gradient[10 x *1]] [conv1.c.W Gradient[32 x 75]] [conv1.c.c.b Gradient[32 x 1]] [conv1.c.c.c Gradient[32 x 32 x 32 x *1]] [conv1.c.c.isd Gradient[32 x 1]] [conv1.c.c.m Gradient[32 x 1]] [conv1.c.c.sc Gradient[32 x 1]] [conv1.c.c.y Gradient[32 x 32 x 32 x *1]] [conv1.y Gradient[32 x 32 x 32 x *1]] [conv2.c.W Gradient[32 x 800]] [conv2.c.c.b Gradient[32 x 1]] [conv2.c.c.c Gradient[15 x 15 x 32 x *1]] [conv2.c.c.isd Gradient[32 x 1]] [conv2.c.c.m Gradient[32 x 1]] [conv2.c.c.sc Gradient[32 x 1]] [conv2.c.c.y Gradient[15 x 15 x 32 x *1]] [conv2.y Gradient[15 x 15 x 32 x *1]] [conv3.c.W Gradient[64 x 800]] [conv3.c.c.b Gradient[64 x 1]] [conv3.c.c.c Gradient[7 x 7 x 64 x *1]] [conv3.c.c.isd Gradient[64 x 1]] [conv3.c.c.m Gradient[64 x 1]] [conv3.c.c.sc Gradient[64 x 1]] [conv3.c.c.y Gradient[7 x 7 x 64 x *1]] [conv3.y Gradient[7 x 7 x 64 x *1]] [featOffs Gradient[1 x 1]] [featScaled Gradient[32 x 32 x 3 x *1]] [features Gradient[32 x 32 x 3 x *1]] [h1.W Gradient[64 x 3 x 3 x 64]] [h1.b Gradient[64 x 1]] [h1.bn Gradient[64 x *1]] [h1.isd Gradient[64 x 1]] [h1.m Gradient[64 x 1]] [h1.sc Gradient[64 x 1]] [h1.t Gradient[64 x *1]] [h1.y Gradient[64 x *1]] [labels Gradient[10 x *1]] [pool1 Gradient[15 x 15 x 32 x *1]] [pool2 Gradient[7 x 7 x 32 x *1]] [pool3 Gradient[3 x 3 x 64 x *1]] }
000000E8A05681E0: {[conv2.c.c.c Value[15 x 15 x 32 x *1]] }
000000E8A0568460: {[conv1.c.c.c Value[32 x 32 x 32 x *1]] }
000000E8A05685A0: {[conv1.c.c.y Value[32 x 32 x 32 x *1]] }
000000E8A0568A00: {[CE Value[1]] }
000000E8A0568AA0: {[conv2.y Value[15 x 15 x 32 x *1]] }
000000E8A0568B40: {[h1.y Value[64 x *1]] }
000000E8A0568D20: {[featScaled Value[32 x 32 x 3 x *1]] }
000000E8A0568DC0: {[pool1 Value[15 x 15 x 32 x *1]] }
000000E8A05694A0: {[conv1.y Value[32 x 32 x 32 x *1]] }
000000E8A0569540: {[pool3 Value[3 x 3 x 64 x *1]] }
000000E8A0569680: {[OutputNodes.t Value[10 x *1]] }
000000E8A0569720: {[OutputNodes.z Value[10 x *1]] }
000000E8A05697C0: {[conv3.y Value[7 x 7 x 64 x *1]] }
000000E8A05699A0: {[conv3.c.c.y Value[7 x 7 x 64 x *1]] }
000000E8A0569D60: {[conv2.c.c.y Value[15 x 15 x 32 x *1]] }
000000E8A0569E00: {[h1.t Value[64 x *1]] }
000000E8A0569F40: {[conv3.c.c.c Value[7 x 7 x 64 x *1]] }
000000E8A056A080: {[Err Value[1]] }
000000E8A056A3A0: {[pool2 Value[7 x 7 x 32 x *1]] }
000000E8A056A620: {[h1.bn Value[64 x *1]] }
000000E8A16A32D0: {[h1.sc Value[64 x 1]] }
000000E8A16A3870: {[conv2.c.c.b Value[32 x 1]] }
000000E8A16A3C30: {[conv1.c.c.isd Value[32 x 1]] }
000000E8A16A3CD0: {[conv2.c.c.sc Value[32 x 1]] }
000000E8A16A3E10: {[conv3.c.c.b Value[64 x 1]] }
000000E8A16A3F50: {[conv1.c.c.b Value[32 x 1]] }
000000E8A16A4090: {[conv2.c.c.isd Value[32 x 1]] }
000000E8A16A4310: {[conv3.c.c.sc Value[64 x 1]] }
000000E8A16A4630: {[conv1.c.c.sc Value[32 x 1]] }
000000E8A16A46D0: {[conv1.c.W Value[32 x 75]] }
000000E8A16A4A90: {[conv1.c.c.m Value[32 x 1]] }
000000E8A16A4B30: {[conv3.c.W Value[64 x 800]] }
000000E8A16A4EF0: {[conv2.c.W Value[32 x 800]] }
000000E8A16A4F90: {[conv3.c.c.m Value[64 x 1]] }
000000E8A16A5030: {[featOffs Value[1 x 1]] }
000000E8A16A50D0: {[conv3.c.c.isd Value[64 x 1]] }
000000E8A16A5350: {[conv2.c.c.m Value[32 x 1]] }
000000E8A16A53F0: {[features Value[32 x 32 x 3 x *1]] }
000000E8A16A5530: {[h1.b Value[64 x 1]] }
000000E8A16A57B0: {[h1.isd Value[64 x 1]] }
000000E8A16A58F0: {[h1.m Value[64 x 1]] }
000000E8A16A5CB0: {[labels Value[10 x *1]] }
000000E8A16A6110: {[OutputNodes.W Value[10 x 64]] }
000000E8A16A61B0: {[OutputNodes.b Value[10]] }
000000E8A16A6930: {[h1.W Value[64 x 3 x 3 x 64]] }
05/13/2016 08:18:52: Final Results: Minibatch[1-625]: Err = 0.84580000 * 10000; CE = 2.27296712 * 10000; perplexity = 9.70816338
05/13/2016 08:18:52: Action "test" complete.
05/13/2016 08:18:52: __COMPLETED__

12
Tests/EndToEndTests/CNTKTextFormatReader/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv/run-test Executable file
Просмотреть файл

@ -0,0 +1,12 @@
#!/bin/bash
. $TEST_DIR/../run-test-common
# cntkrun <CNTK config file name> <additional CNTK args>
cntkrun $CNTKTextConfigDir/02_BatchNormConv.cntk "Train=[SGD=[maxEpochs=5]] Train=[SGD=[epochSize=100]] stderr=-"
ExitCode=$?
# Delete the test data if copied
[[ "$Copied" -eq "1" ]] && rm -rf "$DataDir"
exit $ExitCode

Некоторые файлы не были показаны из-за слишком большого количества измененных файлов Показать больше