64fae7a0bc
Implementation of data-parallel SGD with 1bit quantization of gradients for reduced communication costs
2015-08-05 09:23:33 -07:00
8794397ab9
DiagonalNode and DebugUtil.
2015-08-05 10:44:09 +10:00
d7c06ef221
Small fix to sequence training
2015-08-04 19:45:16 -04:00
21d35a0552
Bug fix: fixing dimensionality mismatch after using MEL for sequence training model
2015-08-03 18:30:51 -04:00
ad5b8a20f7
Fix g++ compilation warnings
...
sprintf_s is snprintf, not sprintf
Need to use more symbols from SGD
Add externs for missing file name generators that seem to be not reachable
and not defined.
2015-08-03 14:36:33 -04:00
d79e00090c
Fixing mixing LSTM state when buffering minibatches
2015-08-01 01:11:53 +00:00
5996b45890
Some cosmetic change in Kaldi2Reader
2015-07-30 18:00:05 +00:00
b417ea8caf
Limit number of overflow warnings in Kaldi2Reader; Add configuration for setting overflow value
2015-07-30 17:16:32 +00:00
d8c5334b41
comment out unnecessary output check and print in Kaldi2Reader that causes problems with non-probability outputs
2015-07-30 04:31:33 -04:00
0000f40a58
Merge branch 'linux-gcc' of https://git01.codeplex.com/cntk into linux-gcc
2015-07-29 10:40:57 -07:00
61514b108a
refine PastValueNode and FutureValueNode. Fixed a bug in FutureValueNode When it's not inside a loop.
2015-07-29 10:40:45 -07:00
95776f8818
Bug fixing regarding m_uttDerivBuffer is NULL
2015-07-29 16:04:21 +00:00
2489646708
change the pastValueNode and FutureValueNode to support NO_FEATURE and NO_LABEL cases better.
2015-07-28 23:55:21 -07:00
a31304adf8
Adding interface for utterance derivative computation in the Kaldi2Reader
2015-07-29 03:30:56 +00:00
94e040a723
Fixing end-of-epoch issue for sequence training
2015-07-27 08:59:16 +00:00
3cdea4e670
Skipping utterances without lattice or alignment in sequence training
2015-07-26 22:14:24 +00:00
c1925bb606
adding verbosity support in DataReader/Kaldi2Reader/HTKMLFWriter.cpp
2015-07-25 09:53:30 +00:00
d98c21f3b3
Changing NO_LABEL to NO_INPUT in Kaldi2Reader
2015-07-25 00:14:48 +00:00
a0bb630616
updating sentence boundary mark for Kaldi2Reader, HTKMLFReader and KaldiReader
2015-07-24 23:04:31 +00:00
4b9c8b25fc
Merge branch 'linux-gcc' of https://git.codeplex.com/cntk into linux-gcc
2015-07-24 16:55:29 +00:00
f0107498aa
Adding suppport for parallelized sequence training in Kaldi2Reader
2015-07-24 16:55:21 +00:00
fb02f1509a
Merge remote-tracking branch 'origin/master' into linux-gcc
2015-07-23 10:17:26 -04:00
a839c118b0
Fix some minor bug in Kaldi2Reader.
2015-07-23 10:07:54 -04:00
fce1212959
Merge branch 'master' of https://git01.codeplex.com/cntk into linux-gcc
...
Conflicts:
DataReader/LMSequenceReader/SequenceReader.cpp
2015-07-22 09:53:02 -07:00
c437f852b6
Rename NDRMParser to SparsePCParser.
2015-07-23 00:15:09 +10:00
41d9515814
Converting NDRM reader to SparsePC reader.
2015-07-23 00:02:39 +10:00
a48278ae08
Remove useless merged code
2015-07-21 21:30:10 -07:00
8c77b0cbe0
Merge branch 'master' of https://git01.codeplex.com/cntk
...
Conflicts:
DataReader/LMSequenceReader/SequenceReader.cpp
modified: DataReader/LMSequenceReader/SequenceReader.cpp
2015-07-21 21:26:24 -07:00
94058b109f
set both for Label Outputs in SequenceReader
2015-07-21 21:24:53 -07:00
a567807aa2
Use CPUdevice only for class-based cross entropy for LM reader
2015-07-21 08:54:52 -07:00
54e9809e42
LM Sequence reader to use CPU device for labels. Also updated LSTM LM training setup.
2015-07-21 08:07:59 -07:00
9ffa9df9c6
Add transfer in getoutput
2015-07-21 06:01:10 -07:00
ae76821812
Set shuffle always true, set m_evalMode under old version
2015-07-21 05:35:16 -07:00
3e46e6e78d
fix dataend issue, and support old-version model for test
...
modified: DataReader/LMSequenceReader/SequenceReader.cpp
modified: MachineLearning/CNTK/TrainingCriterionNodes.h
2015-07-20 20:11:52 -07:00
c74cdb58f9
Merge branch 'master' of https://git01.codeplex.com/cntk into linux-gcc
2015-07-20 16:57:26 -07:00
9bb8d4df3a
minor changes based on code review
2015-07-20 12:50:47 -07:00
45e79ad6d8
fix problems in reader when files have duration mismatch
2015-07-20 12:50:45 -07:00
d1e14783b6
changed feature packing flag names from utteranceXYZ and sentence_XYZ to SequenceXYZ to make them consistent and be understood for both speech and text people.
...
Added flag to indicate the condition of no feature (with assumption that there is also no label) to support efficient bi-directional model training.
Updated the windows version of HTKMLFReader, LM and LU Sequence readers.
Caution: this change will not allow for using past and future label information in the recurrent networks. The ultimate solution is to have a flag for each feature read-in. However, this will require huge change and we will do this after reader refactorization to reduce effort.
2015-07-18 17:28:14 -07:00
bbf49009d1
Merge branch 'master' to 'linux-gcc'
2015-07-17 23:42:22 +00:00
d54e2d30fe
Adding multi-utterance support in Kalsi2Reader for Truncated == false
2015-07-17 21:25:10 +00:00
abc106dd2f
Merge branch 'master' into linux-gcc
2015-07-14 21:52:08 +00:00
8515a8d2f9
Merge branch 'master' of https://git.codeplex.com/cntk
2015-07-13 22:10:34 -04:00
fda6fb62ce
Fix multi-io bug for Kaldi2Reader.
2015-07-13 22:09:43 -04:00
016f112b21
Bugfix for linux reader: 'no labels for xxx.mfc'
2015-07-13 15:03:48 -07:00
c944e0da24
Fix bug when m_truncated=false.
2015-07-13 05:49:16 -04:00
d75b3392ca
Fix build and fix assertion error when we do multi-io training.
2015-07-12 01:17:31 -04:00
6ae4115b4d
Merge branch 'master' of https://git01.codeplex.com/cntk into linux-gcc
...
Conflicts:
MachineLearning/CNTK/ComputationNetwork.h
MachineLearning/CNTK/NetworkDescriptionLanguage.cpp
MachineLearning/CNTK/SynchronousExecutionEngine.h
2015-07-11 14:58:18 -07:00
94e54dac66
fixing sentence boundary issues for sequence training
2015-07-11 08:26:15 +00:00
02b36c822d
First implementation of NDRMReader.
2015-07-09 23:26:00 +10:00
9be47d5562
Fix potential issue in SequenceReader
2015-07-09 02:12:08 -07:00
2c6c940a8f
Build braek fix
2015-07-08 15:49:16 -07:00
f96ac201a6
Fix non-sequence reader.
2015-07-08 16:52:44 -04:00
631857bb94
Fixing feature and labels names in Kaldi2Reader; also adding some explanation in README of how to compile CNTK with Kaldi
2015-07-08 01:56:50 +00:00
776c20b20a
Merge remote-tracking branch 'origin/master' into linux-gcc
...
Conflicts:
Common/Include/DataReader.h
Documentation/CNTK-TechReport/lyx/CNTKBook_CNTK_Chapter.lyx
MachineLearning/CNTK/TrainingCriterionNodes.h
Math/Math/Matrix.cpp
2015-07-07 17:00:17 -07:00
e37ddb8787
Fix Windows builds.
2015-07-07 16:33:14 -07:00
ba1a969e8e
Fix multi-IO.
2015-07-07 16:11:23 -04:00
204b879dfe
Finish GPU NCE training
2015-07-06 20:56:16 -07:00
471f710f49
Add sentence end information to the reader.
2015-07-06 04:04:26 -04:00
f9245c2acc
Merge remote-tracking branch 'origin/dongyu/dev' into linux-gcc
...
Conflicts:
DataReader/HTKMLFReader/HTKMLFReader.cpp
DataReader/HTKMLFReader_linux/HTKMLFReader.cpp
DataReader/KaldiReader/HTKMLFReader.cpp
MachineLearning/CNTK/ComputationNetwork.h
MachineLearning/CNTK/TrainingCriterionNodes.h
Makefile_kaldi2.gpu
2015-07-05 22:53:04 -07:00
f0ceb4b395
Merge with master branch.
2015-07-05 22:19:18 -07:00
b86990e638
Merge remote-tracking branch 'origin/master' into dongyu/dev
...
Conflicts:
CNTK.sln
Common/Include/Basics.h
DataReader/HTKMLFReader/HTKMLFReader.cpp
DataReader/HTKMLFReader_linux/HTKMLFReader.cpp
DataReader/KaldiReader/HTKMLFReader.cpp
Documentation/CNTK-TechReport/lyx/CNTKBook_CNTK_Adv_Chapter.lyx
MachineLearning/CNTK/ComputationNetwork.h
MachineLearning/CNTK/LinearAlgebraNodes.h
MachineLearning/CNTK/NetworkDescriptionLanguage.cpp
MachineLearning/CNTK/NonlinearityNodes.h
MachineLearning/CNTK/SynchronousExecutionEngine.h
MachineLearning/CNTK/TrainingCriterionNodes.h
Makefile_kaldi2.gpu
2015-07-05 22:16:57 -07:00
c2b8eceb4e
change definition of SENTENCE_BEGIN etc.
...
Add operator |= to MinibatchPackingFlag
Add support of UtteranceEnd in MinibatchPackingFlag to support bi-directional model in later
2015-07-05 16:57:34 -07:00
90be0b8d90
Update NCE GPU Code, forward phase is ready
2015-07-02 05:44:19 -07:00
df1d881b1d
Fix linux build (use popen in linux instead of _wpopen).
2015-07-02 04:21:36 -04:00
04c81b434d
Merge remote-tracking branch 'origin/master' into linux-gcc
...
Conflicts:
CNTK.sln
Common/Include/Basics.h
DataReader/HTKMLFReader/HTKMLFReader.cpp
DataReader/HTKMLFReader_linux/HTKMLFReader.cpp
DataReader/KaldiReader/HTKMLFReader.cpp
MachineLearning/CNTK/ComputationNetwork.h
MachineLearning/CNTK/NetworkDescriptionLanguage.cpp
MachineLearning/CNTK/NonlinearityNodes.h
MachineLearning/CNTK/SynchronousExecutionEngine.h
2015-07-01 23:35:59 -07:00
9db9bb2604
Feed several frame into the minibatch if there is no new utterance. not the best way, just make it consistent with other readers.
2015-07-02 00:46:14 -04:00
3aab1d267f
File now supports pipes ("cmd|", "|cmd") and "-" (stdin/stdout) for input and output;
...
stderr redirection now uses File class instead of auto_file_ptr, allowing e.g. stderr="|tee path.log";
new method File::Flush();
class File implementation now uses RuntimeError() and LogicError() instead of throw std::runtime_error() and logic_error();
moved RuntimeError() and class Plugin from basetypes.h to Basics.h;
new VS++ CRT emulation function _wpopen() (Linux only_
2015-07-01 16:17:29 -07:00
e613a79915
Fix Kaldi2Reader for LSTM -- the previous commit fails after couple of epoch due to a bug
2015-07-01 21:13:22 +00:00
f5ce4c8315
Windows->Linux: finalizing reconciliation of Linux vs. Windows version of HTKMLFReader
2015-07-01 14:02:14 -07:00
5c590f7417
Windows->Linux: 85f7553 fix bug in multi utterance reader when mlf and feature file are mismatched in duration
2015-07-01 14:02:14 -07:00
566c71d796
Windows->Linux: 51daad8 moved nocase_compare to Basics.h, and removed....
2015-07-01 14:02:14 -07:00
fe3bcfd1fb
Windows->Linux: 743c5a9 replaced printf format %ws by %ls which is portable between Visual Studio and gcc
2015-07-01 14:02:14 -07:00
c6e2d7147d
Windows->Linux: db7b574 Further depress the message in HKTMLFReader
2015-07-01 14:02:14 -07:00
61ea159c67
Windows->Linux: bc092c4 Disable redundant information output from releaserandomizedchunk when verbosity =0
2015-07-01 14:02:14 -07:00
19926adf49
Windows->Linux - 826ec09 1) Added the missing CNTKMath project dependency...
2015-07-01 14:02:13 -07:00
7da17823be
Windows->Linux fd314f1 Modify HTKMLFReader/fileutil.cpp to support load in MLF file larger than 2 GB.
2015-07-01 14:02:13 -07:00
120e150617
Windows->Linux: 5444c25 bug fixes in HTKMLFReader: randomizationNone for rollingwindow and context window setting for write
2015-07-01 14:02:13 -07:00
36bae006dc
Windows->Linux: a93a2a1 Eliminate "requiredata:..." message when verbosity == 0
2015-07-01 14:02:13 -07:00
20045cd5ea
Windows->Linux: 9e741ec Further remove message in HTKMLFReader when verbosity == 0
2015-07-01 14:02:13 -07:00
56e992b183
Windows->Linux: c3aad5b Using verbosity in HTKMLFReader to control the output.
2015-07-01 14:02:13 -07:00
c88e4cf20d
Whitespace-only change: undoing some whitespace changes
...
Those changes made simulatenously with porting of HTKMLFReader to linux.
Reverting those changes allows better code comparison and helps to
reconcile windows vs. linux versions of the reader
2015-07-01 14:02:13 -07:00
dc6c7db1c9
Clean HTKMLFReader_linux
2015-06-30 16:21:34 -04:00
a3450a7d36
Merge remote-tracking branch 'origin/dongyu/dev' into linux-gcc
2015-06-30 15:00:58 +00:00
a7c5de7b52
Fix all the reader to use the new interface
2015-06-29 02:16:23 -04:00
27b00ea842
linux-gcc: merging master
2015-06-29 02:17:21 +00:00
72326cb325
use vector of flags to indicate each column's packing state.
...
fixed bugs in the original ResetBounds call in the delay node.
2015-06-27 00:02:21 -07:00
8adff3546b
Make Kaldi2Reader to consistent with the new interface
2015-06-26 00:48:34 -04:00
e980861763
Fix the reader speed issue when we have multiple cntk on one machine
2015-06-25 21:34:41 -04:00
d703a2b718
Merge branch 'Dev' of https://git.codeplex.com/cntk into Dev
...
Conflicts:
MachineLearning/CNTK/RecurrentNodes.h
2015-06-21 01:54:54 -04:00
e2f52516be
Add change to support load multiple readers
2015-06-20 16:57:58 -04:00
f332421b7b
Merge with new changes including RowStackNode
2015-06-16 16:01:19 -07:00
99af4139a5
Merge branch 'Dev' of https://git01.codeplex.com/cntk with master
...
fix compilation bug in HTKMLFReader
Conflicts:
.gitignore
MachineLearning/CNTK/ComputationNetwork.h
MachineLearning/CNTK/ComputationNode.h
MachineLearning/CNTK/LinearAlgebraNodes.h
MachineLearning/CNTK/NetworkDescriptionLanguage.cpp
MachineLearning/CNTK/NonlinearityNodes.h
MachineLearning/CNTK/SynchronousExecutionEngine.h
2015-06-16 11:49:19 -07:00
01468f3fb6
Merge branch 'Dev' of https://git01.codeplex.com/cntk into Dev
2015-06-16 07:28:25 -07:00
16c121df3d
Merge branch 'Dev' of https://git01.codeplex.com/cntk
2015-06-14 23:31:22 -07:00
b0fd604d1d
Fix the compile error on windows for htkmlfreader
2015-06-14 23:31:08 -07:00
31171bfc1b
Merge branch 'Dev' of https://git.codeplex.com/cntk into Dev
...
Conflicts:
MachineLearning/CNTK/RecurrentNodes.h
2015-06-15 02:28:07 -04:00
3994dab33e
Fix some compile errors on linux and also change the matrix multiply to for loop for the delay node.
2015-06-15 02:08:04 -04:00
c44aba7aed
Merge branch 'master' of https://git01.codeplex.com/cntk
2015-06-10 13:45:07 -07:00
346dc11cde
fix bugs in the binaryReader and UCIFastReader
2015-06-10 13:44:54 -07:00
32dde12e37
local change in dev branch before merging it to smt branch.
2015-06-08 22:12:44 -07:00
e4424d5697
Fix the comile on linux for kaldi reader
2015-06-07 18:19:26 -04:00
02080fc0f6
fix bug in multi utterance reader when mlf and feature file are mismatched in duration
2015-06-04 13:12:10 -07:00
4b29673fda
Fix the error throw bugs in UCIParser. Now if a file cannot be opened the error will be thrown and caught and cntk will exit gracefully.
2015-06-03 16:57:02 -07:00
1239acfe4c
Make reader to be a meta reader that can loads multiple readers
2015-06-02 17:45:26 -07:00
d436bf2ac8
Further fixes to sequence training on Kaldi lattices (move data from GPU to CPU before lattice computation)
2015-06-02 16:25:45 -04:00
a42229b8a7
bug fix: frameMode was always false in Kaldi2Reader due to my previous commit
2015-06-02 02:46:11 -04:00
026426ad08
In sequencereader move labelOutput to DeviceId and remove useless code in NCE-LSTM, which will move data from GPU to CPU
2015-06-01 23:29:25 -07:00
12e4543373
fix small issues in sequence training; undo one MPI related change in SGD.h that was accidently checked in yesterday.
2015-06-01 17:50:28 -04:00
99d99b3887
Adding code for smbr/mpfe training on Kaldi lattices; this is the initial working version, more will follow
2015-05-31 19:03:44 -04:00
b87775a353
Minor changes to make CNTK compile with gcc-4.7
2015-05-31 18:19:39 -04:00
ebe575481e
LUSequenceReader : if obs is larger than its dimension, check if no-label is set.
2015-05-29 22:44:35 -07:00
f9a851b485
Replace some window specific function
2015-05-28 20:42:19 -04:00
4b12277e0a
merge with master branch code
2015-05-28 11:15:03 -07:00
d3b2007147
Remove OrganizeClass function from sequence reader.
2015-05-28 10:15:07 -07:00
0eddd57f8d
Use CPU matrix in Sequence reader. Specify unk in reader.
2015-05-27 17:12:59 -07:00
a78f80d5b5
serialization of NCE evalMode for NoiseContrastiveEstimationNode. Seperate training and evaluation nodes for NCE because they require different mode. Report error when run NCE in GPU because that is not supported yet.
2015-05-26 16:55:35 -07:00
088ed7c066
LM Sequence Reader uses the new way to reset RNN state
2015-05-26 13:54:01 -07:00
7cc257b45c
add unnormalized evaluation for NCE
2015-05-26 02:27:23 -07:00
0ba2bc05c0
LUSequenceReader changes: use unk word if a word is not i no observed list. Support read word class information from LU Sequence Reader (mode=class)
2015-05-24 23:14:58 -07:00
5863e2e0f5
Minor format specifier cross-platform fixes
2015-05-22 00:25:44 -07:00
862d1f04cc
small comment change for LUSequenceReader. Add to-do in recurrentnodes.h for delaynode.
2015-05-20 17:15:39 -07:00
b1f7263326
Merge from master branch
2015-05-20 17:11:52 -07:00
7014ed24ba
Reader changes to use new way of reseting RNN states and criterion nodes. Use 0/1 to denote sentence begining and middle. -1 to denote no label. Reader needs to pass two matrices of segmentation informatoin. One is a matrix for saving -1/0/1 information and the other is a one-row matrix saving 0/1 mask. In the latter case ,0 means existence of sentence begining or no labels; 1 means no such case.
2015-05-20 14:02:25 -07:00
848104dcda
add comments for BatchSequenceReader::GetLabelOutput
2015-05-20 02:23:31 -07:00
c8666ba3a8
Fix a couple of Linux build issues
2015-05-19 17:00:51 -07:00
5afa3d73e5
Merge branch 'linux-gcc' of https://git01.codeplex.com/cntk
2015-05-19 16:21:56 -07:00
91a27cd9fc
Merge branch 'master' of https://git.codeplex.com/cntk into linux-gcc
...
Conflicts:
Common/Include/minibatchsourcehelpers.h
DataReader/DSSMReader/DSSMReader.cpp
DataReader/HTKMLFReader/HTKMLFReader.cpp
DataReader/LMSequenceReader/SequenceReader.cpp
DataReader/LMSequenceReader/minibatchsourcehelpers.h
DataReader/LUSequenceReader/LUSequenceReader.cpp
DataReader/LUSequenceReader/minibatchsourcehelpers.h
DataReader/LibSVMBinaryReader/LibSVMBinaryReader.cpp
DataReader/UCIFastReader/UCIFastReader.cpp
MachineLearning/CNTK/ComputationNetwork.h
MachineLearning/CNTK/SGD.h
MachineLearning/CNTK/SimpleNetworkBuilder.h
MachineLearning/CNTK/SynchronousExecutionEngine.h
MachineLearning/cn/PTaskGraphBuilder.cpp
Makefile.cpu
Makefile.gpu
Math/Math/NoGPU.cpp
2015-05-19 16:04:47 -07:00
c31fa8bf28
added ... expansion to HTKMLFReader, allows pathnames in SCP files thata re relative to the SCP file itself (... stands for the directory of the SCP file)
2015-05-19 15:45:26 -07:00
a059635386
moved nocase_compare to Basics.h, and removed unnecessary dependencies caused by this on commandArgUtil.h, hoping for less recompilation;
...
renamed PreprocessConfigLine() to StripComments();
updated CNTKBook PDF
2015-05-19 14:32:21 -07:00
0ff74f0405
The commits 32391e1..2a7ab60 should not be pushed to public (since it enables MKL and MPI) revert them in this commit.
...
Reverted commit:
8404e9c8de64a37301f961c6d59e04
2015-05-19 14:08:16 -07:00
61c6d59e04
add a unit test for CNTKEval
...
This unit test is used privately by erw.
Conflicts:
CNTKSolution/EvalUnitTest/CNTKEValTest.cpp
CNTKSolution/EvalUnitTest/EvalUnitTest.vcxproj
CNTKSolution/EvalUnitTest/EvalUnitTest.vcxproj.filters
2015-05-19 12:31:46 -07:00
89ea3f3448
replaced printf format %ws by %ls which is portable between Visual Studio and gcc
2015-05-19 00:24:08 -07:00
3490c66615
Rename to NO_LABELS and ResetForNoLabels
2015-05-18 23:18:49 -07:00
5445952422
(removed a defunct reference from the UCIFastReader VS Project)
2015-05-18 20:37:41 -07:00
178555d0a4
Visual Studio Solution cleanup:
...
moved minibatchsourcehelpers.h to Common/Include, as it is shared amongst readers (exception: the HTK reader, which has a different version);
fixed some incorrect #include of basetypes.h;
copying of DLLs now does not copy if the file already exists and is not out of date;
prebuild.bat now only updates builddate.h if it has changed, to avoid CNTK.cpp to be recompiled each time;
fixed a few build warnings in the CUDA kernels (uninitialized variables);
removed unused PTasks references in Linux makefiles
2015-05-18 20:32:39 -07:00
08625d482d
created a new header file Basics.h, to take on the whole of the legacy basetypes.h (which was to contain a very limited set of generally useful core functions, like UTF8 conversion)
2015-05-18 19:36:04 -07:00
8856fa40e9
Merge branch 'master' of https://git.codeplex.com/cntk
2015-05-18 18:51:31 -07:00
392d64da08
various cleanups:
...
renamed cn.exe to CNTK.exe on Windows and "cntk" (no more .exe!) on Linux;
renamed 'cn' folder to 'CNTK';
moved CNTK.sln out from CNTKSolution folder, which now no longer needs to exist;
removed Win32 build target from Visual Studio Solution and all Projects;
added are documentation and Linux makefiles to the Visual Studio Solution for easy access;
removed unused PTasks files
2015-05-18 18:51:20 -07:00
3dbed2fc89
use default seed for sampler
2015-05-18 05:18:00 -07:00
dad030381e
modified: SequenceReader.cpp
...
modified: SequenceReader.h
2015-05-18 00:46:27 -07:00
0462453a1b
modified: SequenceReader.h
2015-05-16 22:07:38 +08:00
d4e025e35f
Repeat the first couple of labels when set delay targets
2015-05-08 08:34:52 -04:00
204419ba80
Add delay output option to the Kaldi2Reader
2015-05-07 21:03:29 -04:00
edd8c08040
LUSequenceREader use number of samples instead of number of sentences so that it is consistent to what is expected in SGD.h.
2015-05-06 11:04:07 -07:00
9beea90918
deal with no-blank lines at the end of file in LU Sequence Reader
2015-05-06 00:05:20 -07:00
a93ada346b
Fix LU sequence reader bug to get the id of end of sentence symbol
2015-05-05 13:50:38 -07:00
be0f8aaf24
Removed unused code in LUSequenceReader. update getframe function to use sparse input.
2015-05-01 11:55:51 -07:00
0b72e35836
Further depress the message in HKTMLFReader
2015-04-30 14:14:51 -07:00
ef0113ce0d
LUSequenceReader : Use local CPUsparsematrix to read data and then pass data to GPU/CPU sparsematrix as features
2015-04-30 12:02:42 -07:00
04b0b0e1a1
Chanage LU Sequence reader to work on sparse input
2015-04-29 09:31:31 -07:00
3a89639d40
Add nogpu functions for RCRF related functions
2015-04-27 22:22:17 -07:00
93c895ed0e
Merge with master branch
2015-04-22 23:20:34 -07:00
cd42428e07
Fix bug in LUSequence Reader, otherwise read from begining even for single pass of data.
2015-04-21 22:37:06 -07:00
0f74cd8dd2
reset to doc begining when read to end.
2015-04-21 18:09:16 -07:00
fad6322a29
Simplified LU sequence by removing unnecessary code. Also use wstring to support reading wchar strings. Need to make sure that large amount of data can be read, because reader reads at most CACHE_BLOG_SIZE data.
2015-04-20 16:04:39 -07:00
94327e2a18
Fix compile error for old gcc version
2015-04-15 23:06:50 -04:00
ff80ec1b7b
Fix some compile error for old gcc
2015-04-15 23:05:04 -04:00
0ece27ccc5
LUSequenceWriter outputs results to stderr
2015-04-13 15:48:33 -07:00
f694eda4c1
Set label uses CPU Device
2015-04-10 15:43:55 -07:00
43cb5e0e25
Add InitProposals, GetSentencendIdFromOutputLabel and GetProposals in dataread. Update DataEnd in LUSequenceReader so that every minibatch works as sentence-level minibatch.
2015-04-10 14:19:14 -07:00
b6610615a9
Add ignoresentencebegin tag to treat sentence begining frames as in-the-middle-of-sentence frame. Useful for telling BPTT to use state from external, such as states from another network, e.g., encoder network.
2015-04-09 14:49:55 -07:00
9b0b06ec6f
Merge remote-tracking branch 'origin/master' into linux-gcc
2015-04-09 14:35:50 -04:00
9ed33db20b
fixed several warning that happens under gcc.
2015-04-09 02:10:38 -07:00
319ea82d56
Clean the warnings.
2015-04-09 01:09:58 -04:00
14ce7f4f2f
Use -1,0,1 to denote no observation, sentence begining and in-the-middle-of-sentence. changed LU sequence reader. But other readers haven't changed accordingly.
2015-04-08 21:20:12 -07:00
0e5a54bc44
Fix compile error
2015-04-08 13:25:35 -04:00
5ef4bcd790
Merge remote-tracking branch 'origin/master' into linux-gcc
...
Conflicts:
MachineLearning/cn/CompositeComputationNode.h
Math/Math/CPUMatrix.cpp
2015-04-07 00:35:17 -04:00
4683c541b6
Disable redundant information output from releaserandomizedchunk when verbosity =0
2015-04-05 21:41:33 -07:00
66a8ea9e3d
Merge remote-tracking branch 'origin/master' into linux-gcc
2015-04-05 16:31:26 -04:00
ef4c833a4d
Check in the KaldiReader (version2)
2015-04-05 16:28:16 -04:00
13c49fc3ad
LUSequenceReader changes: using multiple io stream.
2015-04-03 14:03:17 -07:00
f4d63f7352
change the ReadClassInfo function in the LM sequence reader so that it also works when the training set is large.
2015-04-01 20:44:02 -07:00
8e5e24fbd6
fix compile errors in LibSVMBinaryReader
2015-03-30 23:12:58 -07:00
6a1ce07561
1) Added the missing CNTKMath project dependency to project LMSequenceReader to enforce correct ordering of the build. 2) A bunch of changesto fix the Windows build. 3) Changed Linux makefiles to use CUDA 7.0
2015-03-27 10:36:32 -07:00
e1c964cfa1
Merge remote-tracking branch 'origin/master' into linux-gcc
...
Conflicts:
DataReader/LMSequenceReader/SequenceParser.h
DataReader/LMSequenceReader/SequenceReader.cpp
MachineLearning/cn/ComputationNetwork.h
Math/Math/CPUMatrix.cpp
Math/Math/CPUSparseMatrix.cpp
Math/Math/GPUMatrix.cu
2015-03-25 22:15:58 -04:00
df6087ba76
Merge remote-tracking branch 'origin/master' into linux-gcc
...
Conflicts:
DataReader/SequenceReader/SequenceReader.cpp
MachineLearning/cn/ComputationNode.h
MachineLearning/cn/cn.cpp
Math/Math/CPUSparseMatrix.cpp
Math/Math/GPUMatrix.cu
Math/Math/GPUSparseMatrix.cu
Math/Math/GPUSparseMatrix.h
2015-03-25 21:49:59 -04:00
77416059e2
fixed bugs when MKL is turned on.
...
change sequencereader folder to LMSequenceReader.
2015-03-24 22:58:11 -07:00
34b46e6499
Add KaldiReader into the source tree. Actually, it is a HTKMLFReader with kaldi-io.
2015-03-20 16:24:46 -04:00
d35570ba6d
Fixed the error for the LM modeling in linux side. Bug: undefined behavior when use index[a] = index.size()
2015-03-20 15:57:31 -04:00
a99c4636fd
Port to CUDA 7.0
...
Changed the writeWordAndClass command to produce the vocabulary, frequency, and word to class mapping information directly from the training set.
2015-03-19 01:04:49 -07:00
bfb0de4764
display file, line, and function name info when a feature is not implemented.
...
display warning on dropout and temp memory for CNN only when non-default value is set.
change the name of Sequencereader project to LMSequenceReader
Change the corresponding config file that use LMSequenceReader.
2015-03-11 23:05:23 -07:00
8beb5249c9
change TrainRNN to Train in all the config files since TrainRNN and Train now do the same thing.
...
Rewrite the RNNLM.NDL for class based Rnn LM.
removed dead code in SequenceReader.cpp
Change SparseLearnableParameters in the simple network builder to Learnable Parameters.
change the argument name and comment in the ClassBasedCrossEntropyWithSoftmaxNode to make it clearer the meaning of each input argument.
2015-03-09 00:28:55 -07:00
cba7c67101
Fix bug in HCIFastReader (segmentation fault because labelIdType is uninit)
2015-03-01 20:59:07 -05:00
c4a1ee1031
Merge remote-tracking branch 'origin/master' into linux-gcc
...
Conflicts:
DataReader/SequenceReader/SequenceReader.cpp
MachineLearning/cn/ComputationNetwork.h
MachineLearning/cn/NetworkDescriptionLanguage.cpp
MachineLearning/cn/SGD.h
MachineLearning/cn/cn.cpp
Math/Math/CPUMatrix.cpp
Math/Math/CPUSparseMatrix.cpp
Math/Math/CPUSparseMatrix.h
Math/Math/GPUMatrix.cu
Math/Math/GPUSparseMatrix.cu
Math/Math/GPUSparseMatrix.h
2015-03-01 20:01:26 -05:00
e33afa7b79
Add Kaldi header to cntk
2015-03-01 14:56:07 -05:00
89a9e8935f
Fix a bug in HTKMLFWriter for Linux.
2015-02-28 22:52:16 -05:00
8ebdd580fa
add ranomization in lu sequence reader
2015-02-24 21:30:30 -08:00
c9f7164ef6
Set sentence begining by looking at the input sequence instead of output sequence. Ignore Sparse matrix in gradient check (to-do to fix this)
2015-02-24 16:22:41 -08:00
5a2ae4fd8a
Full commit of the LibSVMBinaryReader. Somehow the cpp/h files were not added before.
2015-02-17 16:05:47 -08:00
7fe0af4446
Merge branch 'thhoens' of http://vstfpg07:8080/tfs/msrspeech/msrspeech/_git/cntk
2015-02-16 19:05:49 -08:00
5ecc77cb06
Merge branch 'master' of https://git01.codeplex.com/cntk into DSSM-Internal
...
Conflicts:
CNTKSolution/CNTKSolution.sln
MachineLearning/cn/ComputationNetwork.h
MachineLearning/cn/TrainingCriterionNode.h
Math/Math/CPUMatrix.cpp
Math/Math/CPUMatrix.h
Math/Math/GPUMatrixCUDAKernels.cu
Math/Math/Matrix.h
2015-02-16 13:35:59 -08:00
22d7769b04
make Resize and SwitchToMatrixType to use explicit specification of arguments to avoid unexpected behavior.
...
Completely remove PTask. This makes code clean and much easier to understand.
Added SetMatrixFromCSCFormat to CPU sparse matrix.
Reimplemented Set(Get)MatrixFromCSC(CSR)Format functions to take CPUSPARSE_INDEX_TYPE. changed GPU sparse matrices' persistence code accordingly.
2015-02-16 13:18:57 -08:00
5651eeb210
Update sequence reader to use CPUDEVICE for label always.
2015-02-15 22:25:44 -08:00
775c0e501a
Update sequence reader to assign class information to output label.
2015-02-14 22:36:21 -08:00
41a1338092
Update ClassBasedCrossEntropyNode. Modified sequence reader to use dense matrix to represent labels at each time. Cleaned SGD.h to remove SetOtherInfo function that calls class-based LM setup functions, which should be done in reader. Updated simple network builder for class-based LM. Added MinusOneAt function for Matrix. Updated CNTK book chapter.
2015-02-14 21:38:23 -08:00
25674402db
Merge branch 'thhoens' of http://vstfpg07:8080/tfs/msrspeech/msrspeech/_git/cntk
2015-02-13 12:22:46 -08:00
f2004053a3
Output missing label information when the label is not found.
2015-02-13 11:17:16 -08:00
da76314061
Merge branch 'thhoens' of http://vstfpg07:8080/tfs/msrspeech/msrspeech/_git/cntk into DSSM-Internal
2015-02-12 18:02:30 -08:00
b67ca20a9f
Merge branch 'master' of http://vstfpg07:8080/tfs/msrspeech/msrspeech/_git/cntk into DSSM-Internal
...
Conflicts:
Common/ConfigFile.cpp
Common/File.cpp
Common/fileutil.cpp
MachineLearning/cn/cn.cpp
Math/Math/GPUMatrix.cu
Math/Math/GPUMatrixCUDAKernels.cu
Math/Math/GPUSparseMatrix.cu
2015-02-11 17:35:08 -08:00
c1e94a0ce0
Update LUSequence Parser to suport blank lines as sentence ending
2015-02-11 10:10:15 -08:00
Amit Agarwal
UnderdogGeek
chenguoguo
Scott Cyphers
Hakan Erdogan
Dong Yu
Yu
Yinggong ZHAO
ksyao
Mike Seltzer
Vladimir Ivanov
Yinggong ZHAO
Yu Zhang
Frank Seide
erw
Chiyuan Zhang
thhoens