40d0b79451
Add trace for nodes with abs sum to make compare easier, with trace level 1000
2017-01-05 13:09:12 -08:00
cfd6083a01
Fix autoPad issue using GetAutoPad().
2017-01-04 16:26:33 -08:00
02291c55b4
Bug fix when lowerPad and upperPad has lower ranks than kernelPad.
2017-01-04 16:26:33 -08:00
a0ad56ba75
Revert automatic format change by VS2013.
2017-01-04 16:26:33 -08:00
1ed0f08dce
Bug fix for asymmetric padding for convolution. Added a test too.
2017-01-04 16:26:33 -08:00
ca6b17c377
Revert automatic format change by VS2013.
2017-01-04 16:26:33 -08:00
439fd4eb49
Bug fix for asymmetric padding for convolution. Added a test too.
2017-01-04 16:26:33 -08:00
b9e06b77c0
Incorporated CR feedback
2016-12-29 18:29:58 -08:00
60d4e09ae4
bug fix which affects x.assignSumOf(x,y)
2016-12-29 15:22:49 -08:00
48acae2e48
Remove redundant code.
2016-12-22 14:03:55 -08:00
3bbffd9ae6
Bug fixes for convolution.
2016-12-22 14:03:55 -08:00
e2fd7255af
Align randomization across platforms
2016-12-21 17:31:22 +01:00
124c44b5f3
CNTK V2 library: Added Value::Create method to batch sparse sequences
2016-12-16 01:27:22 -08:00
e26cde816f
Integrate alrezni/8byte_seed into master
2016-12-13 15:05:07 -08:00
eabd0e37d8
Fixing omp
2016-12-13 16:48:28 +01:00
6cde1e3b7d
Use 8 bytes for RNG(Handle) seed on both platforms.
...
Change seed type from unsigned long (4 bytes on windows, 8 bytes on
linux) to uint64_t. Bump cnkt model version to 17 (v1). Change a few
int64_t to uint64_t along the way. Close #1124 .
2016-12-13 12:37:18 +01:00
4fe22b81c5
CNTK core: Implemented an optimization to elide the initial zeroing and subsequent accumulation into gradients for nodes with just one parent/accestor node
2016-12-12 12:06:17 -08:00
a0377462c3
Remove statics from prefetch transferer
2016-12-07 11:57:48 +01:00
c12979b245
Fix sparse matrix MultiplyAndAdd bug that old value may not be kept
2016-12-06 23:49:01 -08:00
20dada981a
cuda related vs2015 changes
2016-11-29 11:57:01 +01:00
ce0d0c7d46
Properly checkpoint stateful nodes
...
Properly checkpoint stateful nodes (Dropout, RandomSample) in both v1/v2
formats:
To ensure reproducibility, nodes that inherit from RngUser need to
properly save and restore their internal stated (rng seed and offset).
V2 additionally needs to copy state from the checkpoint into the function
graph and the underlying network if it already exists.
2016-11-28 15:06:48 +01:00
faef2b540a
Quantized multiplier v1
2016-11-27 21:25:26 -08:00
b47974642b
Fix gpu sparse tests
2016-11-27 00:28:55 +01:00
f77f0fb587
Merge branch 'master' into fmegen/vs2015/general-changes
2016-11-25 18:29:35 +01:00
a5de27940f
Support for GPU/sparse
2016-11-25 15:26:09 +01:00
1823c0fe24
adressing more comments from CR
2016-11-25 14:26:52 +01:00
dac6b6d738
adressing comments from CR iteration 4
2016-11-25 12:06:38 +01:00
b5be3edc7d
getting rid of "variable hides ... " warnings.
...
# Conflicts:
# Source/Common/Include/Platform.h
# Source/Common/Include/ScriptableObjects.h
# Source/ComputationNetworkLib/ComputationNetworkAnalysis.cpp
# Source/ComputationNetworkLib/RNNNodes.cpp
# Source/Readers/HTKMLFReader/msra_mgram.h
# Source/Readers/HTKMLFReader/utterancesourcemulti.h
# Source/Readers/LMSequenceReader/SequenceParser.cpp
# Source/SGDLib/SGD.cpp
2016-11-25 01:04:50 +01:00
e686d1f79e
addressing CR comments.
...
# Conflicts:
# Source/CNTK/BrainScript/BrainScriptEvaluator.cpp
# Source/Common/ExceptionWithCallStack.cpp
# Source/Common/Include/Config.h
# Source/Common/Include/Platform.h
# Source/Common/Include/latticearchive.h
# Source/Common/Include/numahelpers.h
# Source/Math/CPUMatrix.cpp
# Source/Math/Matrix.cpp
# Source/Math/QuantizedMatrix.cpp
2016-11-23 23:12:32 +01:00
7f841125b3
fix more warning descriptions
2016-11-23 12:01:31 +01:00
9e2c1fa4a9
added reasons for various #pragma warning disable.
2016-11-23 11:35:37 +01:00
cae429ff23
make parameter explicit by adding namespace
2016-11-23 10:39:35 +01:00
1e1d437143
fix typo again (curLocation2)
2016-11-23 10:38:12 +01:00
8865dedddf
fix typo (variable name misspelled)
2016-11-23 10:35:25 +01:00
88208f1cfc
disable warnings for new issues.
2016-11-23 10:32:35 +01:00
3c2e112045
add /bigobj to enable compile.
2016-11-22 14:06:49 -08:00
3b2df9d370
Add assert based on code review.
2016-11-22 11:24:37 -08:00
6804156bc2
Max pooling and convolution boundary fix.
2016-11-22 10:39:51 -08:00
a06c88cf6a
added TempDeviceBuffer to CommonMatrix.h
2016-11-19 21:53:24 +01:00
913dcd31d4
fixes for dense times sparse
2016-11-19 21:53:24 +01:00
b555be9303
Incorporated CR feedback
2016-11-17 17:14:34 -08:00
4e88307aba
added a bunch of ops to load legacy models
2016-11-16 16:34:27 -08:00
88065644f7
Merge remote-tracking branch 'origin/master' into mahilleb/pr/874
2016-11-09 14:01:33 +00:00
b8d668ecf9
Source/Math/NcclComm.cpp: tune logging
2016-11-09 14:00:53 +00:00
814381b9a8
1. Fixed some known merging problems
...
2. Fixed a bug of EndForwardProp in BatchNormalization
2016-11-09 17:26:00 +08:00
4670158510
Added NCCL support to SimpleDistGradAggregator
2016-11-08 18:15:01 -08:00
01be41a2b3
Merge branch 'master' into v-yuxgu/memory-optv2
2016-11-08 16:14:12 +08:00
f5af601244
Supple comments and fix several bugs of memory optimization
2016-11-08 15:41:08 +08:00
9c02129e2b
First version of all reduce aggregator
2016-10-24 19:25:36 +02:00
e6036b2735
CNTK v2 library: Fix a bug in convolution inference and other misc bug fixes
2016-10-20 15:37:15 -07:00
KeDengMS
Cha Zhang
Nikos Karampatziakis
Mark Hillebrand
Amit Agarwal
Project Philly
Eldar Akchurin
Alexey Reznichenko
Friedel van Megen
Vadim Mazalov
Willi Richert
fmegen@microsoft.com
Thilo Will
yuxiao.guo
Nathan Luehr