https://github.com/Microsoft/CNTK

sort by:
Revision Author Date Message Commit Date
eb89c00 License change 18 January 2016, 08:32:22 UTC
57f50f4 Update solution, update the baseline for Speech/DNN/DiscriminativePreTraining, update control data for reader tests 16 December 2015, 14:18:27 UTC
4954e4f Update baselines for Speech/DNN/SequenceTraining 16 December 2015, 10:46:56 UTC
cb46fa8 Update baselines 16 December 2015, 08:51:10 UTC
5a99309 Fix random shuffle 15 December 2015, 16:28:17 UTC
e178277 Merge branch 'master' into alrezni/RandomUniform 15 December 2015, 09:48:50 UTC
bed1eea Address uniform random inconsistencies * Use mt19937 instead of ranlux64_base_01. Replace std random with boost random. * Fix floating point issues in _rescaleToRange. Flip range to [min, max). Add CUDA intrinsics and a unit test for doubles. * Use mt19937_64 in readers instead of std::rand * Update baselines 15 December 2015, 09:40:23 UTC
d273078 merged with master 15 December 2015, 00:13:51 UTC
dc32e4f (comments) 15 December 2015, 00:10:07 UTC
7a57f8b (added logging to calgammaformb() to track down an error) 14 December 2015, 23:40:50 UTC
f4a9155 bug fix, MBLayout::IsEnd() 14 December 2015, 22:48:26 UTC
af41546 Merge branch 'master' of https://git.codeplex.com/cntk Conflicts: MachineLearning/CNTKSGDLib/DataReaderHelpers.h 14 December 2015, 21:12:36 UTC
e9fb669 Merge branch 'master' of https://git.codeplex.com/cntk into fseide/mblayout 14 December 2015, 21:11:51 UTC
abbc28e Coup de grâce: MBLayout reimplemented: the old bitmap-based implementation is gone. MBLayout now also stores sequence ids (so we can line up multiple MBLayouts, instrumental to sequence-to-sequence models), and it naturally allows for testing whether a delayed value cross a sequence boundary; deleted definition of MinibatchPackingFlags enum; old Is() and Get() functions are fully replaced by IsGap() and IsBeyondStartOrEnd(); HasGaps() test is now a one-liner 14 December 2015, 21:11:29 UTC
5ba34cf also removed generic MBLayout::Get(). Now tests are always for Gap or sequence boundary; bug fix: ImageReader must now initialize MBLayout with InitAsFrameMode() 14 December 2015, 19:43:48 UTC
cf931e2 removed MBLayout::Is(...t...) and Get(..t...). Now one must use Is(FrameRange), or some places can be more efficient now by avoiding to use this at all; new methods MBLayout::HasSequenceBeyondBegin() and likewise for End, for testing whether state needs to be preserved across minibatches 14 December 2015, 19:35:49 UTC
fe3076f Fixed ImageReader and updated sample. 14 December 2015, 19:33:26 UTC
25c5a12 deleted MBLayout::Set() and the old flag matrices--yay; deleted DelayedValueNodeBase::m_shiftedMBLayout (in lieu of the new MBLayout implementation) 14 December 2015, 17:01:57 UTC
0d808a5 Merge branch 'master' of https://git.codeplex.com/cntk 14 December 2015, 16:04:20 UTC
b8eb51d Fix kaldi Reader. 1) Using the new MBLayout interface. 2) Modify the configParameter to configRecordType to consistent with HTKMLFReader. 3) Clean the warning message. Next: refactor to make it consistent with HTKMLFReader. 14 December 2015, 16:00:20 UTC
8559820 Fixing boost version 14 December 2015, 15:05:29 UTC
e573dd7 Fixing greater and less than symobls 14 December 2015, 14:07:34 UTC
7c251eb Fixing special characters 14 December 2015, 13:56:49 UTC
0bfa4c8 Changing notification message for boost installation 14 December 2015, 12:07:56 UTC
642947a Merge branch 'master' of https://git.codeplex.com/cntk into fseide/mblayout 13 December 2015, 21:11:11 UTC
e066467 ColumnRangeWithMBLayoutFor(), ValueFor() etc. now understand FrameRange::m_timeOffset. DelayedValueNode has been changed to use this mechanism 13 December 2015, 21:07:52 UTC
88fe6b2 DelayedValueNode no longer uses m_pShiftedMBLayout (still kept around for use inside asserts to test the new code) 13 December 2015, 20:19:27 UTC
e6f8108 CUDA bringing changes in elementwise functions to NoGPU.cpp 13 December 2015, 19:46:02 UTC
d0e587d CUDA fix bug I accidentallty introduced in _scaleAndAddScalar kernel 13 December 2015, 19:45:59 UTC
5a74551 CUDA moved common patterns into macros for elementwise functions, changed op* names to match. 13 December 2015, 19:45:58 UTC
259d00f Reverting to old calculation of _assignSigmoidOf() 13 December 2015, 19:45:56 UTC
3ab2491 CUDA renamed macro to be more descriptive. 13 December 2015, 19:45:55 UTC
85e7d78 CUDA kernels were missing some const declarations, GPUSparseMatrix kernels now access m_pArray indirectly through NzValues() helper function. 13 December 2015, 19:45:54 UTC
990f1fe CUDA Matrix ScaleAndAdd additionally uses a third parameter for the output 13 December 2015, 19:45:52 UTC
1c6f9d0 GPU*Matrix code cleanup. 13 December 2015, 19:45:51 UTC
bd9fa22 CUDA kernels changed from _inplace*() to _elementWise() in anticipation of speeding up the GPUMatrix::Assign*() implementations. 13 December 2015, 19:45:49 UTC
a45365c made gcc happy (variable set but not used) 13 December 2015, 09:04:27 UTC
1fc1ad0 made gcc happy (two missing exports for Matrix<char> template) 13 December 2015, 08:39:07 UTC
785bc15 DelayedValueNode::ForwardProp() now works without the shifted layout (it is still there for comparing the new method); MBLayout::Get() methos are now all private; MBLayout::Is() method split into per-flag methods. Soon, Is() itself will be private, then removed 13 December 2015, 08:34:24 UTC
489dcb0 old MBLayout flags no longer used except for testing the new versions (to be removed soon) 13 December 2015, 07:15:15 UTC
059d21b MBLayout::GetColumnsValidityMask() now returns a Matrix & instead of a shared_ptr, and is now sliced like the data matrix with DataWithMBLayoutFor(). it also now assumes there are gaps. Caller (MaskMissingColumnsTo()) must check. More regular and shorter code 13 December 2015, 06:53:26 UTC
a2eb692 removed left-overs of lazy-alloc logic in MBLayout 13 December 2015, 04:24:28 UTC
eff836b MBLayout no longer lazily allocating 13 December 2015, 04:09:15 UTC
1ef00c9 Temp convo tests disabled. 12 December 2015, 18:02:05 UTC
8c9080b Fixed Win build. 12 December 2015, 18:01:55 UTC
76d3da3 cudnn: added check for CC version. 12 December 2015, 18:01:45 UTC
217c9f9 Enabled defatul OpenCV location search on Linux. 12 December 2015, 18:01:34 UTC
7c01500 cudnn: fixed CPUONLY. 12 December 2015, 18:01:24 UTC
a0a8e27 Fix for gcc. 12 December 2015, 18:01:14 UTC
057512d Added buffer size heuristic to UCIFastReader to avoid segfaults on long lines. 12 December 2015, 18:01:03 UTC
ddf6db1 cudnn: addressed review comments. 12 December 2015, 18:00:53 UTC
4e96192 Fixed Linux build issues in ImageReader. 12 December 2015, 18:00:42 UTC
1ad4e29 cudnn: resolved merge issues. 12 December 2015, 18:00:32 UTC
107867f cudnn: refactored ImageReader prefetcher to use OpenMP. 12 December 2015, 18:00:22 UTC
d93be91 cudnn: added prefetching to ImageReader. 12 December 2015, 18:00:11 UTC
254b04c cudnn: added OpenCV libs to Linux build. 12 December 2015, 18:00:01 UTC
ea72ada cudnn: fixed workspace bug, minor refactoring. 12 December 2015, 17:59:51 UTC
fef4b5c cudnn: moved tests to Boost framework, updated samples. 12 December 2015, 17:59:40 UTC
6fe07ed cudnn: fixed integration issues. 12 December 2015, 17:59:30 UTC
334d645 cudnn: added bias forward/backprop implementation for default engine. 12 December 2015, 17:59:19 UTC
3023f11 cudnn: enabled build on Linux with cuDNN. 12 December 2015, 17:59:09 UTC
895c10a cudnn: merge with master, fix Linux compile errors. 12 December 2015, 17:58:58 UTC
6c7c617 cudnn: implemented batch norm backprop, updated samples, added VGG_E net. 12 December 2015, 17:58:48 UTC
39a8a8c cudnn: added batch norm forward implementation. 12 December 2015, 17:58:38 UTC
8430f88 cudnn: added BatchNormalizationNode. 12 December 2015, 17:58:27 UTC
5698333 cudnn: bug fixes, samples update. 12 December 2015, 17:58:17 UTC
80ebfb4 cudnn: added bias forward/backward. 12 December 2015, 17:58:06 UTC
7300306 cudnn: refactored to use NCHW format. 12 December 2015, 17:57:56 UTC
afae131 cudnn: completed pooling nodes implementation, fixed bugs, added unit tests. 12 December 2015, 17:57:46 UTC
f1eb5d8 cudnn: added pooling engine, unit tests and refactoring. 12 December 2015, 17:57:35 UTC
c2736bf cudnn: added padding support, clean up and refactoring. 12 December 2015, 17:57:25 UTC
91e3378 cudnn: added auto-tuning. 12 December 2015, 17:57:14 UTC
5db082e cudnn: added backprop data/filter implementation, unit tests. 12 December 2015, 17:57:04 UTC
da822fd cudnn: added filter format conversion and backprop. 12 December 2015, 17:56:54 UTC
3f7a3ac cudnn: minor changes to tensor/filter formats. 12 December 2015, 17:56:43 UTC
57eea8c cudnn: add forward implementation, unit tests. 12 December 2015, 17:56:33 UTC
1e9b061 cudnn: add filter descriptor, refactor 12 December 2015, 17:56:23 UTC
106dd17 cudnn: Fixed linker issue. 12 December 2015, 17:56:12 UTC
f241eb2 Add missing files. 12 December 2015, 17:56:02 UTC
aee952e Add cuDNN to VS project. 12 December 2015, 17:55:52 UTC
f703cc2 cuDNN: introduce tensor, convolition options and some ConvNode refactoring. 12 December 2015, 17:55:41 UTC
ede0ff0 MBLayout::Get(t) also folded into Get(FrameRange). Only operator== and a legacy specialized operation stop us from removing the old flags 12 December 2015, 09:00:23 UTC
4a92bd7 Merge branch 'master' of https://git.codeplex.com/cntk into fseide/mblayout 12 December 2015, 08:16:12 UTC
37a827f made Release build happy 12 December 2015, 08:15:47 UTC
f55e6e4 adding an MPI init test in case of that MPI was initialized repeatedly 12 December 2015, 07:33:44 UTC
8787174 Merge branch 'master' of https://git01.codeplex.com/cntk into amitaga/outputValuesMemShare 12 December 2015, 02:11:53 UTC
439b8f1 Removed an unneeded change accidentally added 12 December 2015, 02:06:01 UTC
270077e MBLayout::Get(s,t) now implemented by calling Get(FrameRange) to allow for time offsets 12 December 2015, 00:53:14 UTC
9b41d1b implemented MBLayout::Get() to use the new structure and validate against the old 12 December 2015, 00:38:32 UTC
a8c2e1a Fixed linux build issues 11 December 2015, 22:17:14 UTC
8f0c8c7 updated all frame-mode readers to initialize MBLayout following the new AddSequence() style 11 December 2015, 22:11:28 UTC
60c32fb Merge branch 'master' of https://git01.codeplex.com/cntk into amitaga/outputValuesMemShare 11 December 2015, 22:04:28 UTC
ff6444b Merge branch 'fseide/getmbfix' of https://git.codeplex.com/cntk into fseide/mblayout 11 December 2015, 21:47:47 UTC
f785ff4 Merge branch 'master' of https://git.codeplex.com/cntk into fseide/mblayout 11 December 2015, 21:43:48 UTC
f729a8e MBLayout now validates that AddSequence() was called for all samples, and also keeps track of a gap count to accelerate HasGaps(); new method MBLayout::InitAsFrameMode() for easy updating of frame-mode readers 11 December 2015, 21:41:34 UTC
f366d2e Merge branch 'master' of https://git01.codeplex.com/cntk into amitaga/fixGPUDeviceSelection 11 December 2015, 21:11:53 UTC
3f6d50d Merge branch 'master' of https://git.codeplex.com/cntk into fseide/getmbfix 11 December 2015, 20:54:46 UTC
7a038d9 Merge branch 'master' of https://git.codeplex.com/cntk into fseide/mblayout 11 December 2015, 20:53:54 UTC
fc348af made MBLayout::IsAllNode() private 11 December 2015, 20:53:41 UTC
c995994 Fixed a bug in device selection enforcement. The enforcement function was file static instead of global due to which each source file was getting its own copy of the function and the static variable inside it. 11 December 2015, 19:48:04 UTC
back to top