59d382f | Mark Hillebrand | 18 January 2016, 08:31:47 UTC | License change | 18 January 2016, 08:31:47 UTC |
69ccd3f | Amit | 24 October 2015, 21:31:34 UTC | Remove debug info generation option for CUDA compilation in debug flavor builds and also enable fast-math optimizations . These changes have been done to eliminate differences in GPU results for the E2E tests between debug and release flavors. Setting environment variable CNTK_CUDA_DEVICE_DEBUGINFO=1 will enable debug info generation. The baselines for all E2E tests have also been updated in accordance with this change | 24 October 2015, 21:34:01 UTC |
204d4be | Amit Agarwal | 24 October 2015, 04:36:05 UTC | Added newline to a meesage absence of which was causing test comparisons against baselines to fail | 24 October 2015, 04:36:05 UTC |
978dc4f | Alexey Kamenev | 22 October 2015, 05:37:29 UTC | Fix for make_unique in GCC. | 22 October 2015, 05:37:29 UTC |
4ba13c2 | Amit Agarwal | 22 October 2015, 02:01:56 UTC | Merge branch 'master' of https://git01.codeplex.com/cntk into amitaga/removeGPUMatrixDefaultDeviceId | 22 October 2015, 02:01:56 UTC |
abc12c4 | erw | 21 October 2015, 23:36:35 UTC | Improving on DumpNode command. Adding an nodeNameRegex option so that user can specify a gropu of nodes (determined by the regex) that he/she wants to dump. | 22 October 2015, 01:36:48 UTC |
8d68e33 | Amit Agarwal | 22 October 2015, 01:08:31 UTC | Bug fix: Fixed an edge case in gradient aggregation during parallel training, when gradient matrices are unsized due to a rank not having processed any samples | 22 October 2015, 01:08:31 UTC |
6a635f1 | Amit Agarwal | 22 October 2015, 00:42:15 UTC | Fixed a reader bug in the handling of no data being available from the source | 22 October 2015, 00:42:15 UTC |
e72b4fd | Amit Agarwal | 21 October 2015, 22:18:11 UTC | Bug fix: The guard for filling lattice was checking for presence of lattice on an updated mbiter instead of the previous one from which the utterance was actually filled | 21 October 2015, 22:18:11 UTC |
7979372 | Amit Agarwal | 21 October 2015, 19:11:54 UTC | Optimized copy performance between CNTK matrices and SSE matrices in sequence training gamma calculation code - improves DNN sequence training performance further by 10% | 21 October 2015, 19:11:54 UTC |
d85c26e | Amit Agarwal | 21 October 2015, 18:03:56 UTC | Bug fix: The sequence training code for 1 parallel utterance case was not populating the logLLs on the CPU whis is required in the forwardbackward computation | 21 October 2015, 18:03:56 UTC |
68c1fe5 | Amit Agarwal | 21 October 2015, 07:35:12 UTC | Another performance fix for DNN sequence training | 21 October 2015, 07:35:12 UTC |
507922b | Amit Agarwal | 21 October 2015, 01:41:07 UTC | Performance improvement for DNN sequence training | 21 October 2015, 01:41:07 UTC |
f9418c3 | Amit Agarwal | 20 October 2015, 22:38:22 UTC | Added a MPI barrier after the Training epoch loop to make sure model writing is finished before proceeding to the next command | 20 October 2015, 22:38:22 UTC |
265f9d1 | Chris Basoglu | 20 October 2015, 21:42:10 UTC | Fix minibatch progress percenatge | 20 October 2015, 21:42:24 UTC |
7b4b257 | Amit Agarwal | 19 October 2015, 22:15:02 UTC | Moved NDL/MEL config files for the DPT speech test from the data directory to the test directory where the main config already exists | 19 October 2015, 22:15:02 UTC |
5723c5c | Chris Basoglu | 19 October 2015, 19:04:50 UTC | Fix optional return value | 19 October 2015, 19:04:50 UTC |
17951ea | Chris Basoglu | 19 October 2015, 18:41:14 UTC | Fix return value | 19 October 2015, 18:41:14 UTC |
4832e65 | Chris Basoglu | 19 October 2015, 18:28:03 UTC | Fix windows conversion warning | 19 October 2015, 18:28:03 UTC |
bf5af27 | Chris Basoglu | 19 October 2015, 17:57:55 UTC | Add progress measurement features | 19 October 2015, 17:58:35 UTC |
2c72bc3 | erw | 18 October 2015, 22:57:39 UTC | Allow users to specify a PrefixPathInSCP option , so there is no need to revise scp after moving | 18 October 2015, 22:57:39 UTC |
4777f5f | Amit Agarwal | 18 October 2015, 19:40:43 UTC | Rename SequenceTraining project to CNTKSequenceTrainingLib to be consistent with other libs | 18 October 2015, 19:40:43 UTC |
180fae5 | Amit Agarwal | 18 October 2015, 03:35:35 UTC | Removed the no longer needed MANAGEDEXTERN device option | 18 October 2015, 04:53:51 UTC |
794911d | Amit Agarwal | 18 October 2015, 02:10:02 UTC | Removed default parameter value of 0 for deviceId in GPUMatrix. This is error prone and may accidentally lead to mismatch in deviceIds for different matrixes. The deviceId to a GPUMatrix should always be explicitly specified | 18 October 2015, 02:10:02 UTC |
4731a74 | Amit Agarwal | 16 October 2015, 22:02:42 UTC | Fixed a bug in sequence training code when copying of data between CNTK and SSE aligned matrices - the code previously was not accounting for alignment padding in the SSE matrix | 16 October 2015, 22:02:42 UTC |
03358d9 | erw | 11 October 2015, 06:03:11 UTC | Disable legacy usage. | 16 October 2015, 18:32:20 UTC |
96cb4b9 | Amit Agarwal | 16 October 2015, 00:00:41 UTC | Temporary disabling Release-GPU configuration for Speech/DNN/DiscriminativePreTraining test because of a known bug causing large variance between Release and Debug configurations for this test. This will be re-enabled after the bug has been addressed. This only effect the nighly-suite and keeps the BVT checkin suite unchanged | 16 October 2015, 00:51:23 UTC |
3b59002 | Amit Agarwal | 15 October 2015, 23:16:46 UTC | Added run-test-common EOL setting to .gitattributes to make cygwin happy | 15 October 2015, 23:16:46 UTC |
08119b8 | Amit Agarwal | 15 October 2015, 22:10:01 UTC | Added command line spew in the --dry-run option for TestDriver and also refactored common code across run-test scripts for different tests into a run-test-common script | 15 October 2015, 22:10:01 UTC |
547015a | Frank Seide | 15 October 2015, 21:04:50 UTC | bug fix: RowStackNode should call AssignToRowSliceValuesOf() instead of AssignRowSliceValuesOf() (see the difference?) | 15 October 2015, 21:04:50 UTC |
601b43e | U-FAREAST\fseide | 15 October 2015, 20:26:10 UTC | merged GPUMatrix.cu | 15 October 2015, 20:26:10 UTC |
c0e3c41 | Frank Seide | 15 October 2015, 20:08:49 UTC | updated DNN test case to narrower tolerances again | 15 October 2015, 20:08:49 UTC |
124d7f3 | Frank Seide | 15 October 2015, 20:00:51 UTC | CNTK now includes its own EXE path in the command line it logs; added code, currently #if-0'ed out, to set the random seed correctly (but slowly) for GPU | 15 October 2015, 20:00:51 UTC |
2fae709 | Alexey Kamenev | 15 October 2015, 17:43:34 UTC | Fix for ImageReader and MBLayout. | 15 October 2015, 18:33:19 UTC |
3a6fc85 | Frank Seide | 15 October 2015, 04:06:39 UTC | disabled initOnCPU again | 15 October 2015, 04:06:39 UTC |
dd41e50 | Frank Seide | 15 October 2015, 03:30:40 UTC | increased per-MB tolerance even more | 15 October 2015, 03:30:40 UTC |
ba8ad50 | Frank Seide | 15 October 2015, 03:03:25 UTC | changed DPT tolerances to 5%. Even between the baselines the first minibatch already differs by 3.2% | 15 October 2015, 03:03:25 UTC |
2ebc529 | Frank Seide | 15 October 2015, 02:34:22 UTC | set tolerance of DPT test to 1% until we figured out why Debug and Release are so (relatively) different. | 15 October 2015, 02:34:22 UTC |
e0b17e5 | Frank Seide | 15 October 2015, 01:30:47 UTC | made gcc happy | 15 October 2015, 01:30:47 UTC |
76d9539 | Frank Seide | 15 October 2015, 01:18:24 UTC | bug fix: MeanNode and InvStdDevNode must mask gaps in multi-sequence input to zero before reducing over them; revived MBLayout::DetermineActualNumSamples() | 15 October 2015, 01:18:24 UTC |
146140e | Frank Seide | 14 October 2015, 22:16:52 UTC | merged with fseide/winlstm; fixed a few error throws to XXXError() calls; moved one Validate() dimension check under isFinal condition | 14 October 2015, 22:16:52 UTC |
a4ed652 | Frank Seide | 14 October 2015, 22:02:58 UTC | changed Parameter initialization in ndl\macros.txt (used by DNN DPT test) to do randomization on the CPU, to avoid differences between configurations; added a comment to document the unexpected behavior of Parse() and what the previous fix works around | 14 October 2015, 22:02:58 UTC |
8127874 | Frank Seide | 14 October 2015, 21:31:06 UTC | added a bad hack to Parse() to stop it from nuking leading . and .. in pathnames--aargh; added Speech\DiscriminativePreTraining cmd line to README.txt for easy debugging | 14 October 2015, 21:31:06 UTC |
f829b4e | Frank Seide | 14 October 2015, 20:52:40 UTC | added the new test files to the Solution | 14 October 2015, 20:52:40 UTC |
fc9897b | Alexey Kamenev | 14 October 2015, 00:49:45 UTC | Merge fix. | 14 October 2015, 19:30:31 UTC |
a0e60f6 | Alexey Kamenev | 14 October 2015, 00:24:33 UTC | Correct handling of OpenCV path. | 14 October 2015, 19:30:21 UTC |
bb62af7 | Alexey Kamenev | 13 October 2015, 19:02:35 UTC | ImageReader fixes for Linux build. | 14 October 2015, 19:30:11 UTC |
085c29f | Alexey Kamenev | 13 October 2015, 00:54:13 UTC | Add CUB to Linux build. | 14 October 2015, 19:30:01 UTC |
78d2e0b | Alexey | 12 October 2015, 23:36:40 UTC | Add CUB path. | 14 October 2015, 19:29:51 UTC |
64fec93 | Alexey Kamenev | 09 October 2015, 22:28:36 UTC | Enabled check for CUB and conditional build for OpenCV. | 14 October 2015, 19:29:41 UTC |
3a6abed | Alexey Kamenev | 09 October 2015, 01:06:20 UTC | Merge fixes. | 14 October 2015, 19:29:31 UTC |
edab352 | Alexey Kamenev | 05 October 2015, 22:57:06 UTC | Merge fixes. | 14 October 2015, 19:29:21 UTC |
797c727 | Alexey Kamenev | 21 September 2015, 19:58:36 UTC | Fix merge issues. | 14 October 2015, 19:29:11 UTC |
a5acf9d | Alexey Kamenev | 21 September 2015, 19:01:24 UTC | Refactor to use conc_stack. | 14 October 2015, 19:29:01 UTC |
98324b8 | Alexey Kamenev | 17 September 2015, 19:19:04 UTC | Add unit tests for TopK-related Matrix changes. | 14 October 2015, 19:28:50 UTC |
e0ee2b3 | Alexey Kamenev | 16 September 2015, 23:45:34 UTC | Fix for NoGPU. | 14 October 2015, 19:28:40 UTC |
3129a35 | Alexey Kamenev | 16 September 2015, 19:35:21 UTC | Minor changes in ImageReader and TopK eval. | 14 October 2015, 19:28:30 UTC |
7b751bf | Alexey Kamenev | 16 September 2015, 17:30:09 UTC | Add TopK error evaluation | 14 October 2015, 19:28:20 UTC |
fa4383e | Alexey Kamenev | 10 September 2015, 22:24:26 UTC | Add TopK error evaluation | 14 October 2015, 19:28:10 UTC |
35d4a54 | Alexey Kamenev | 09 September 2015, 03:52:30 UTC | Minor changes | 14 October 2015, 19:28:00 UTC |
6cd863e | Alexey Kamenev | 09 September 2015, 00:58:45 UTC | Refactoring to use transforms | 14 October 2015, 19:27:50 UTC |
ecdd9ae | Alexey Kamenev | 05 September 2015, 00:51:01 UTC | Add crop transform | 14 October 2015, 19:27:39 UTC |
9aba364 | Alexey Kamenev | 27 August 2015, 01:50:56 UTC | Add ImageReader | 14 October 2015, 19:27:29 UTC |
264f3c7 | Frank Seide | 14 October 2015, 15:22:17 UTC | further simplified RowStackNode, e.g. we actually do not need m_startRowIndices[num children] | 14 October 2015, 15:22:17 UTC |
19e448d | Frank Seide | 14 October 2015, 15:03:48 UTC | RowStackNode() now uses Matrix::AssignRowSliceValuesOf() instead of AssignRowStackValuesOf() because that allows to use ValueSlice() like all other nodes. The previous method passed the column indices directly, but it is no longer of nodes' concern that they operate on column slices (and this caused a bug for Abdo's config after eliminating RowStackNode::EveluateThisNodeMap()); also #if-0'ed out AssignRowStackValuesOf() from all matrix libs as it is both very complex and no longer used; removed RowStackNode::m_inputMatrices[] since it was redundant | 14 October 2015, 15:03:48 UTC |
0c7eb75 | Frank Seide | 14 October 2015, 01:09:34 UTC | added a workaround for Check_t() (which will be fixed properly after memshare merge) | 14 October 2015, 01:09:34 UTC |
b0b4278 | Frank Seide | 13 October 2015, 23:16:16 UTC | added missing DiagonalToDense() stub to NoGPP.cpp; parallelforwardbackward.cpp now includes BestGpu.h which prevents it from referencing CNTKMathCUDA.lib in CPUONLY builds | 13 October 2015, 23:16:16 UTC |
91f85af | Frank Seide | 13 October 2015, 21:59:38 UTC | Merge branch 'master' of https://git.codeplex.com/cntk into fseide/winlstm | 13 October 2015, 21:59:38 UTC |
864d1c1 | Frank Seide | 13 October 2015, 21:58:16 UTC | removed all calls to EvaluateThisNodeMap(), all are now using the FrameRange code branch | 13 October 2015, 21:58:16 UTC |
f904d91 | Amit Agarwal | 13 October 2015, 21:29:47 UTC | Removed testing for EvalErr in each mini-batch - too flaky | 13 October 2015, 21:29:47 UTC |
f02cf9c | Amit Agarwal | 13 October 2015, 17:23:16 UTC | Loosen the tolerance of the Discriminative pretraining E2E test | 13 October 2015, 17:23:16 UTC |
450b16b | Amit Agarwal | 13 October 2015, 05:10:06 UTC | Fixed a bug where labelIDs assigment was incorrectly being done unconditionally when it only applies when lattices exist | 13 October 2015, 05:12:21 UTC |
3a90db9 | Amit | 13 October 2015, 02:05:23 UTC | Merge branch 'master' of https://git.codeplex.com/cntk into amitaga/newSpeechTests | 13 October 2015, 03:59:24 UTC |
57cd8b8 | Amit Agarwal | 13 October 2015, 01:45:04 UTC | Added a more comprehensive Speech DNN training test that uses discriminative pre-training, NDL and MEL | 13 October 2015, 01:59:23 UTC |
5d13127 | Frank Seide | 13 October 2015, 01:36:49 UTC | Merge branch 'master' of https://git.codeplex.com/cntk into fseide/winlstm | 13 October 2015, 01:36:49 UTC |
d4e5455 | Yu | 13 October 2015, 01:26:32 UTC | Fixed linux build | 13 October 2015, 01:26:32 UTC |
d80046b | Frank Seide | 13 October 2015, 01:15:45 UTC | fixed a refactoring error that caused variable shadowing in ComputeGradient() which broke the last commit | 13 October 2015, 01:15:45 UTC |
5dfda1f | Frank Seide | 12 October 2015, 21:57:33 UTC | bug fixes in Debug mode/NaN checks--now resets gaps to 0 before computing a gradient. This is a stop-gap, the correct encapsulation is to clear gradient gap columns wherever they are *used* instead of where they are *generated*; If the final model exists, SGD now prints the path to log; ValidateSubNetwork() now tracks changes in m_needsGradient flag | 12 October 2015, 21:57:33 UTC |
c9b4987 | Frank Seide | 11 October 2015, 05:02:02 UTC | Merge branch 'master' of https://git.codeplex.com/cntk into fseide/winlstm | 11 October 2015, 05:02:02 UTC |
0da6aec | Frank Seide | 11 October 2015, 04:56:25 UTC | split m_needGradient into m_parameterUpdateRequired (node says whether it is to be updated) and m_needsGradient (network says whether this node or any of its children require a gradient to be passed in and/or passed through). m_needsGradient is propagated in Validate() (also in its old init location, which is wrong); added noexcept to ConfigValuePtr constructor and assignment | 11 October 2015, 04:56:25 UTC |
acc741d | Amit | 11 October 2015, 01:58:55 UTC | Fixed Linux warnings | 11 October 2015, 01:58:55 UTC |
ae528fb | Amit | 11 October 2015, 01:45:45 UTC | Fixed linux build | 11 October 2015, 01:45:45 UTC |
b8adcb5 | Amit Agarwal | 11 October 2015, 01:07:19 UTC | Replaced all bare throw calls with Centralized exception functions that print the call stack | 11 October 2015, 01:38:59 UTC |
5998260 | Frank Seide | 11 October 2015, 01:29:41 UTC | Readded lost latticefunctionskernels.h. What is wrong with VS?? | 11 October 2015, 01:29:41 UTC |
d4ca7b2 | Frank Seide | 11 October 2015, 01:23:44 UTC | removed an extern definition of fileno(), which somehow seems to no longer be needed for Linux (probably some other change brought in a missing header) | 11 October 2015, 01:23:44 UTC |
8c6ca65 | Frank Seide | 11 October 2015, 01:18:59 UTC | added _wchdir() emulation for Linux to Platform.h | 11 October 2015, 01:18:59 UTC |
704faf3 | Frank Seide | 11 October 2015, 01:14:26 UTC | added NaN checks (commented out) and NaN gap-blasting to gradient computation; added NaN check to SGD right before UpdateWeights() (_DEBUG only) | 11 October 2015, 01:14:26 UTC |
f4bcfc5 | Frank Seide | 11 October 2015, 00:45:00 UTC | reapplied the fix to atomicLogAdd() where the result of atomicCAS() cannot be compared as floats, but must be compared as bit patterns, as to make it work for NaNs | 11 October 2015, 00:45:00 UTC |
9cb50e3 | Frank Seide | 11 October 2015, 00:19:40 UTC | bug workaround: MinusNode had m_needGradient false while a child had it true, which caused use of an empty gradient matrix. Worked around by going back to old CNTK way of always resizing gradients as well; UpdateFunctionAndGradientMBSize() new command-line option currentDirectory=... to set the CD. Use this as the first option. This allows for self-contained debug command lines; implemented Yu Zhang's recent bug fix differently by moving the MB reset into LoadPersistableParametersFromFile() | 11 October 2015, 00:19:40 UTC |
df5269a | Frank Seide | 10 October 2015, 23:14:49 UTC | Merge branch 'master' of https://git.codeplex.com/cntk into fseide/winlstm | 10 October 2015, 23:14:49 UTC |
d9cdf9e | Amit Agarwal | 10 October 2015, 18:29:22 UTC | Replace use of basetypes.h header in HTKMLFReader with Basics.h | 10 October 2015, 19:12:16 UTC |
807f021 | Amit | 10 October 2015, 09:13:25 UTC | Fixed linux build | 10 October 2015, 09:13:25 UTC |
4d31376 | Amit Agarwal | 10 October 2015, 09:06:03 UTC | Merged Common/basetypes.h and HTKMLFReader/basetypes.h and deleted the HTKMLFReader/basetypes.h copy | 10 October 2015, 09:06:03 UTC |
3d2d626 | Amit Agarwal | 10 October 2015, 07:22:36 UTC | Centralized PrintCallStack calls | 10 October 2015, 07:22:36 UTC |
9a4c84b | Amit | 10 October 2015, 06:54:37 UTC | Fixed Linux build | 10 October 2015, 06:54:37 UTC |
0150884 | Amit Agarwal | 10 October 2015, 06:37:44 UTC | Merge branch 'master' of https://git01.codeplex.com/cntk into bmitra/Changes | 10 October 2015, 06:37:44 UTC |
542f704 | Amit Agarwal | 10 October 2015, 06:35:18 UTC | Adapted the new DiagonalNode to recent structural changes in computation node pattern | 10 October 2015, 06:35:18 UTC |
64ee4ba | Yu | 10 October 2015, 06:14:48 UTC | Reset the MBLayout when we reload the best model. Hasn't carefully read the MBLayout class, so just call the init function to set it to (1,0,false). Also let the loadFromNetwork function in delay node do not resize the cols. | 10 October 2015, 06:14:48 UTC |
349f07f | Amit Agarwal | 10 October 2015, 05:58:15 UTC | Merge branch 'master' of https://git01.codeplex.com/cntk into bmitra/Changes Conflicts: Common/Include/Basics.h DataReader/LMSequenceReader/LMSequenceReader.vcxproj DataReader/SparsePCReader/SparsePCReader.vcxproj MachineLearning/CNTKComputationNetworkLib/NonlinearityNodes.h Math/Math/CNTKMathCUDA.vcxproj.filters Math/Math/CommonMatrix.h Math/Math/GPUMatrix.cu Math/Math/GPUSparseMatrix.cu Math/Math/Matrix.h | 10 October 2015, 06:07:34 UTC |
4a133d2 | Frank Seide | 10 October 2015, 00:05:31 UTC | Merge branch 'guoguo/linuxBuildFix' of https://git.codeplex.com/cntk into fseide/winlstm | 10 October 2015, 00:05:31 UTC |