06b7626 | Frank Seide | 25 August 2016, 22:07:19 UTC | (made VC happy) | 25 August 2016, 22:07:19 UTC |
16dd77e | Frank Seide | 25 August 2016, 21:23:12 UTC | Merge branch 'slayton58-divmod' into fseide/cudnn5 | 25 August 2016, 21:23:12 UTC |
828b393 | Simon Layton | 27 July 2016, 17:22:41 UTC | Initial divmod implementation | 25 August 2016, 17:29:21 UTC |
4cb011b | Eldar Akchurin | 25 August 2016, 16:28:20 UTC | Adapting the sister test | 25 August 2016, 16:28:20 UTC |
e1cf519 | Mark Hillebrand | 25 August 2016, 14:26:03 UTC | CuDNNv5: enable for build v5.1 is preferred over v5. | 25 August 2016, 14:26:03 UTC |
e459613 | Mark Hillebrand | 25 August 2016, 14:25:18 UTC | Baselines | 25 August 2016, 14:25:18 UTC |
d9e9c88 | Mark Hillebrand | 25 August 2016, 13:44:53 UTC | Merge remote-tracking branch 'origin/fseide/cudnn5' into mahilleb/CuDnn5Test | 25 August 2016, 13:44:53 UTC |
58f18f8 | Mark Hillebrand | 25 August 2016, 13:42:24 UTC | Merge remote-tracking branch 'origin/master' into mahilleb/CuDnn5Test | 25 August 2016, 13:42:24 UTC |
6827182 | Mark Hillebrand | 25 August 2016, 13:38:08 UTC | Merge remote-tracking branch 'origin/mahilleb/CuDnn5Test' into mahilleb/CuDnn5Test | 25 August 2016, 13:38:08 UTC |
fc3a071 | Mark Hillebrand | 25 August 2016, 13:37:23 UTC | CntkBatchNormalization.cuh: fix for batchSize == 1 | 25 August 2016, 13:37:23 UTC |
7e82e9b | Philipp Kranen | 19 August 2016, 15:43:51 UTC | Added e2e test for image regression example | 25 August 2016, 09:26:23 UTC |
e7a273d | Philipp Kranen | 18 August 2016, 15:47:47 UTC | added simple image regression examples for CIFAR-10 | 25 August 2016, 09:26:12 UTC |
a885a43 | Mark Hillebrand | 16 August 2016, 07:38:25 UTC | Log batch-normalizatoin runInvStdDev values after model loading. | 25 August 2016, 07:50:08 UTC |
d6da982 | Frank Seide | 25 August 2016, 05:53:32 UTC | fixed cudnn5 RNN tests | 25 August 2016, 05:53:32 UTC |
af423dd | Frank Seide | 25 August 2016, 05:48:13 UTC | disabled cudnn5 RNN and Layers RNN test on Linux until we create the baselines | 25 August 2016, 05:48:13 UTC |
9c0a602 | Frank Seide | 25 August 2016, 05:44:09 UTC | FullUtteranceCuDNN5 now restricted to run on gpu, actually | 25 August 2016, 05:44:09 UTC |
2be0dc6 | Frank Seide | 25 August 2016, 05:14:25 UTC | updated Truncated test | 25 August 2016, 05:14:25 UTC |
b814ff8 | Frank Seide | 25 August 2016, 05:01:37 UTC | updated Truncated momentum value and missign baselines | 25 August 2016, 05:01:37 UTC |
6a1c9a7 | Frank Seide | 25 August 2016, 04:51:41 UTC | added RecurrentLSTMLayerStack() | 25 August 2016, 04:51:41 UTC |
f3942f1 | Frank Seide | 25 August 2016, 03:54:08 UTC | added missing RNN bits to Linux Makefile | 25 August 2016, 03:54:08 UTC |
4e19029 | Frank Seide | 25 August 2016, 03:42:48 UTC | RNNnodes.cpp missing in Linux Makefile; changed test tolerance to accept a momentum printed as 0.899998 instead of the original 0.9 | 25 August 2016, 03:42:48 UTC |
e576c3d | Frank Seide | 25 August 2016, 03:24:31 UTC | missing NoGPU.cpp entries for RNN node; fixed shared_ptr to incomplete CuDnnRNNExecutor | 25 August 2016, 03:24:31 UTC |
686078f | Frank Seide | 25 August 2016, 02:51:02 UTC | (made gcc happy) | 25 August 2016, 02:51:02 UTC |
0bff544 | Frank Seide | 25 August 2016, 02:48:46 UTC | created baselines for LSTM E2E with Layers and cudnn5 | 25 August 2016, 02:48:46 UTC |
32c9d11 | Frank Seide | 25 August 2016, 02:29:46 UTC | added e2e tests for LSTM Layers and cudnn5, missing baselines | 25 August 2016, 02:29:46 UTC |
7558fc4 | Frank Seide | 25 August 2016, 02:18:44 UTC | repaired LSTM e2e test for old model | 25 August 2016, 02:18:44 UTC |
5f14fca | Frank Seide | 25 August 2016, 01:45:39 UTC | merged from mahilleb/CuDnn5Test | 25 August 2016, 01:45:39 UTC |
a144404 | Frank Seide | 25 August 2016, 01:33:36 UTC | RecurrentLSTMLayer{} now allows to select the optimized engine | 25 August 2016, 01:33:36 UTC |
769b260 | Frank Seide | 25 August 2016, 01:15:09 UTC | updated SLUHandsOn tests | 25 August 2016, 01:15:09 UTC |
9e73e6f | Frank Seide | 25 August 2016, 00:37:53 UTC | bug fix in RNN node input swapping; LSTM e2e for RNN node now uses dimension inference | 25 August 2016, 00:37:53 UTC |
d9c7e82 | Frank Seide | 25 August 2016, 00:17:23 UTC | OptimizedRNNStackNode: renamed some variables, renamed recurrentOps to camelCase, added weigth inference | 25 August 2016, 00:17:23 UTC |
179b821 | Mark Hillebrand | 24 August 2016, 23:39:00 UTC | Tests/EndToEndTests/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv: baseline update | 24 August 2016, 23:39:00 UTC |
8a86da8 | Frank Seide | 24 August 2016, 23:10:01 UTC | renamed RNNNode to OptimizedRNNStackNode, also updated parameter names | 24 August 2016, 23:10:01 UTC |
0285fa9 | Mark Hillebrand | 24 August 2016, 22:29:10 UTC | Merge remote-tracking branch 'origin/master' into mahilleb/CuDnn5Test Conflicts: Source/ComputationNetworkLib/ComputationNode.h Source/ComputationNetworkLib/TrainingNodes.h Tests/EndToEndTests/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv/baseline.linux.txt Tests/EndToEndTests/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv/baseline.windows.txt Tests/UnitTests/MathTests/ConvolutionEngineTests.cpp | 24 August 2016, 22:37:49 UTC |
e5f0bb9 | Frank Seide | 24 August 2016, 22:32:24 UTC | disabled peephole connections by default in RecurrentLSTMLayer{} | 24 August 2016, 22:32:24 UTC |
b2a18dd | Frank Seide | 24 August 2016, 22:10:45 UTC | LSTMBlock{} now combining 4 matrices into one | 24 August 2016, 22:10:45 UTC |
150aedd | Frank Seide | 24 August 2016, 21:55:58 UTC | towards optimized LSTMBlock | 24 August 2016, 21:55:58 UTC |
00e0249 | Frank Seide | 24 August 2016, 21:35:47 UTC | added cudnn5 emulation version to LSTM e2e test | 24 August 2016, 21:35:47 UTC |
0b049c0 | Frank Seide | 24 August 2016, 19:25:43 UTC | added cudnn to LSTM e2e | 24 August 2016, 19:25:43 UTC |
90bbcc1 | Frank Seide | 24 August 2016, 18:06:24 UTC | (updated param init in LSTM e2e) | 24 August 2016, 18:06:24 UTC |
453bd59 | Frank Seide | 24 August 2016, 17:39:10 UTC | (a good layer-norm setup for LSTM e2e) | 24 August 2016, 17:39:10 UTC |
e64dba1 | Frank Seide | 24 August 2016, 17:25:31 UTC | added layers version to LSTM e2e test | 24 August 2016, 17:25:31 UTC |
737fe4c | Frank Seide | 24 August 2016, 17:06:42 UTC | updated LSTM test to newer style | 24 August 2016, 17:06:42 UTC |
d8c3c15 | Frank Seide | 24 August 2016, 16:44:29 UTC | merged from CuDnn5Test and cudnn-rnn | 24 August 2016, 16:44:29 UTC |
9d6345b | Frank Seide | 24 August 2016, 15:05:13 UTC | Merge branch 'master' of https://github.com/Microsoft/CNTK into fseide/kdd | 24 August 2016, 15:05:13 UTC |
493744d | Mark Hillebrand | 24 August 2016, 12:00:22 UTC | Source/Math/CntkBatchNormalization.cuh: fix variance conversion | 24 August 2016, 12:00:22 UTC |
e27af11 | Project Philly | 24 August 2016, 07:47:02 UTC | Integrate zhouwang/rebase-mtNonUniform2 into master | 24 August 2016, 07:47:02 UTC |
d1b1127 | Frank Seide | 24 August 2016, 07:20:13 UTC | Merge branch 'jdroppo/cudnn-rnn-lstm' of https://github.com/Microsoft/cntk into fseide/cudnn5 | 24 August 2016, 07:20:13 UTC |
835985b | Frank Seide | 24 August 2016, 06:39:28 UTC | fixed test | 24 August 2016, 06:39:28 UTC |
bd5f9cb | Frank Seide | 24 August 2016, 04:54:00 UTC | addressed CR feedback; modifed an e2e test to cover inferInputRankToMap | 24 August 2016, 04:54:00 UTC |
7aa9b83 | Frank Seide | 24 August 2016, 01:12:36 UTC | merged from master | 24 August 2016, 01:12:36 UTC |
3c5aa75 | Frank Seide | 24 August 2016, 00:59:31 UTC | changed inferInputRank to inferInputRankToMap, and updated {Linear,Dense}Layer{} to accept the mutually exclusive inputRank and mapRank parameters | 24 August 2016, 00:59:31 UTC |
de864bf | Frank Seide | 24 August 2016, 00:32:03 UTC | renamed Times inputRank to inferInputRank | 24 August 2016, 00:32:03 UTC |
8dfb1e2 | Mark Hillebrand | 23 August 2016, 23:28:21 UTC | Report on converting minibatch count to samples | 23 August 2016, 23:33:03 UTC |
bb155ef | Mark Hillebrand | 23 August 2016, 22:06:49 UTC | tune | 23 August 2016, 22:06:49 UTC |
e1a9cab | Mark Hillebrand | 23 August 2016, 15:57:00 UTC | Address CR comments | 23 August 2016, 18:32:03 UTC |
373adfd | Jasha Droppo | 23 August 2016, 17:40:34 UTC | Changes Addressing Code Review for CUDNN RNNStack Node | 23 August 2016, 17:40:34 UTC |
86ec025 | Zhou Wang | 17 August 2016, 13:27:15 UTC | Use mt19937 instead of rand() | 23 August 2016, 17:24:45 UTC |
705951a | Zhou Wang | 11 August 2016, 13:02:59 UTC | adjust baseline with new boost random distribution function change std random distribution to boost random distribution | 23 August 2016, 17:23:46 UTC |
3179559 | Zhou Wang | 11 August 2016, 12:40:58 UTC | use boost random distribution functions instead of std | 23 August 2016, 17:23:42 UTC |
c11ad31 | Zhou Wang | 09 August 2016, 13:51:57 UTC | Using slash instead of backslash | 23 August 2016, 17:23:26 UTC |
a2565ab | Zhou Wang | 08 August 2016, 19:05:50 UTC | fix bug that removes path prefix in the file name on Linux Use comparision with tolerance use '/' instead of '\' in path in data files. | 23 August 2016, 17:22:47 UTC |
6744b96 | Zhou Wang | 08 August 2016, 12:52:00 UTC | check output with tolerance optionally allow tolerance when comparing file content | 23 August 2016, 17:21:43 UTC |
aeee8f6 | Momcilo Vasilijevic | 09 June 2016, 14:12:24 UTC | Add bilinear initialization of learnable parameter Enable filling learnable parameter with coefficients for bilinear interpolation. A common use is with deconvolution acting as upsampling. | 23 August 2016, 15:26:25 UTC |
6aea378 | Mark Hillebrand | 23 August 2016, 11:12:45 UTC | Baseline update | 23 August 2016, 11:12:45 UTC |
66498cf | Mark Hillebrand | 23 August 2016, 09:12:35 UTC | Merge remote-tracking branch 'origin/master' into mahilleb/CuDnn5Test Note: baselines need to be fixed for Tests/EndToEndTests/BatchNormalization and Tests/EndToEndTests/Examples/Image/Miscellaneous/CIFAR-10/02_BatchNormConv. | 23 August 2016, 09:12:35 UTC |
d4f40dd | jeanfad | 23 August 2016, 08:21:42 UTC | adapt the swig wrapper to the change in getNextMinibatch method | 23 August 2016, 08:21:42 UTC |
10271ba | jeanfad | 23 August 2016, 07:38:24 UTC | Merge branch 'master' into blis/PythonV2 | 23 August 2016, 07:38:24 UTC |
f46e911 | jeanfad | 23 August 2016, 07:37:03 UTC | move nn.py to examples folder | 23 August 2016, 07:37:03 UTC |
1f9c539 | Frank Seide | 23 August 2016, 03:11:07 UTC | Merge branch 'jdroppo/cudnn-rnn-lstm' of https://github.com/Microsoft/cntk into fseide/cudnn5 | 23 August 2016, 03:11:07 UTC |
4983c46 | Frank Seide | 23 August 2016, 01:53:35 UTC | fixed up two new baselines w.r.t. ErrorPrediction -> ClassificationError | 23 August 2016, 01:53:35 UTC |
54f0960 | Frank Seide | 23 August 2016, 01:51:22 UTC | merged from mahilleb/CuDnn5Test | 23 August 2016, 01:51:22 UTC |
abe6624 | Frank Seide | 23 August 2016, 00:35:15 UTC | re-fixed baselines from ErrorPrediction to ClassificationError | 23 August 2016, 00:35:15 UTC |
320f2f5 | Frank Seide | 23 August 2016, 00:27:23 UTC | re-merged baseline files from master, seems I mixed them up | 23 August 2016, 00:27:23 UTC |
f32db3e | Frank Seide | 23 August 2016, 00:11:23 UTC | undid the test case of BrainScript from ClassificationError to ErrorPrediction | 23 August 2016, 00:11:23 UTC |
ed7e602 | Frank Seide | 22 August 2016, 21:53:27 UTC | reapplied renaming from ErrorPrediction to ClassificationError for baseline files after merge | 22 August 2016, 21:53:27 UTC |
5b969ba | Frank Seide | 22 August 2016, 21:36:28 UTC | merged from master. Undid the ClassificationError baseline updates due to merge conflicts | 22 August 2016, 21:36:28 UTC |
1e68b3c | Frank Seide | 22 August 2016, 20:32:19 UTC | changed some more baselines from ErrorPrediction to ClassificationError | 22 August 2016, 20:32:19 UTC |
2fa1b70 | Jasha Droppo | 22 August 2016, 20:28:42 UTC | Merge commit 'origin/master' 8493f11 into jdroppo/cudnn-rnn-lstm | 22 August 2016, 20:28:42 UTC |
363f1e5 | Frank Seide | 22 August 2016, 20:27:40 UTC | changed some more baselines from ErrorPrediction to ClassificationError | 22 August 2016, 20:27:40 UTC |
8493f11 | Amit Agarwal | 22 August 2016, 18:54:21 UTC | CNTK v2 library: Clone constants to right device when user supplied value is not on the target compute device | 22 August 2016, 18:54:21 UTC |
ba01512 | Frank Seide | 22 August 2016, 18:23:09 UTC | updated MatrixQuantizerTests to new >1-bit SGD defaults; addressed CR feedback | 22 August 2016, 18:23:09 UTC |
536e264 | Frank Seide | 22 August 2016, 18:16:39 UTC | renamed ErrorPrediction to ClassificationError, second attempt | 22 August 2016, 18:16:39 UTC |
37b6897 | Amit Agarwal | 22 August 2016, 17:48:54 UTC | Merge branch 'master' of https://github.com/Microsoft/CNTK into amitaga/cntkv2Library | 22 August 2016, 17:48:54 UTC |
9bd5738 | Frank Seide | 22 August 2016, 17:30:17 UTC | renamed ErrorPrediction to ClassificationError | 22 August 2016, 17:30:17 UTC |
f3ce815 | Mark Hillebrand | 22 August 2016, 16:34:04 UTC | Baseline updates | 22 August 2016, 16:54:53 UTC |
1fa5b23 | jeanfad | 22 August 2016, 16:41:27 UTC | resnet cfair first complete network run | 22 August 2016, 16:41:27 UTC |
f76afa2 | Mark Hillebrand | 08 August 2016, 07:13:39 UTC | Switch to CuDNN v5 For batch normalization, running inverse standard deviation becomes running variance. We mirror this CuDNN v5 change in the CNTK batch normalization engine. Model version is bumped. When old models are loaded, this parameter is (approximately) converted. In the same model version change, let batch normalization count samples seen rather minibatches (this deals with incorrect averaging when minibatch size is varied across epochs). For batch normalization averaging and blending handle initialization cases, don't rely on mean and variance initial values (set in NDL/BrainScript). Update Windows / Linux / Docker build. With this commit, CuDNN v4 is not supported anymore. | 22 August 2016, 15:55:10 UTC |
9521d7f | jeanfad | 22 August 2016, 14:34:41 UTC | move layer creation functions to a separate module | 22 August 2016, 14:34:41 UTC |
4df1389 | jeanfad | 22 August 2016, 14:13:28 UTC | convert the loss value to float | 22 August 2016, 14:13:28 UTC |
d2eed70 | jeanfad | 22 August 2016, 14:12:41 UTC | turn mnist and ffnet python example into py.test tests | 22 August 2016, 14:12:41 UTC |
4e174f2 | Project Philly | 22 August 2016, 13:45:48 UTC | Integrate eldak/mtNonUniform2 into master | 22 August 2016, 13:45:48 UTC |
a1528b9 | jeanfad | 22 August 2016, 13:22:54 UTC | fix to_numpy in swig so it reversed the shape to row major, and fix the shape reversal in ndshape.dimensions() | 22 August 2016, 13:22:54 UTC |
6a063e2 | Project Philly | 22 August 2016, 13:01:46 UTC | Integrate zhouwang/fix-deviceid-threadlocal into master | 22 August 2016, 13:01:46 UTC |
de1c1e5 | jeanfad | 22 August 2016, 12:23:48 UTC | adapt the python api to the new Learner c++ api | 22 August 2016, 12:23:48 UTC |
941931d | jeanfad | 22 August 2016, 09:29:14 UTC | Merge branch 'master' into blis/PythonV2 | 22 August 2016, 09:29:14 UTC |
e34528e | jeanfad | 22 August 2016, 09:24:55 UTC | fix file paths in examples, in order to work with py.test | 22 August 2016, 09:24:55 UTC |
6566012 | Zhou Wang | 22 August 2016, 08:04:47 UTC | Add comment to use thread_local after uprade to VS2015 | 22 August 2016, 08:04:47 UTC |
d5f47f1 | Eldar Akchurin | 10 August 2016, 16:09:01 UTC | Updating sources and baselines to use MT | 22 August 2016, 07:49:08 UTC |
b856ab7 | Eldar Akchurin | 10 August 2016, 10:43:58 UTC | Adding support for mersenne twister rand | 21 August 2016, 19:08:25 UTC |