474570d | Frank Seide | 11 August 2016, 06:47:44 UTC | first version of ResNet in TutorialImage.cntk; breaking change: ConvolutionalLayer{} and XXPoolingLayer{} now both default to padding=false, like Keras | 11 August 2016, 06:47:44 UTC |
8ecc7b1 | Frank Seide | 11 August 2016, 03:33:19 UTC | readded BN to TutorialImage.CNTK | 11 August 2016, 03:33:19 UTC |
8a79eda | Frank Seide | 11 August 2016, 01:53:39 UTC | got a new baseline for TutorialImage.cntk/simple network, with latest reader architecture | 11 August 2016, 01:53:39 UTC |
6e32e40 | Frank Seide | 11 August 2016, 01:18:34 UTC | switched convnet to new reader architecture | 11 August 2016, 01:18:34 UTC |
ea3b31f | Frank Seide | 11 August 2016, 00:42:46 UTC | added new reader to TutorialImage.cntk | 11 August 2016, 00:42:46 UTC |
e1d0cb2 | Frank Seide | 10 August 2016, 23:36:41 UTC | bug fix in GetMatrixInfo: should check gradient matrix ptr for being NULL instead of asking NeedsGradient() | 10 August 2016, 23:36:41 UTC |
19233d7 | Frank Seide | 10 August 2016, 19:06:40 UTC | changed shareNodeValueMatrices to true by default | 10 August 2016, 19:06:40 UTC |
34cb4a9 | Frank Seide | 10 August 2016, 18:08:09 UTC | created ConvNet.cntk | 10 August 2016, 18:08:09 UTC |
9f6306c | Frank Seide | 10 August 2016, 18:07:19 UTC | replaced 01_Convolution.bs.cntk with TutorialImage.cntk | 10 August 2016, 18:07:19 UTC |
03143a0 | Frank Seide | 10 August 2016, 17:47:25 UTC | printing matrix info no longer requires ElemType and has streamlined output; added parameter dimension inference to BatchNormalizationNode | 10 August 2016, 17:47:25 UTC |
d071351 | Frank Seide | 10 August 2016, 06:56:24 UTC | fragments of parameter inference for BN, but failed due to existing bug; added BN to TutorialImage | 10 August 2016, 06:56:24 UTC |
3be8fab | Frank Seide | 10 August 2016, 04:39:29 UTC | (made gcc happy) | 10 August 2016, 04:39:29 UTC |
7948ec4 | Frank Seide | 10 August 2016, 04:33:47 UTC | (minor cleanup) | 10 August 2016, 04:33:47 UTC |
6772d27 | Frank Seide | 10 August 2016, 04:33:21 UTC | added a BS version of 01_Conv | 10 August 2016, 04:33:21 UTC |
4cd66b6 | Frank Seide | 10 August 2016, 04:26:20 UTC | minor cleanup of TutorialImage.cntk | 10 August 2016, 04:26:20 UTC |
ce550d9 | Frank Seide | 10 August 2016, 04:24:17 UTC | added Sequential style to TutorialImage | 10 August 2016, 04:24:17 UTC |
5567398 | Frank Seide | 10 August 2016, 04:08:28 UTC | implemented dimension inference for Convolution() | 10 August 2016, 04:08:28 UTC |
afb0175 | Frank Seide | 10 August 2016, 00:54:20 UTC | ND-Pooling no longer requires to specify reduction dimensions for kernel etc. (they will be inferred) | 10 August 2016, 00:54:20 UTC |
e44b53e | Frank Seide | 09 August 2016, 23:38:42 UTC | adapted seqcla.cntk to a renamed optional parameter | 09 August 2016, 23:38:42 UTC |
b32a04e | Frank Seide | 09 August 2016, 23:35:53 UTC | bug fix: Times() should infer tensor ranks correctly | 09 August 2016, 23:35:53 UTC |
9046f9a | Frank Seide | 09 August 2016, 23:01:16 UTC | towards new structure for 01_Conv CIFAR-10 | 09 August 2016, 23:01:16 UTC |
27c3697 | Frank Seide | 09 August 2016, 22:33:20 UTC | created a first BS version of CIFAR-10 01_Conv | 09 August 2016, 22:33:20 UTC |
30ff99b | Frank Seide | 09 August 2016, 17:01:06 UTC | tried bidirectional intent classifier | 09 August 2016, 17:01:06 UTC |
53d5222 | Frank Seide | 09 August 2016, 05:51:17 UTC | added intent classification to ATIS sample | 09 August 2016, 05:51:17 UTC |
f9c10a8 | Frank Seide | 09 August 2016, 05:08:54 UTC | created slot tagger in new style from Derek's example; added a layer-like LSTM called LSTMBlock{} and fixed RecurrentLSTMLayer{} with that; ComputationNetworkBuilder from BrainScript now force-uses top-level node names for all nodes that are reachable from outer record, as that is the name the user intended | 09 August 2016, 05:08:54 UTC |
71c2a94 | Frank Seide | 09 August 2016, 02:14:06 UTC | switched to new CTF files (not cleaned up yet); fixed a CTF preparation error | 09 August 2016, 02:14:06 UTC |
b96a508 | Frank Seide | 09 August 2016, 01:54:57 UTC | added pre-processed ATIS files and first step towards tutorials | 09 August 2016, 01:54:57 UTC |
595c7ab | Frank Seide | 08 August 2016, 17:04:02 UTC | addressed review feedback | 08 August 2016, 17:04:02 UTC |
d6fb378 | Frank Seide | 08 August 2016, 04:27:54 UTC | bug fix: CPUSparseMatrix<ElemType>::MultiplyAndWeightedAdd() should handle transposed inputs in all combinations | 08 August 2016, 04:27:54 UTC |
784d72e | Frank Seide | 08 August 2016, 02:39:34 UTC | Merge branch 'master' of https://github.com/Microsoft/cntk into fseide/layerslib | 08 August 2016, 02:39:34 UTC |
8d7819c | Frank Seide | 08 August 2016, 02:13:53 UTC | deleted remnants | 08 August 2016, 02:13:53 UTC |
efae278 | Frank Seide | 08 August 2016, 02:11:03 UTC | seqcla changed to define the model through layer composition | 08 August 2016, 02:11:03 UTC |
441f828 | Frank Seide | 08 August 2016, 01:47:52 UTC | seqcla now using layers but in graph style | 08 August 2016, 01:47:52 UTC |
5cd5ec8 | Frank Seide | 08 August 2016, 01:34:14 UTC | converted seqcla to sparse input | 08 August 2016, 01:34:14 UTC |
8bfd693 | Frank Seide | 08 August 2016, 01:07:38 UTC | (more refacoring of seqcla test) | 08 August 2016, 01:07:38 UTC |
e4fd6c8 | Frank Seide | 08 August 2016, 00:59:59 UTC | added EmbeddingLayer{}; tuned LayerNormalizationLayer{}; added ReLU alias; bug fix: SGD should print the dimensions of the trained nodes; seqcla changed to model() pattern | 08 August 2016, 00:59:59 UTC |
c2f5a29 | Amit Agarwal | 07 August 2016, 23:36:48 UTC | CNTK v2 library: Minor fix in a comment | 07 August 2016, 23:36:48 UTC |
ca2f33f | Amit Agarwal | 07 August 2016, 23:33:51 UTC | Merge branch 'master' of https://github.com/Microsoft/CNTK into amitaga/cntkv2Library | 07 August 2016, 23:33:51 UTC |
2ceb7d8 | Frank Seide | 07 August 2016, 04:32:30 UTC | deprecated C++ implementation of PerDimMeanVarNormalizationNode in favor of a BrainScript one-liner. One more C++ node down; first cut of a layer library (partially tested) | 07 August 2016, 04:32:30 UTC |
9cbb456 | Frank Seide | 07 August 2016, 01:01:54 UTC | moved 'standard functions' section in CNTK.core.bs further to the back | 07 August 2016, 01:01:54 UTC |
f4c4f63 | Frank Seide | 05 August 2016, 21:22:47 UTC | bug fix: any setting of LearnableParameter size must lazy-init | 05 August 2016, 21:22:47 UTC |
db5fff2 | Frank Seide | 05 August 2016, 21:11:38 UTC | merged from master | 05 August 2016, 21:11:38 UTC |
98749f9 | Frank Seide | 05 August 2016, 20:41:10 UTC | disabled call to LazyInitParameters() in Validate() because it does nothing | 05 August 2016, 20:41:10 UTC |
76f0d64 | Frank Seide | 05 August 2016, 20:36:38 UTC | bug fix: SVD must use InitLearnableParameters() first, to avoid its result to be initializer over later | 05 August 2016, 20:36:38 UTC |
b6cfca3 | Zhou Wang | 05 August 2016, 14:58:10 UTC | fix missing path for nvidia-ml in brainscripttests | 05 August 2016, 14:58:10 UTC |
4e17fd5 | Thilo Will | 05 August 2016, 12:40:51 UTC | Fixing typo in reduction test and reformatting | 05 August 2016, 12:40:51 UTC |
253e65b | Thilo Will | 05 August 2016, 12:08:54 UTC | ReduceLogSum: beautifications | 05 August 2016, 12:08:54 UTC |
6947079 | Thilo Will | 05 August 2016, 10:33:39 UTC | merged from master | 05 August 2016, 10:33:39 UTC |
e6ca0c0 | Frank Seide | 05 August 2016, 00:57:23 UTC | bug fix: LearnableParameter constructor must use correct iniString | 05 August 2016, 00:57:23 UTC |
44bb46e | Frank Seide | 04 August 2016, 23:58:03 UTC | cleaned up logging of lazy Parameter initialization; removed compat hack; fix in w.r.t. a printf format; | 04 August 2016, 23:58:03 UTC |
6ebdc6b | Frank Seide | 04 August 2016, 23:29:12 UTC | bug fix: PostInitParameters() should not shadow randomSeed with a local variable | 04 August 2016, 23:29:12 UTC |
194f80c | Frank Seide | 04 August 2016, 22:15:33 UTC | BrainScript now optimally allows { } instead of [ ] for record definition, to make it look more like other languages; replaced experimental { } array syntax with F#-style syntax, to avoid syntax clash; stratified logging of lazy Parameter initialization; BatchNormalizationNode changed to -NonLooping since it does not support recurrence | 04 August 2016, 22:15:33 UTC |
683896a | Dong Yu | 04 August 2016, 18:36:25 UTC | Merge branch 'master' of https://github.com/Microsoft/CNTK into dongyu/stoplr | 04 August 2016, 18:36:25 UTC |
588dae4 | Frank Seide | 04 August 2016, 17:08:35 UTC | more test code for lazy init | 04 August 2016, 17:08:35 UTC |
76e2f61 | Frank Seide | 04 August 2016, 16:51:20 UTC | more test code for lazy init | 04 August 2016, 16:51:20 UTC |
dc262af | Eldar Akchurin | 04 August 2016, 10:36:50 UTC | Addressing CR | 04 August 2016, 10:36:50 UTC |
ae2f3b7 | Eldar Akchurin | 03 August 2016, 11:34:20 UTC | Addressing CR comments | 04 August 2016, 10:29:42 UTC |
38ebf57 | Eldar Akchurin | 03 August 2016, 11:21:52 UTC | Fixing potential race between loading/unloading of chunks | 04 August 2016, 10:29:42 UTC |
73f607a | Eldar Akchurin | 30 June 2016, 15:15:34 UTC | Implementing io prefetch | 04 August 2016, 10:29:42 UTC |
b11593f | Dong Yu | 03 August 2016, 23:29:43 UTC | allow nodes' output value matrix to be shared during training time when gradient is not needed for that node. | 04 August 2016, 00:58:28 UTC |
73109af | Frank Seide | 03 August 2016, 19:42:10 UTC | (minor fix) | 03 August 2016, 19:42:10 UTC |
008c357 | Frank Seide | 03 August 2016, 18:51:16 UTC | added test code for lazy init | 03 August 2016, 18:51:16 UTC |
1a4f4a4 | Zhou Wang | 03 August 2016, 18:40:17 UTC | fix typo | 03 August 2016, 18:40:17 UTC |
2ead1b4 | Frank Seide | 03 August 2016, 18:10:17 UTC | bug fix: InitLearnableParameters() cannot be a <float>/<double> template that is selected based on an argument that can be passed as 0 (which would select <int>) | 03 August 2016, 18:10:17 UTC |
f4cb19b | Zhou Wang | 03 August 2016, 14:01:45 UTC | disable convertion of LF to CRLF for test data file, and update the file offset value in the control file. change the file name to take .gitattributes in effect without cleanup workspaces in Jenkins | 03 August 2016, 17:55:03 UTC |
cf0022d | Zhou Wang | 03 August 2016, 06:51:05 UTC | change the output format of exponent number from e[+-][0-9][0-9][0-9] to e[+-][0-9][0-9]; adapt base line files of reader unit tests Update format e[+-][0-9][0-9][0-9] to e[+-][0-9][0-9] for Linux use two-digit for exponenet number remove files that slipped in by accident add #ifdef and comments change exponent output format change exponent output format | 03 August 2016, 17:52:53 UTC |
bdad298 | Project Philly | 03 August 2016, 17:23:02 UTC | Integrate eldak/fixingBSTestsDependencies into master | 03 August 2016, 17:23:02 UTC |
4542b48 | Eldar Akchurin | 03 August 2016, 09:04:05 UTC | Adding the unit test | 03 August 2016, 14:44:16 UTC |
b8a7895 | Eldar Akchurin | 02 August 2016, 16:54:23 UTC | Fixing crash | 03 August 2016, 14:44:16 UTC |
03539cd | Eldar Akchurin | 03 August 2016, 12:39:11 UTC | Fixing dependencies | 03 August 2016, 12:39:11 UTC |
63a5580 | Ivan Rodriguez | 02 August 2016, 07:18:08 UTC | Update linux build | 03 August 2016, 10:56:19 UTC |
2546b72 | Ivan Rodriguez | 01 August 2016, 14:35:04 UTC | Remove old BrainScript tests | 03 August 2016, 10:56:19 UTC |
74fffe3 | Ivan Rodriguez | 01 August 2016, 12:41:12 UTC | Use input and output files for each test | 03 August 2016, 10:56:19 UTC |
8003758 | Ivan Rodriguez | 26 July 2016, 12:00:19 UTC | Clean solution | 03 August 2016, 10:56:19 UTC |
cf80994 | Ivan Rodriguez | 26 July 2016, 09:29:10 UTC | Parser tests added | 03 August 2016, 10:56:19 UTC |
d9a348f | Ivan Rodriguez | 20 July 2016, 09:13:27 UTC | Start for BrainScriptTests module | 03 August 2016, 10:56:19 UTC |
621561a | Ivan Rodriguez | 20 June 2016, 09:51:37 UTC | check of original tests | 03 August 2016, 10:56:19 UTC |
f951e9a | Thilo Will | 03 August 2016, 09:35:59 UTC | Fixing typo in comment | 03 August 2016, 09:35:59 UTC |
0d3b9e5 | Thilo Will | 03 August 2016, 08:59:12 UTC | Added comment regarding default axis values in python bindings | 03 August 2016, 08:59:12 UTC |
01915bc | Zhou Wang | 29 July 2016, 16:58:49 UTC | Link libcntkmath to libeval Use GDK_NVML_LIB_PATH Simplify Boost libraries Use ORIGINDIR in rpath of libeval remove GDK_NVML_LIB_PATH from cppevalclient and Add -lcntkmath to libeval.so Use GDK_NVML_LIB_PATH to avoid using system-wide one, Simplify use in libeval.so | 03 August 2016, 06:43:47 UTC |
aaac609 | Frank Seide | 03 August 2016, 05:07:46 UTC | bug fix: LearnableParameter::CopyTo() must exist, now that we have the lazy-init variables | 03 August 2016, 05:07:46 UTC |
c213625 | Frank Seide | 03 August 2016, 04:22:07 UTC | NDL version of lazy Parameter init (for dim inference) | 03 August 2016, 04:22:07 UTC |
e80562f | Frank Seide | 03 August 2016, 02:24:42 UTC | trying a fix to lazy init | 03 August 2016, 02:24:42 UTC |
bc06c3c | Frank Seide | 03 August 2016, 02:13:30 UTC | CNTK BatchNorm engine Backprop() should honor blendFactor | 03 August 2016, 02:13:30 UTC |
838717f | Frank Seide | 02 August 2016, 21:51:56 UTC | bug fix in lazy random init of Parameter: Validate() is called too early, must do it in InferInputDimsFrom() | 02 August 2016, 21:51:56 UTC |
6232550 | Dong Yu | 02 August 2016, 21:11:38 UTC | Merge branch 'master' of https://github.com/Microsoft/CNTK into dongyu/stoplr | 02 August 2016, 21:11:38 UTC |
0e3b5da | Zhou Wang | 02 August 2016, 16:35:08 UTC | include crtdbg.h only on Windows | 02 August 2016, 16:35:08 UTC |
f61cc4e | Thilo Will | 02 August 2016, 15:21:32 UTC | in __init__.py boved ReduceLogSum to bottom. Renamed python operation to reduce_log_sum | 02 August 2016, 15:21:32 UTC |
8db871e | Wolfgang Manousek | 02 August 2016, 15:08:12 UTC | addressed final code review comments | 02 August 2016, 15:09:54 UTC |
e440132 | Wolfgang Manousek | 02 August 2016, 14:44:36 UTC | moved to _CrtSetReportHook2 | 02 August 2016, 15:09:54 UTC |
e7fd489 | Wolfgang Manousek | 02 August 2016, 10:58:43 UTC | use CrtSetReportHook on windows to disable the popup of assert boxes | 02 August 2016, 15:09:54 UTC |
f84fdc2 | Thilo Will | 02 August 2016, 14:57:56 UTC | ReduceLogSum: testcases for backward path. Documentation | 02 August 2016, 14:57:56 UTC |
e2e15e0 | Project Philly | 02 August 2016, 13:53:24 UTC | Integrate mahilleb/AssertRemoval into master | 02 August 2016, 13:53:24 UTC |
fa7befb | Mark A. Hillebrand | 02 August 2016, 13:48:54 UTC | Address CR comment | 02 August 2016, 13:48:54 UTC |
6c7f7c1 | Thilo Will | 02 August 2016, 11:38:47 UTC | Forward path test for ReduceLogSum works | 02 August 2016, 11:38:47 UTC |
cf51d75 | Thilo Will | 02 August 2016, 07:43:30 UTC | Adding code for reduce_logsum test. | 02 August 2016, 07:43:30 UTC |
5a7e77b | thilow | 02 August 2016, 07:32:57 UTC | ElementwiseProductWithExpOffDiff | 02 August 2016, 07:32:57 UTC |
2115ac1 | Dong Yu | 01 August 2016, 22:30:57 UTC | allow to customize minLearningRatePerSample to control early stopping. | 01 August 2016, 22:30:57 UTC |
c81eb6f | Vadim Mazalov | 27 July 2016, 00:01:25 UTC | Add Array struct, quantizer unit tests, minor fixes. | 01 August 2016, 19:33:03 UTC |
7cfc3f3 | Vadim Mazalov | 27 June 2016, 16:23:56 UTC | Remove LearnableParameterQuantized and MEL command to quantize a node | 01 August 2016, 19:33:03 UTC |