Граф коммитов

16078 Коммитов

Автор SHA1 Сообщение Дата
Arijit Das 586953e286
Fix typo "evalutaion" => "evaluation" 2018-10-10 18:02:23 +05:30
Spandan Tiwari 149d87bad3 Adding ONNX export support for OneHotOp. 2018-10-05 14:06:56 -07:00
kokuB 1aefedb8af
Fix percent typo 2018-10-03 15:24:06 +05:30
kokuB dde1869563
fix typo 2018-10-03 12:32:05 +05:30
Bowen Bao 6063866a4f Merge branch 'bowbao/pooling_pad_offset' 2018-10-02 03:41:06 +00:00
Bowen Bao bf37aadc53 Fix pad offset computation for pooling
* Compute keyInterior according to the updated algorithm for computing
cell offset key.
* Update unittest of avg_pooling/max_pooling for cases that requires
auto_padding = True. Previous test cases cover only those that do not
need padding.
2018-10-01 17:21:30 -07:00
TJ 7c838d3b6c fix csharp examples and references 2018-10-01 15:45:13 -07:00
Bowen Bao fcdeef63d0 Support crop_manual export & import. 2018-09-29 13:46:17 -07:00
Bowen Bao a36fae88bb Support logPlus(log_add_exp) export to ONNX
* ONNX supports similar op ReduceLogSumExp. Conversions are added when
exporting.
* Refactored CNTKToONNXHelper::BroadcastInputsIfNeeded to support more
generalized cases.
2018-09-28 15:59:55 -07:00
Spandan Tiwari c2072cc4ab Add support for ONNX export of StraightThrough op. 2018-09-27 10:25:11 -07:00
Spandan Tiwari 1aab76af99 Updating ONNX submodule hash to include defs for ConstantLike and EyeLike ops. 2018-09-26 18:01:34 -07:00
Peyman Manikashani ce503f8dd7 pooling export fix for backward compatibility 2018-09-25 17:09:39 -07:00
Ke Deng 9165fd06f8 Merge branch 'kedeng/fixCrash' 2018-09-25 00:26:34 +00:00
liqfu 58f810fed0 update with ONNX1.3 and latest onnxruntime 2018-09-22 09:53:27 -07:00
KeDengMS 1489de8de8 Fix a crash in transpose_times simplification to element times 2018-09-21 22:33:41 -07:00
Bowen Bao 93cc680d56 Update reshapeing_test to accomodate python 2.7 2018-09-21 10:24:57 -07:00
Bowen Bao d626c2a466 Merge branch 'bowbao/gather_backward' 2018-09-21 05:02:57 +00:00
Bowen Bao da6b0bc71f GatherNode backward: add check for no dynamic axis
Previously, to resolve issue of gather producing incorrect gradient
values, validity mask check was added to ensure we don't count non-valid
cells as 0.
However, this check is needed only for input that has dynamic axis, i.e.
inputs that have MBLayout.
2018-09-20 14:54:39 -07:00
Peyman Manikashani b2c28cc98e updating iteration documentation file 2018-09-20 14:39:58 -07:00
Bowen Bao 0a3eb3b813 Update onnx_model_test with tests on cntk pretrained models 2018-09-19 11:27:55 -07:00
Liqun Fu 6f09c398b9 Merge branch 'release/2.6' 2018-09-18 02:33:30 +00:00
liqfu 4ed1896332 set public_build to "no"/false 2018-09-17 16:10:48 -07:00
liqfu 1be3b64195 update readme for .net support 2018-09-14 18:07:45 -07:00
TJ d355c1c700 Updated current_iteration with .net support 2018-09-14 13:07:24 -07:00
TJ 04caa9deaf Updated current_iteration with .net support 2018-09-14 12:29:51 -07:00
liqfu 7c1b0fadb6 udpate readme with current iteration 2018-09-13 17:01:45 -07:00
Bowen Bao da31ba04fa Update current_iteration.md 2018-09-13 16:57:44 -07:00
Sergii Dymchenko e4d708118d Update current_iteration.md. 2018-09-13 16:57:30 -07:00
liqfu 4f965aaaf7 update version # in cntk_common.cmake 2018-09-13 16:34:30 -07:00
liqfu 82d350d0ac bump up version number 2018-09-13 15:52:47 -07:00
Bowen Bao d264a26034 Update current_iteration.md 2018-09-13 13:44:42 -07:00
Sergii Dymchenko be28e864cc Update current_iteration.md. 2018-09-13 11:18:53 -07:00
Bowen Bao deda94b67b Support pooling(cpu) where kernel center is on pads.
- Previous implementation has the assumption that (0 <= dk < width).
This assumption doesn't stand when lo >(kernel - 1) / 2.
    The updated calculation supports arbitrary lo & hi non-negative
    integer value. The new calculation has dk in range (0, width + hi +
    lo].
- Enables onnx backend test {averagepool_2d_pads, maxpool_2d_pads} to
pass.
2018-09-12 21:37:21 -07:00
Bowen Bao 62e18f4854 Improve clarity in pads calculation for conv/pool
- Refactor function CalcPaddingForSameLowerOrUpperAutoPad in conv/pool import,
  changing parameter "const Variable& input" to "const NDShape& inputWithBatchAxisShape",
  to specify the required shape format as [N x C x H x W].
2018-09-12 21:37:21 -07:00
Sergii Dymchenko 35e370170a Merge branch 'sedymche/onnx-min-max' 2018-09-13 02:42:25 +00:00
liqfu d33b7b44e8 update iteration plan 2018-09-12 17:24:45 -07:00
Peyman Manikashani 6a4ec05b19 Merge branch 'peykash/batchnorm_and_pooling_fixes' 2018-09-13 00:13:46 +00:00
Sergii Dymchenko 61d7dab912 Support more than 2 inputs for ONNX Min/Max import. 2018-09-12 15:12:14 -07:00
Spandan Tiwari 8b48976bed
Adding CNTK 2.6 release work summary to current_iteration.md 2018-09-12 11:12:10 -07:00
Peyman Manikashani b374e149b4 fixes on Batchnorm and Pooling for v1 pretrained models after removal of sequence axis from input 2018-09-12 10:02:52 -07:00
Bowen Bao 5897265366 small patch on conv/pooling export
- when pads are all zero, check if autopad is true.
- when pads are all zero, check if ceilOutDim is true, and extra cells
are needed.
2018-09-11 08:52:22 -07:00
Bowen Bao 61572e89f8 Update onnx_model_test skip list 2018-09-11 08:51:45 -07:00
Bowen Bao 0754b38e34 update onnx_model_test with tests from onnx backend test 2018-09-09 13:53:26 -07:00
Bowen Bao 2f52f2219f update conv/convtranspose/pooling import.
pad values are explicitly computed based on ONNX spec equations during import in the following cases:
- case 1: when auto_pad is SAME_UPPER | SAME_LOWER for convolution, convolution transpose and pooling.
- case 2: when output_shape is explicitly set for convolution transpose.
	  note: output_shape in ONNX spec can have the two below format:
	  	1. [X1 * X2 * ... * Xn]
		2. [N * O * X1 * X2 * ... * Xn]
2018-09-09 13:41:04 -07:00
liqfu d877233979 Make broadcast ops compitable between CNTK and ONNX,
Enable ONNX export/import for optimizedRNN op,
More ONNX support for Sequence ops
2018-09-09 08:59:33 -07:00
Bowen Bao fcf9f48895 Overhaul conv/convTrans/pooling pads value export
- Update exporting of conv/pooling to always export pad values.
- Enable correct exporting of multiple pretrained models (ResNet50/ResNet101/ResNet152_ImageNet_Caffe, etc).
- Overhaul convtranspose pads exporting
- Support conv weight export with omitted out channel axis (LRN).
- Add tests in onnx_op_test to cover the above changes
2018-09-06 11:46:14 -07:00
Bowen Bao e3a1acfdf0 Resolve dependencies and build issues
-Temporary add importorskip around import onnx
-bump up .yml matplotlib version
2018-09-05 15:02:23 -07:00
Bowen Bao dc5e482d54 fix onnx average pooling export.
- this fix solves the issue that ceilOutDim == true will enforce exporting auto_pad as true, even if autoPadding is explicitly set to false.
2018-08-30 18:19:04 -07:00
Bowen Bao 77a8c4992f Temporarily skip onnx_model_test if import onnx fail 2018-08-30 10:52:53 -07:00
Liqun Fu 94a43edd24 Merge branch 'liqun/liqun/RNN2.6.Stage' 2018-08-30 07:40:20 +00:00