Граф коммитов

16055 Коммитов

Автор SHA1 Сообщение Дата
Yang Chen f310127f72 enable loading onnx from buffer 2019-07-16 10:23:24 -07:00
Yang Chen 9fe872b413 Added a C API to load a model from memory
Added a new C API CNTK_LoadModel_FromArray that loads a model
from memory
2019-06-20 16:26:41 -07:00
Yang Chen 238eb2d055 pick up fix from liqun/scan14
cherry-pick commits:

ad1709c: fix reduce op keepdim setting
1fe0bf1: reduction leep dim edge cases
03a8c41: default keep dim is true
2019-06-19 14:10:24 -07:00
Yang Chen da9f2e5ac6 manually pick Liqun's fix to past values
commit 248ff7 from liqun/scan13
2019-05-11 23:00:41 -07:00
Yang Chen ab62f886ca fixed build for py36
manually picked up commit 0e172d from master
2019-04-30 13:02:45 -07:00
Yang Chen eb9786aa63 manually picked up commit from liqun/scan12
onnx export: fix cases when scan input needs to be broadcasted

commit: 22a86bf3
2019-04-30 00:53:03 -07:00
KeDengMS 2905ba6613 Manually pick up changes from Liqun/Scan10 2019-03-26 11:28:49 -07:00
Yang Chen 351fddf54e Fixed an issue for generating name pairs 2019-02-27 13:07:43 -08:00
Yang Chen c6dccd344d bump default protobuf version to 3.6.1 2019-02-27 11:59:49 -08:00
Yang Chen 627df8723d added a mapping of onnx output to cntk output into the graph's description 2019-02-27 11:55:56 -08:00
Yang Chen e919013090 handle more kinds of dictionary values 2019-02-27 01:16:19 -08:00
Yang Chen b3778c9309 Turn custom attributes of a node into the converted node's description
Also disabled an asserts that seemed to be too strong.
2019-02-26 00:35:21 -08:00
Thiago Crepaldi 06469f7798 Merge branch 'pull/3558' 2019-02-25 18:46:26 +00:00
liqfu a5fb3d6f8d add more onnx rnn model tests, enable optimizedRNN op test 2019-02-19 12:48:41 -08:00
Bowen Bao 00b80377fe Update onnx submodule, support ConstantOfShape & OneHot
* Support SequenceIs[First,Last] with ConstantOfShape
* Update bypass load test in verify\_one\_input & add test for one hot
op
* Update export for one hot op. Migrate from exporting
onnx.ml.OneHotEncoder to onnx.OneHot.

Op fixes
* Fix topk onnx\_op\_test
* Support MVN export using ONNX function
* Fix LayerNormalization
* Skip tests for sequence slice float16: not supported in cntk
* Support gather export & import with float16
    - Fix cntk gather issue with float16 inputs.
    - Support exporting constant float16 tensor.
    - Support importing int32 indices input for gather.
* Enable more passed op tests
2019-02-15 15:12:10 -08:00
Liqun Fu aa828194db Merge branch 'liqun/remove-crosstalkcaffe-test' 2019-02-14 21:07:34 +00:00
liqun fu 3bef0f4214 protobuf 3.6.0 to 3.6.1 2019-02-13 11:08:06 -08:00
Amit Agarwal 9ce6c1c2f9
Merge pull request #3550 from Microsoft/yaronwe/hide-protobuf
hide protobuf library symbols
2019-02-08 08:37:36 -08:00
liqfu 2ccdff45e3 enable onnx rnn test, name input with proper name 2019-02-03 21:18:50 -08:00
Liqun Fu 979c8bfe2a Merge branch 'liqun/liqun/scan6Stage2' 2019-01-31 04:27:20 +00:00
Thiago Crepaldi 9093a4def1 Cache CNTK dependencies at CNTK Azure Blob 2019-01-29 19:47:10 -08:00
liqfu 6b2323c4e4 onnx support of sequential convolution 2019-01-28 14:48:13 -08:00
Peyman Manikashani 4ef44170ab Adding sequence models to CI and skipping failing ones 2019-01-24 15:01:14 -08:00
Peyman Manikashani fc839bc371 Moving CI test dependencies to python environments yml files 2019-01-22 16:24:29 -08:00
Aaron Schweiger d67a5ae6cf Free nnzPerRowOrCol at end of SetValue 2019-01-16 17:00:33 -05:00
Vadim Mazalov fea6e9dd47 Merge branch 'vadimma/bmlfex' 2019-01-12 00:01:46 +00:00
Vadim Mazalov 62fa8182d7 Enable frame mode for the binary MLF reader 2019-01-10 18:02:34 -08:00
Liqun Fu 0be4bcb949 Merge branch 'liqun/scan5Stage' 2019-01-09 02:39:18 +00:00
liqfu fe939aa3cb support ONNX scan(9) 2019-01-08 12:16:32 -08:00
Bowen Bao a5fcc80d91 Resolve Linux binary drop & windows oobe test conda dependency 2019-01-07 15:07:04 -08:00
yaron weinsberg 2274e024ed hide protobuf library symbols 2019-01-06 10:04:34 +00:00
Mickaël Schoentgen b732aeb0fd Fix several DeprecationWarning: invalid escape sequence
Signed-off-by: Mickaël Schoentgen <contact@tiger-222.fr>
2019-01-05 22:36:45 +01:00
Vadim Mazalov 322293c8fd Merge branch 'vadimma/adambs' 2019-01-04 00:21:27 +00:00
liqfu 6c56fdd2f9 Revert "refactor"
This reverts commit 47345c6c54.
2019-01-03 15:45:28 -08:00
liqfu 47345c6c54 refactor 2019-01-03 15:26:03 -08:00
Vadim Mazalov 05435d0db2 Expose Adam in BS 2019-01-03 12:58:17 -08:00
Bowen Bao 08130dbb92 update linux binary drop cudnn path 2019-01-03 10:57:44 -08:00
Bowen Bao e2d79d7da0 Submodule onnxruntime, and remove previous drop.
* A few patches are required to build cntk_uwp.
* Use proto from onnxruntime/protobuf instead of from onnx.
* TODO: Some issues with onnx_op_test RNN and OptimizedRNNStack from shape inference.
2019-01-02 17:09:08 -08:00
Bowen Bao 254a3362f5 Bump cuda version in nuget spec 2018-12-21 10:40:18 -08:00
Bowen Bao bc708633c8 Add Maxunpooling export 2018-12-14 11:24:26 -08:00
Bowen Bao d087acc279 Fix typo & update .yaml files
* pyqt is required by matplotlib and various other libraries. In our
.yaml files we default to install the latest version of pyqt. However
the current latest version (5.9.2) is exhibiting some problems. Thus we
downgrade to the stable version 5.6.0.
2018-12-13 10:37:03 -08:00
Deyu Fu ef2f039b53 fix break for cub 1.8 API change 2018-12-12 16:14:15 -08:00
Bowen Bao f1781446d1 Support CUDA 10
* Move to support CUDA 10, cudnn 7.3, cub 1.8.
* Fixed a bug related to "pointer to pin pointer is disallowed" #3063,
which is exposed in newer version vctools.
* Added workaround for a potential vs2017 15.9 bug with cntk Debug
version.
2018-12-12 16:10:31 -08:00
liqfu 93e10096cb comment debug code 2018-12-11 19:17:16 -08:00
liqfu de15bda40d passed prod models and unfold test. handle delay op. 2018-12-11 17:49:23 -08:00
Spandan Tiwari 22e869ec42 Add ONNX support for zeros_like, ones_like, and eye_like. 2018-12-05 13:51:43 -08:00
Peyman Manikashani 5fc6c2a26a sequence_isfirst_islast test fix for python 2.7 2018-11-30 11:22:25 -08:00
Peyman Manikashani 75b0141b48 Adding support for exporting CNTK's Sequence::IsFirst and Sequence::IsLast nodes 2018-11-29 15:57:43 -08:00
Bowen Bao d1d113322c ConvTranspose asymmetric padding
* Temporarily reverse the extra padding location in case of
SAME_UPPER vs SAME_LOWER for convTranspose to match with onnxruntime.
* In case of importing asymmetric padding convTranspose, use
symmetric pads by alter the output_shape and pads, and attach a slice
node afterwards to enable cudnn.
* Fix a bug in slice/squeeze attribute axes export.
2018-11-27 09:33:15 -08:00
Emad Barsoum 196df1a143 Update image layout in readme. 2018-11-26 12:49:56 -08:00