tensorflow-onnx

Summary

This is a minor bug fix release on top of tf2onnx-v1.6.2.

Changes since v1.6.2

• Implemented MaxPool3D and AvgPool3D 1020 (TomWildenhain-Microsoft)
• Remove unnecessary deepcopy when accessing output from class Node 1019 (xadupre)
• Fixes #1011, add support for atan2 1016 (xadupre)
• Fixed LSTM conversion for TF2 1012 (TomWildenhain-Microsoft)
• Fixes #500, replaces operator FakeQuantWithMinMaxVars 1009 (xadupre)
• Adjust spatial shape to meet rank requirements for Conv*D 1005 (jignparm)
• Fix transpose optimizer when handling "if" operator subgraphs 1004 (jignparm)
• Added an import onnx statement to init.py to fix MacOS bug 1003 (TomWildenhain-Microsoft)
• Make 'serve' tag default again. 1000 (jignparm)
• Fixed dropout rewriter to properly read ratio 995 (TomWildenhain-Microsoft)
• Added 'Toutput_types' to ignored_attr 989 (TomWildenhain-Microsoft)
• Added 'Toutput_types' to ignored_attr 988 (TomWildenhain-Microsoft)
• Add support for TF2.x saved_models from TFHub, as well as --tag & -concrete_function cmd line parameters 984 (jignparm)
• Enhance Transpose optimizer "Mul" handler 978 (jignparm)
• handle 2 switch output consumers 975 (guschmue)

A huge thank you to our contributors for this release!

Summary

This is a minor bug fix release on top of tf2onnx-v1.6.1.

Changes since v1.6.1

• Unpack operator: fix incorrect shape #974
• Fix graph deepcopy issue #972
• Fix issue with shape of LSTM node #970
• UnifyConst #969
• support QueueDequeueManyV2 #968
• Set output shape of ConvTranspose (Conv2DBackpropInput) correctly #966
• update the out-dated shape in _handle_node_having_branches() #963

A huge thank you to our contributors for this release!

Buddha Puneeth Nandanoor, daquexian

Summary

• support for tf-1.15
• experimental support for tf-2.1 and tf-2.2
  (there is a remaining issues with lstm support for tf-2.x)
• support for opset 12

Changes since v1.5.6

• add support for tf.math.is_finite #936
• tutorial how to convert efficientdet to onnx #937
• enable all cond ut for tf-2.x #962
• dims can be a list #956
• Enable tf22 ci #955
• MatrixDiagV1&V2&V3&MatrixSetDiagV3 #935
• MatrixDiagPartV3: Change consts to dynamic ops #948
• Add new pattern for RandomStandardNormal op in TF2 #949
• Update version support for opset 12 operators #947
• matrixdiagpartv3 #942
• fix transpose optimizer for slice op #934
• Add half pixel transformation to resize bilinear op #932
• Add stacked LSTM support #925
• Activate opset12 tests #923
• Multiple fixes for Bert Model (fine-tuned) #929
• Add 2 more handlers for Tranpose: Exp and Log #928
• Support for QuantizeAndDequantize operation #919
• Ensure scalar values only in MatrixDiagPart->Range() function #924
• Fix UnicodeDecode error #922
• Ignore shape inference warnings for FusedBatchNormV3:5 #916
• Fix LSTM pattern matching for version between 1.15.0 and 2.x. #913
• handle softplus in transpose optimizer, needed for mish #908
• fix split in case of splits are negavitve #891
• move some ops to generators.py, new version of supported ops doc #888
• add tf_optimize back to tf2onnx since apps are using it #882
• ReverseV2 - fix shape computations #909
• Fix Transpose + Pad handler, for Keras app MobilenetV2 model #907
• Fix GEMM to check for shape broadcast compatibility of A*B and C #906
• Some ops for opset 12. #903
• opset 12 support #897
• Fix NonMaxSuppression #895
• Support MatrixDiagPart v2 and v3 #890
• Adds Sum(Transpose(x1), Transpose(x2),...) optimizer. #884
• Add Keras apps, ResNet50 model test #880
• Add getting started section to README #877
• resolve warnings and recommendations from LGTM.com #879
• map bfloat to float16 #878
• refactor resize #874
• Fix typo (no function change) #873
• Fix scatternd - inputs bound to different type #870
• Add fusion for Conv2D+ BatchNormalization #871
• dynamic random #869
• zero like bool #866
• use same opset->ir mapping as in r1.5 branch #867

A huge thank you to our contributors

Anders Huss, Buddha Puneeth Nandanoor, Chin Huang, Dheeraj Peri, Emma Yu, Holger Finger, Johannes Dobler, Nikita Pokidyshev, PreethaVeera, Satyajith, Tian Jin, Vincent Delaitre, alexG, anttisaukko, dheerajperi, dirkbrink, mindest, simpeng, ziyuang

This will prevent us from using IR7 when onnx-1.7 is released.