tf2onnx-1.12.0

Summary

• added support for python 3.10, onnx 1.12, onnxruntime 1.12, tensorflow 2.9
• added --outputs_as_nchw option to transpose output to from nhwc to nchw
• enhanced optimizer and exception handing
• improved keras pattern and ci
• misc fixes

What's Changed

• Transpose optimization for Softmax and LogSoftmax (fixes #1716) by @janbernloehr in #1964
• The from_tflite() function should accept None as default value of input_names and output_names. by @fatcat-z in #1967
• Change Loop op with maximum iterations input M equals to empty string by @hwangdeyu in #1971
• Update README.md by @andife in #1976
• Update CONTRIBUTING.md by @andife in #1978
• Update Keras related tests to support latest TF version. by @fatcat-z in #1980
• Fix a test issue in keras2onnx_application_tests.yml. by @fatcat-z in #1982
• L2_NORMALIZATION support for tflite by @shesung in #1989
• Replace deprecated np.object with object by @vvolhejn in #1990
• Implement --outputs_as_nchw option to transpose output to from nhwc to nchw by @hwangdeyu in #1979
• Fix transpose split optimize attr when opset >=13 by @hwangdeyu in #1996
• Skip existing const node in _parse_graph_input by @q-ycong-p in #2000
• Add handling of HardSigmoid recurrent activation for Keras LSTM by @q-ycong-p in #2001
• Add more tests for tf 2.9.x into CI pipelines. by @fatcat-z in #2009
• Fix problem with adding more than one tf.newaxis at the same time by @southfreebird in #2007
• Improve ZerosLike implementation and optimize for opset >= 9 by @hwangdeyu in #2003
• Add newly required dependencies for latest ORT version. by @fatcat-z in #2012
• Add support for Python310 and ORT 1.12 by @hwangdeyu in #1975
• Increment main to 1.12, TF to 2.9 by @hwangdeyu in #2015

A huge thank you to our community for this release!

janbernloehr, andife , shesung, vvolhejn, southfreebird

Full Changelog: v1.11.1...v1.12.0