[ONNX] Fix interpreting auto_pad parameters in ConvTranspose operator

[padreofthegame]

Fix in interpreting auto_pad parameters SAME_UPPER and SAME_LOWER in ConvTranspose version 1 and 11 operator according to documentation on link.

Tests that were failing are also added in test_forward.py.

Also, enable unsupported test for ONNX: test_convtranspose_autopad_same.

[padreofthegame]

@tvm-bot rerun

[padreofthegame]

Hello,

looks like that the problem with this PRs failure is due to the different implementations of convTranspose operator in different versions of onnxruntime. Precissely, onnxruntime 1.12 and 1.15 gives different results when the parameters SAME_UPPER and SAME_LOWER are used. On my local machine, this tests passes when onnxruntime 1.15 is used, and fail with 1.12 (probably due to wrong implementation), although the operator is implemented according to ONNX documentation.

What could be the next step for this PR? What is the common practice here:
- to skip this test, as it was done with test_convtranspose_autopad_same?
- to mark this test as flaky?
- to add part of the code to check onnxruntime version before test running?

Sorry if I tagged someone not related to the topic.

@masahi @Hzfengsy @jikechao

[padreofthegame]

Hello,

looks like this PR is staying opened this way for quite some time now. Any feedback related to this PR?

[padreofthegame]

Hello,

Any updates on this one? Can we get this merged as it has been approved for quite some time now, and is passing all the tests?

[yongwww]

@padreofthegame it was merged! Thanks for the fix!