[ONNX] Support ConvTranspose + auto_pad

[stakemura]

TVM Relay (latest master branch) couldn't compile ONNX including ConvTranspose + auto_pad.

Here is the code and the attached
ZIP-compressed ONNX file updated on Jul 25
to reproduce the error.

import tvm
import tvm.relay as relay
import onnx

onnx_model = onnx.load('upconvert.onnx')
 
shape_dict = {'input__0': (1, 3, 360, 640)}
mod, params = relay.frontend.from_onnx(onnx_model, shape_dict)

This issue is critical for us.
Would you like to support this attribute?

[tqchen]

It would be great if you can look a bit into what is going on, and the community can provide help on possibly actions

[vinx13]

@stakemura this means we need to infer padding in case auto_pad is set in https://github.com/dmlc/tvm/blob/e7fb2d4d05858ad64f27cc06b944d5abcdaaef3e/python/tvm/relay/frontend/onnx.py#L209

[stakemura]

@vinx13 Thank you for the supplementary explanation!
@abuccts I deeply appreciate your help. I would like to try your patch.

[abuccts]

Hi @stakemura, could you help to verify the patch? @vinx13 could you please review the PR, thx!

[stakemura]

@abuccts I'm sorry for the late reply. I confirmed #3645 patch can load the attached onnx model successfully.

[cchung100m]

Hi @zhiics

This PR should be closed by #4563, too. I would appreciate that if you can help to manage it, thanks :)

[zhiics]

@cchung100m Thanks for resolving the issue.