Fix SAME_UPPER/SAME_LOWER (auto_pad attribute) in ConvTranspose

[jcwchen]

Description:
To sync the definition of SAME_UPPER/SAME_LOWER among all operators and make it same as ONNX definition, switch the logic of SAME_UPPER and SAME_LOWER in ConvTranspose.

Definition of SAME_UPPER and SAME_LOWER should be as follows:

if auto_pads == 'SAME_UPPER':
  pad_head = paddings / 2  # smaller one 
  pad_tail = paddings - paddings / 2  # larger one
elif auto_pads == 'SAME_LOWER':
  pad_head = paddings - paddings / 2  # larger one
  pad_tail = paddings / 2  # smaller one

Motivation and Context
The auto_pad attribute, SAME_UPPER and SAME_LOWER of ConvTranspose is different from other operators' (pool and conv related operators) auto_pad attribute. The behavior of same attribute should be the same among all operators. Also, it does not meet the definition in ONNX.

• SAME_UPPER and SAME_LOWER in other operators
  

  onnxruntime/onnxruntime/core/providers/cpu/nn/pool_attributes.h

  Line 149 in c20fcf2

  *pad_head = pad_needed / 2;

• ConvTranspose definition in ONNX:
  https://github.com/onnx/onnx/blob/b2ed660d0a065b8346816f2c3a95d79ca79b88c9/onnx/defs/nn/defs.cc#L1222
• A related PR in ONNX (There is a contradiction of this in ONNX):
  Update spec for Convtranspose to make it sync onnx/onnx#3019

cc @askhade

[hariharans29]

#4271 - is it good to include that as well or is that a bit out of sync ?

[jcwchen]

#4271 - is it good to include that as well or is that a bit out of sync ?

@hariharans29 Thank you for bringing this up. I have seen that PR before.
The current modification for this PR is breaking something in CIs. Bringing more modification might complicate debugging. I am not sure what is that PR for, but I think the purpose should be quite different from this one.

BTW, is there any reason why that PR has not been merged? Did it pass all the CIs?

[hariharans29]

#4271 - is it good to include that as well or is that a bit out of sync ?

@hariharans29 Thank you for bringing this up. I have seen that PR before.
The current modification for this PR is breaking something in CIs. Bringing more modification might complicate debugging. I am not sure what is that PR for, but I think the purpose should be quite different from this one.

Thanks Jacky. I bought it up because that was also trying to address some aspect of the auto_pad attribute for the ConvTranspose op.

BTW, is there any reason why that PR has not been merged? Did it pass all the CIs?

I couldn't find a reviewer and it sort of slipped through. Yes, I think it passed all CIs. If you could review it and provide feedback, that will be great.

[hariharans29]

Could you please share some context on this toggle ? - As far as i can see, the way it currently is in master seems all right as per the spec - https://github.com/onnx/onnx/blob/master/docs/Operators.md#ConvTranspose.

I think this change is causing tests to break.

[jcwchen]

Here is more context: onnx/onnx#3019
Actually the document in ONNX needs to be updated as well. I think that's the reason why ORT here has incorrect behavior in the first place...
I remember there is one deprecated model test failed. Let me rerun CIs. Thank you.

[jcwchen]

/azp run Windows GPU CI Pipeline, Linux CPU CI Pipeline, Linux CPU Minimal Build E2E CI Pipeline, Linux CPU x64 NoContribops CI Pipeline, Linux GPU CI Pipeline, Linux GPU TensorRT CI Pipeline, Linux OpenVINO CI Pipeline, MacOS CI Pipeline, MacOS NoContribops CI Pipeline, Windows CPU CI Pipeline

[azure-pipelines]

Azure Pipelines successfully started running 4 pipeline(s).

[jcwchen]

/azp run Linux CPU CI Pipeline

[jcwchen]

I have prepared the updated cntk_simple_seg by the corrected ORT. I plan to filter it out in this PR. Then after merge I can update this model directly and propose another PR to add it back. In that case it won't cause other PRs to fail.