Fix default value of RoiAlign's attribute coordinate_transformation_mode.

[i4oolish]

The default semenatics of RoiAlign's attrtibute coordinate_transformation_mode in opset16 is not same with opset10.

For opset10, other framework like pytorch will manual insert an operation to process half_pixel, so it treat coordinate_transformation_mode as output_half_pixel, while in opset16, the default value for RoiAlign's attrtibute coordinate_transformation_mode in Operators.md is half_pixel. And at the same time, onnxruntime treat this attribute as no half_pixel in its cpu providers, below is the link.

https://github.com/microsoft/onnxruntime/blob/main/onnxruntime/core/providers/cpu/object_detection/roialign.h

[daquexian]

Hi @i4oolish , thanks for your contribution. However, I'm afraid I do not understand what this PR does. It is completely ok that the behavior of opset16 is not the same with that of opset10. Could you please elaborate a bit more on this PR?

[gramalingam]

Hi @i4oolish I have the same question as @daquexian. It is unclear what the issue is. Is there a bug in the pytorch converter? Or, a bug in the onnxruntime kernel? Please note that the version-converter will do the right thing for this change in spec (see this)

[justinchuby]

Closed as the default was by design