The ONNX graph converted from tf.image.crop_and_resize() function will cause a different result by ORT nightly build.

[fatcat-z]

Describe the bug
Install ORT nightly job. Try to execute the example script provided in this issue which is trying to inference an ONNX file generated by converting tf.image.crop_and_resize() op to ONNX via tf2onnx. The result of ORT is different than tensorflow.

Urgency
This is breaking tf2onnx CI pipeline so please fix it soon.

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Linux Ubuntu 20.04
• ONNX Runtime installed from (source or binary): pip install --index-url https://test.pypi.org/simple/ ort-nightly
• ONNX Runtime version: ort-nightly 1.11.0.dev20220228001
• Python version: Python 3.7
• Tensorflow version: 1.13.1

To Reproduce

1. Download the attached zip file crop_and_resize_failure.zip.
2. Unzip it and execute the command "python crop_resize_issue.py" under this folder.

Expected behavior
It should display Pass without any exceptions.

[fatcat-z]

@HectorSVC ,

Actually, this issue itself has no relationship with convertor. The example code in attachment is for tf 1.x version, so the Tensorflow version is 1.13.1.

This issue actually shows that for the same ONNX file, different ORT version will cause different results. Please re-consider the tag and issue.

Thanks.

[pranavsharma]

• can you attach the onnx model?
• what vers of ORT gave different results?
• how much was the difference?

[yufenglee]

It should be caused by #10554. Here is the fix #10824 in review.

[RandySheriffH]

@yufenglee: I have build the fix and tried it out with the script it seems the diffs are still there.
https://aiinfra.visualstudio.com/Lotus/_build/results?buildId=198775&view=results

[RandySheriffH]

@fatcat-z : now with #10917 the diffs are gone.