Support tf ResizeArea op.

[fatcat-z]

Add support for ResizeArea op, and add tests.
Fix #2239

[ZachZhao96]

Thank you very much. So, does the current version on the PyPI support the ResizeArea after this commit got merged? or do I expect to see the area method support in the next release?

[fatcat-z]

Thank you very much. So, does the current version on the PyPI support the ResizeArea after this commit got merged? or do I expect to see the area method support in the next release?

This commit has been merged and will be available in next release which is not in a plan yet. If you want this function, you can install from source for your local usage.

[ZachZhao96]

Thank you very much. So, does the current version on the PyPI support the ResizeArea after this commit got merged? or do I expect to see the area method support in the next release?

This commit has been merged and will be available in next release which is not in a plan yet. If you want this function, you can install from source for your local usage.

I successfully converted my model to onnx using the new version, but when using onnxruntime for inference, the prediction result is way off compared to the result from the regular keras model.predict. I guess I will have to wait for onnx or onnxruntime to have the ResizeArea method in Resize operation as well. Look forward to the next release. Really appreciate your time.

[ZachZhao96]

@fatcat-z Hi, has resize_area method been included in the latest ONNX and supported by tf2onnx yet? I couldn't find anything regarding the resize_area in either onnx or tf2onnx documentation except for the test_backend file you committed last time.

I dont see any error message when i convert my keras model (with resize_area method) to onnx using tf2onnx.convert.from_keras(model,opset=13), however it seems that the "resize_area" method is recognized as "bilinear" because the inference result from the converted ONNX model is different from the original keras model.

[fatcat-z]

No, resize_area is not an ONNX op.

[ZachZhao96]

Thanks, I also tried resized_spectrogram = tf.image.resize(spectrogram, [256, 256],method = "bilinear", antialias = True), then I got this error:

I wonder if the antialias supported in Resize Op is the same as the tf.image.resize(..., antialias = True).