tf-onnx introducing many transpose operations about tensorflow-onnx HOT 5 CLOSED

Hi Guenther, checking how we avoid the transpose operations for the cpu tf2onnx graph?

from tensorflow-onnx.

If the model is nhwc we wrap the ops with transpose. There are only 2 optimizations: if the input is a const ant we apply the transpose to the constant and if we find back to back transpose we remove it. Pretty conservative and I'm sure we can do better.
Let me find some time to look at it.

from tensorflow-onnx.

hi I am also seeing that Intel MKL is able to handle NHWC (similar to GPU). This is what I found and posted on the onnx-tf page - onnx/onnx-tensorflow#31 (comment)

from tensorflow-onnx.

I think this only creating issues for tensorflow models ... wish they would have not done nhwc. Adding nhwc to onnx would create similar issues there.
I have some ideas how to get rid of some ... going to get to this soon.

from tensorflow-onnx.

This will help:
#108

from tensorflow-onnx.