How to convert onnx(NCHW) to NHWC(without adding transpose) and write back to onnx file about onnx2tf

[PINTO0309]

I don't know what you expect me to do. ONNX works only in NCHW format. It is impossible in principle.

If you can't tolerate only one Traspose, give up. I cannot change the ONNX specifications.

[PINTO0309]

onnx2tf -i aaa.onnx -kt {input_op_name}

onnx2tf -i aaa.onnx -kat {input_op_name}

https://github.com/PINTO0309/onnx2tf?tab=readme-ov-file#cli-parameter

[abhishek27m1992github]

I don't know what you expect me to do. ONNX works only in NCHW format. It is impossible in principle.

If you can't tolerate only one Transpose, give up. I cannot change the ONNX specifications.

So sorry if i'm unable to make you understand. I will try my best to explain.

First i'm not using onnxruntime. we have our own runtime SDK & NN hardware which accepts NHWC format.

if i convert pytorch models to tflite via onnx ---> so many transpose layer gets added. so I used your tool openvino2tflite which nicely removes transpose layers.

Now i want to add support .onnx to my SDK. My assumption is if we somehow only change memory format to NHWC - aka how tensor laid in memory i.e. weights & biases etc... in onnx file then i feed same .onnx to my SDK & H/W it should work. but i know it will not work in onnx runtime. Please correct me if i'm wrong.

So that's what i wanted to do with your tool.