Force output to have chw layout

[devalexqt]

Even if I specify my input is chw, output still has hwc layout.
How to fix this?

python -m tf2onnx.convert  --opset 13 --saved-model  model --output  model.onnx --inputs input:0[1,256,256,3] --inputs-as-nchw input:0 --rename-inputs input

[hwangdeyu]

hi @devalexqt Could you share your save_model then we can take a look if it‘s something about optimizer in tf2onnx?

[devalexqt]

Model can be downloaded here:
https://drive.google.com/file/d/1tdim-rnz6xK2zbZ-_NO9FW6__wF0DRCU/view?usp=sharing
It's just simple autoencoder model with bunch of Conv/Deconv layers, nothing special.

[hwangdeyu]

Thanks @devalexqt for providing it, the --inputs-as-nchw is only for specifying input as nchw to match the ONNX standard.
Now tf2onnx don't support --outputs-as-nchw so far, we have not seen much models where this applies.
We can do it manually or add an new option. Could you tell us which scenario requires it?

[devalexqt]

Thanks.
In my case it's about performance, as you can see provided model is for upscale image by 4x, where output is about 1024x1024 or even 2048x2048. And output transpose layer can take up to 50% of total model execution time, and then I need transpose it again to chw, so I just lost lot of milliseconds.
In production I use converted onnx model on CPU where chw format is up to 2x faster in comparison to hwc.

Please add option to force output to chw.

[andife]

I would also be very interested in this functionaliy, also because of speed issues. My use case would be semantic segmentation.
With for example using the dnn module of opencv with onnx, it is quite nice using selecting the output layer for different classes when I have the nchw format. For the current hwc format I've to manually reorder the output, which is quite slow.

[hwangdeyu]

I would also be very interested in this functionaliy, also because of speed issues. My use case would be semantic segmentation. With for example using the dnn module of opencv with onnx, it is quite nice using selecting the output layer for different classes when I have the nchw format. For the current hwc format I've to manually reorder the output, which is quite slow.

Thanks for sharing your case, we'll add this feature later.

[hwangdeyu]

@devalexqt We have added --outputs_as_nchw option which is supported it in latest main branch. Please tell me if it's work for your model demand Thanks!