Convert to onnx after script removal?

[christian-holland]

I want to convert a trained model to onnx. It looks like previously there was a pytorch2onnx.py script, but it looks like this doesn't exist anymore. I assume this change was made, so that the mmdeploy repo is used instead.

However it is not entirely clear to me from the docs there, how it would be used in that way. Trying to follow the steps as I understood them from the mmdeploy docs, I tried:

• Build the docker image like described here and installed the mmpretrain repo from source.
• Run the conversion like as described here.

In particular I tried:

python tools/torch2onnx.py \
    configs/mmcls/classification_onnxruntime_dynamic.py \
    /data/mmpretrain/work_dirs/swinv2-tiny-w16_16xb64_in1k-256px/swinv2-tiny-w16_16xb64_in1k-256px.py \
    /data/mmpretrain/work_dirs/swinv2-tiny-w16_16xb64_in1k-256px/epoch_300.pth \
    /data/FST/221130_person_crops/val/yellow_person/img_2022_09_14_10_31_45t578460.jpg \
    --work-dir /data/mmpretrain/work_dirs/swinv2-tiny-w16_16xb64_in1k-256px

I get this error ImportError: mmcls has not been installed. Import mmcls failed. This is just the old name of this repo? However, this made me question if I am trying to do the correct thing in the first place here for what I am trying to achieve.

Can you give me some guidance what would be intended way if I want to convert a trained model to onnx with this repo?

[RunningLeon]

@christian-holland hi，could use mmdeploy after this PR open-mmlab/mmdeploy#2003

[christian-holland]

Thanks. The Dockerfile uses master as default where these changes are not yet present. Building from the main branch resolves the issue.

[RunningLeon]

Good to hear that. If you mean the dockerfile in mmdeploy, pls. use the latest main branch, which has already updated to use main branch as per here and here.