How to migrate existing models from 0.4.0 to 0.5.0?

[ChristianWiele]

Hi,

I trained a lot of models on version 0.4.0. After upgrade to 0.5.0 I can't load them anymore, as I get errors like

AttributeError: Can't get attribute 'DecoderBlock' on <module 'segmentation_models_pytorch.decoders.unet.decoder' from '/home/christian/dev/HSP/libraries/atnc_mcv/venv/lib/python3.12/site-packages/segmentation_models_pytorch/decoders/unet/decoder.py'>

So, how can I keep using my trained models with the new version?

Thanks, Christian

[qubvel]

Hi @ChristianWiele, it depends on the models you have trained. state_dicts should be backward compatible, so saving weights in 0.4.0 and loading them in 0.5.0 should work. In case you save the entire model (troch.save(model)), I would recommend loading it in v 0.4.0, saving with save_pretrained and then loading in 0.5.0 with from_pretrained - this should work.

[qubvel]

Sorry for the inconvenience 🤗 that's the way to bring new features to the library

[ChristianWiele]

Hi @qubvel, thanks for the quick response. This seems to work. I use mlflow to save my models which uses torch.save() . I guess there is no direct way to use save_pretrained with mlflow!?

[qubvel]

I'm not very familiar with MLflow, not sure if there is a direct way to do that. However, I hope some callbacks exist to override the default behaviour.