[mertalev]

Description

Fixes #2947 by adding error handling when loading models. If a model file fails to load, the folder for that model is deleted and the model is redownloaded. If this fails as well, then it will still error out.

Testing

This bug was reproduced by corrupting the first byte of both the CLIP model and one of the facial recognition models (commands adapted from this). Adjust the model paths as needed.

export CLIP_MODEL_PATH=/cache/clip/clip-ViT-B-32/sentence-transformers_clip-ViT-B-32/0_CLIPModel/pytorch_model.bin
export FACE_MODEL_PATH=/cache/facial-recognition/buffalo_l/models/buffalo_l/1k3d68.onnx
echo -ne \\xFF | dd conv=notrunc bs=1 count=1 of=$CLIP_MODEL_PATH
echo -ne \\xFF | dd conv=notrunc bs=1 count=1 of=$FACE_MODEL_PATH

The models should be redownloaded on startup and otherwise load without error.

[vercel]

The latest updates on your projects. Learn more about Vercel for Git ↗︎

1 Ignored Deployment

Name	Status	Preview	Comments	Updated (UTC)
immich	⬜️ Ignored (Inspect)			Jun 27, 2023 6:18pm

[mertalev]

Tested and I can confirm that it works for both Insightface and HF models. One thing to note with the Insightface models is that if you just touch an empty model file for testing, it will give a different error Fail instead of InvalidProtobuf. I didn't include Fail since it seems too broad - if downloading again doesn't solve the issue, then it will just keep downloading on container restart.

[mertalev]

I managed to get the InvalidProtobuf error by corrupting the first byte instead of completely replacing it. It's a crude way to test it but it works.