Issue with importing safetensors generated by PyTorch

[the-pawn-2017]

The network structure I want to load is a bit complex, with non-tensor fields in the model, such as Int, and also torch.views. I encountered problems during import and can't find a way to ignore these fields

[antimora]

Have you tried SafetensorsStore? It is in burn-store

Also use PyTorchAdapter with safetensors

[the-pawn-2017]

thank you for you reply. what's the different of burn-import and store？I know that in Candle, layers can be configured one by one via varbuild(I forget the exact name, but it's something like a hashmap), but I'm not sure how to perform such an operation in Burn.

[antimora]

We are moving towards burn-store.

If you want individual tensors from the file, you can use this: https://github.com/tracel-ai/burn/tree/main/crates/burn-store#direct-tensor-access

But I would recommend reading the README file in its entirety so you might take advantage of other features. I don't have full context what you are doing. So it might be better if you learn of new features. We can also get some feedback from you.

[the-pawn-2017]

Thank you for your patient reply! I’ll go and explore the new module thoroughly.