How to load pytorch tensor

[med1844]

I'm trying to port a model from pytorch to burn. I want to use pytorch pre-calculated tensors & model weights to test if the burn implementation of the same model components works identically as the torch counterpart. This requires me to load tensors.

I haven't seen such an example in the documentation, issues, discussions, or in the examples folder. Any help on how to do this idiomatically is appreciated!

[antimora]

We should have a special book section on this. Please file an issue and mentione me. Meanwhile you can use the following resources:

1. Book section https://burn.dev/burn-book/import/pytorch-model.html

2. Use this repo to test model output:

https://github.com/antimora/face-onnx/blob/main/src/bin/with_weights_test_outputs.rs

This uses outputs from pytorch saved as pt file.

[med1844]

Thank you for the info. The example from face-onnx helps me deserialize tensors.

Further investigation suggests this might be caused by candle-core::pickle::read_pth_tensor_info only supports dict-structured .pt file, where non-dict tensors are simply discarded.

https://github.com/huggingface/candle/blob/2f49e1b5349f4e56677ec0d3dc3fe98f9cbb87c7/candle-core/src/pickle.rs#L703-L713

That explains why we must derive Module to load a tensor and wrap torch.Tensor with dict before torch.save.

If they can modify these lines to handle not only Object::Dict but also Object::Reduce then this problem could be fully solved.