Can burn execute ptl files for inference?

[knma1992]

I am using the flutter rust bridge, and I am experimenting with implementing my tensorflow and pytorch models in my rust code, so I can call my models from the same codebase without using any flutter packages that are platform specific. My tensorflow model is onnx convertible so I might get lucky to use it in burn. However, for my pytorch mobile (.ptl) models, I would love to just be able to call them using burn for inference.

[laggui]

You can import your pytorch weights with the PyTorchFileRecorder, but I'm not sure that works for pytorch mobile models.

To be honest, I am not too familiar with the ptl model format. After a quick glance at the docs it seems that it is a scripted module that is saved for the lite interpreter. Our PyTorchFileRecorder just needs the state_dict, but it requires an equivalent model to be defined in Burn so that the weights can be loaded in the correct modules.