Hi@mosh98 ,

Not silly at all. When you say BERT pytorch models do you mean an implementation of BERT in pytorch that is not part of the Huggingface package ?

If that's the case this package unfortunately won't work out of the box as much of the implementation leans heavily on HF's model architecture.

I would imagine though that Captum could be made work with a straight pytorch implementation.

Hope that helps :-)