Sorry if I missed some of the other discussions on this – In a lot of my projects I have custom PyTorch Functions defined and have been getting a lot of deprecation warning errors about updating to the staticmethod interface. Here’s an example of one with a lot of attributes that aren’t necessarily Tensors that I would like to access in the forward/backward pass and I would like to return some information from inside of the forward pass that also aren’t Tensors, which I don’t know how to do with the new interface.
Here’s another example that I’m thinking about updating: http://github.com/locuslab/qpth/blob/master/qpth/qp.py
And here are others that I have updated that I’ve done quite messily by passing all of the attributes into the
forward method and putting them back into the context – a hack to get nearly the same functionality as the old interface. Is returning
None in the backward pass for all of the additional flags the right thing to do?
Do you have any recommendations/thoughts before I start making more updates for this? Also how should I return auxiliary information that will never be backpropped through from a