Module that returns something other than tensor

Hi Pytorch
I have a module that returns a dict of outputs (I like to keep them labeled so I can disregard some of them during testing). This apparently breaks everything, including __call__ (I use forward instead, i understand this is bad. It also breaks data parallelism. I was wondering if there were best practices for such a thing (where I want to return several tensors, each with their own loss, keep them organized, etc). Would hooks be applicable here, for example?

It’s easiest for now to return a tuple; if you need labels maybe use a namedtuple?

Thanks I ended up submitting a pull request myself, it was due to a bug.