Codecov Report

Merging #248 (a9f9182) into dev (0ed6f78) will decrease coverage by 0.75%.
The diff coverage is 76.19%.

@@            Coverage Diff             @@
##              dev     #248      +/-   ##
==========================================
- Coverage   66.01%   65.26%   -0.75%     
==========================================
  Files           5       29      +24     
  Lines        2401     6664    +4263     
  Branches      641     1573     +932     
==========================================
+ Hits         1585     4349    +2764     
- Misses        520     1472     +952     
- Partials      296      843     +547     

@gbaydin I think this can be merged, or you can take the change and apply it into your own work on transformers?

It raises separate questions of generalizing differentiable objects to beyond tensors but I think we can deal with those later, this seems like a reasonable minimal start?

Actually this is indeed very relevant for the type of things I'm working on. For example in recurrent models I had to think about a design to pass and return hidden state tensors in addition to the input and return values of .forward. Thanks for pinging, I'm going to have a close look and merge.

@dsyme I added a couple of fixes.

I think the most important one was that Model<'In, 'Out>.clone was returning Model<Tensor, Tensor> which I fixed to return Model<'In, 'Out>.

I also moved a few things that were not involving <'In, 'Out> to BaseModel.

I'm really happy about this PR in general. I think this is quite nice and powerful! For example, we can do things like:

• Model<string, string> for machine translation between languages
• Model<unit,Tensor> for some generative models etc.

Also after reading through all <'In, 'Out>s this reminds me of https://www.in-n-out.com/ :)

This is great, thanks for refining it!