[MRG] Fix batch issue when generating features + add sample_weight in deep models

[YanisLalou]

Before when doing a model.predict_features(X) we were passing all the input in a single batch.
This created CUDA out of memory issues when working with big datasets.
Thus here I tried to mimic the behaviour of skorch models to use batching.

Might not be the best way to fix that issue though, would love your opinion @tgnassou

[tgnassou]

I'm wondering if you can not just create a dataloader on your own, without using skorch function. Like that you only create a dataloader with X, iterate on the dataloader and that's it. It seems very long to do something simple, but maybe I'm wrong

[antoinecollas]

sample_weight is everywhere but in practice we only use it to reweight the loss, right? IMO, it should used only when calling the loss but maybe I misunderstood something.

If you run coverage run -m pytest -v -s && coverage html && open htmlcov/index.html on your branch and on the main branch, you will see that you have added 24 lines in skada/deep/base.py that are not covered by tests.

[antoinecollas]

From Skorch FAQ: when X is a dict, its keys are passed as kwargs to forward, thus our forward has to have the arguments 'data' and 'sample_weight'; usually, sample_weight can be ignored here.

[codecov]

Codecov Report

Attention: Patch coverage is 74.26471% with 35 lines in your changes missing coverage. Please review.

Project coverage is 96.31%. Comparing base (28aaf82) to head (6aeda7e).

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #220      +/-   ##
==========================================
- Coverage   97.01%   96.31%   -0.70%     
==========================================
  Files          54       54              
  Lines        5429     5486      +57     
==========================================
+ Hits         5267     5284      +17     
- Misses        162      202      +40     

[tgnassou]

Very good PR, the only thing missing is to modify also CDANModule