feat(skorch): add an inherited class from skorch.NeuralNet that is compatible with PyTorch Frame

[34j]

Closes #147

~~@MacOS Please continue from here if it helps. Sorry for being so loud, but this took me a whole day, so I would appreciate it very much if you could make me as a co-author if you used this code.~~

[codecov[bot]]

Codecov Report

All modified and coverable lines are covered by tests :white_check_mark:

Project coverage is 93.52%. Comparing base (ee98b87) to head (aa5484d). Report is 6 commits behind head on master.

:exclamation: Current head aa5484d differs from pull request most recent head cb76e8d. Consider uploading reports for the commit cb76e8d to get more accurate results

Additional details and impacted files

@@           Coverage Diff           @@
##           master     #375   +/-   ##
=======================================
  Coverage   93.52%   93.52%           
=======================================
  Files         124      124           
  Lines        6456     6456           
=======================================
  Hits         6038     6038           
  Misses        418      418           

:umbrella: View full report in Codecov by Sentry.
:loudspeaker: Have feedback on the report? Share it here.

[MacOS]

Sure - on both!

[34j]

@weihua916 Would you mind reviewing if you think this is a good way to implement it?

Also, it is strange that mypy in pre-commit does not raise errors, but mypy in CI does. I don't think there is any way to deal with this.

[34j]

@weihua916 @zechengz @yiweny Would appreciate your review, thank you very much in advance.

[weihua916]

A kind check-in. Is there any progress here?

[34j]

No progress, sorry

[34j]

• First, torch_frame.DataSet cannot properly handle DataFrane with indices not [0,1,...]. dataset.py L736 must be modified to indices = np.where(self.df[self.split_col] == SPLIT_TO_NUM[split])[0].
• ~~Second, though I am not aware of this before, torch_frame models require col_stats, col_names_dict in the constructor and unfortunately do not allow modifications after instantiation. Furthermore, they do not follow the standards of scikit-learn as they do not save the constructor parameters, making it impossible to instantiate the module again.~~ Resolved by allowing to pass a function to create a module

[34j]

Almost done

[34j]

@weihua916 Thanks for your patience, I've added tests.

[34j]

@weihua916 Removed all tutorials and added examples/sklearn_api.py instead. The failing tests are probably due to a pandas version update and are not related to this PR.

[34j]

@weihua916 Would you please reconsider merging this PR?

This change makes it easy to try out torch-frame based neural networks on code that already uses scikit-learn. (which was demonstrated in https://github.com/pyg-team/pytorch-frame/pull/375#discussion_r1673507323).

(This PR is not intended to save training of Pytorch models that do not use scikit-learn. lightning or fastai should be used for such applications.)

Thank you in advance.

[34j]

@weihua916 @yiweny @akihironitta Any chance that this PR could be merged?

[34j]

Thanks for your positive reply. In that case I would consider trying to make another package, but I am very busy right now, so please leave this as it is for a while. Note that in any case one-line change I made (in dataset.py) needs to be merged for that to work and you may want to review that specific line.