Request: Support for Multilabel

[vody-am]

Hi! Thanks for this cool project. I think it'd be very useful to add in support for multilabel classification as well. This might require training with BCEWithLogitsLoss, and accepting a different dataset format. I attempted to add it in but ran into some bugs. If interested, this feature would be appreciated. If I manage to get it working, I'll send a PR.

[micedre]

Hi! This would be a very nice feature indeed. If you already have some code you are willing to share, don't hesitate to open a PR (even as draft), we may be able to help you.

We are also in the process of refactoring this package so that we can provide different classifier type (see InseeFrLab/torch-fastText#50). Some work on that is available here : https://github.com/InseeFrLab/torchTextClassifiers , it might be easier to add your feature there (and it will be the package maintained long term).

Thanks a lot for your contribution.

[vody-am]

Hey! Thanks for that, I can take a look at the classifiers package, it is similar to what I want as well! I got a very basic version (in pure Pytorch) working with BCE and a few other losses for multi-label, and when I get time, I will try to port it over to your project as well. Thanks for sharing code, it was helpful even if to double-check that my work was in the right direction!

[micedre]

Hi,

This package wil soon be archived in favor of the new (and shiny) https://github.com/InseeFrLab/torchTextClassifiers.

If you don't mind, can I transfer your issue to the new repository ?

[vody-am]

@micedre let's do it! I actually have some time now so I can dive into your project (I was busy training a classifier 😂 ).

[ericjeangirard]

Also interested in this feature to switch from fasttext to torchTextClassifiers

[meilame-tayebjee]

I am on it. Will be available soon.

[meilame-tayebjee]

Feel free to reopen if something is wrong, but now multilabel is supported. You can refer to this notebook for an example. We have not pushed the new version to PyPI for now so make sure you are installing the package from the GitHub repository to have the latest version.