Support multi-dimensional y data

tvdboom / ATOM

Automated Tool for Optimized Modelling

https://tvdboom.github.io/ATOM/

MIT License

148 stars 14 forks source link

Support multi-dimensional y data #33

Open no-clue-what-to-do opened 2 years ago

no-clue-what-to-do commented 2 years ago

Contribution guidelines

[X] I've read the contribution guidelines.
[X] The documentation does not mention anything about this feature.
[X] There are no open or closed issues that are related to this feature.

Description

I have a lot of project where I am trying to predict multiple targets at once. Atom has a bunch of nice features that I'd like to use, but I'm not sure its worth it if I have to build a separate model for each target dimension

Use Cases

No response

Screenshots / Mockups

No response

tvdboom commented 2 years ago

This would be a nice extra feature indeed. Since this is a lot of work, I will consider adding this to the release after the next one. The next release is almost finished and will be out soon. After that, I'll do some digging to see how the code could be refactored to support multi-target models. I'll keep you updated in this page.

no-clue-what-to-do commented 2 years ago

Great! I would love to contribute if I can. I've been struggling for a long time to find a package that works for my DL workflow and this package is the closest I've come. Sklearn pipelines are cool but they don't handle the validation data that well, and don't have a lot of the tooling you've put in.

tvdboom commented 1 year ago

Hi. Now that 4.14 has been released I was looking at the features for next release. I think it's going to be a big one (probably v5.0.0) so this seems as good a moment as any to implement the multioutput support. I added a list of things that need to be done. Feel free to pick up any of them or add your own if you think I forgot one.

Code

[ ] Refactor _prepare_input in basetransformer.py to accept multidim y
[ ] Refactor models to use classifier/regressor chain if multioutput
[ ] Check all transformers and adapt to multioutput where needed
[ ] Refactor the way metrics are calculated
[ ] Refactor the evaluate method
[ ] Add multioutput support class attribute to models
[ ] Add multioutput support to get_available_models

Documentation

[ ] Add multioutput section to user guide
[ ] Add multioutput example
[ ] Add multioutput support banner to models in API

tvdboom commented 1 year ago

@no-clue-what-to-do multioutput datasets are supported sicne version 5.1.0. Read about it in the user guide.