Skip to content

Conversation

@dholzmueller
Copy link
Contributor

@dholzmueller dholzmueller commented Aug 14, 2025

Issue #, if available:

Description of changes: Changed the RealMLP search space.

  • tune some categorical hyperparameters as well
  • use per-fold ensembles (n_ens=8). This should not make it much slower on GPU for most datasets.
    Note that this requires using pytabkit>=1.6.1.

Regarding n_ens, I used 8 as a compromise. One may use 16 - it uses 22 GB RAM on a dataset with 10K numerical features, but it is already 2x slower on datasets with 500 numerical features.

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

@LennartPurucker
Copy link
Collaborator

Benchmarking Results

image

Results data will be shared here soon: https://data.lennart-purucker.com/tabarena/leaderboard_submissions/

(I think the memory estiamte is still not perfect but we can improve it over time)

@LennartPurucker
Copy link
Collaborator

LennartPurucker commented Aug 26, 2025

@dholzmueller I think we can merge this, no?

It now represents the state of the LB submission.

@dholzmueller
Copy link
Contributor Author

In principle yes, though as long as the new version is not on the leaderboard it might be misleading for people who want to check the version on the leaderboard.

@LennartPurucker
Copy link
Collaborator

We also merged the rest that is not on the LB yet. They can look at the commit history, or we can create a tag for LB versions in the future.

@LennartPurucker LennartPurucker merged commit 087e181 into autogluon:main Aug 26, 2025
1 check passed
atschalz pushed a commit to atschalz/tabarena that referenced this pull request Sep 5, 2025
Co-authored-by: LennartPurucker <contact@lennart-purucker.com>
atschalz pushed a commit to atschalz/tabarena that referenced this pull request Sep 5, 2025
Co-authored-by: LennartPurucker <contact@lennart-purucker.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants