lbotecur/tabomics-v2.5

TabOmics v2.5

TabOmics is a tabular foundation model for wide omics data. It is pretrained on synthetic tables and uses in-context learning: given labelled training rows and unlabelled test rows, it predicts the test rows in a single forward pass, with no training or tuning on your data. It is designed for wide tables (thousands of features, few samples, as in omics), also works on ordinary tables, and provides row embeddings and feature importance.

from tabomics import Classifier
clf = Classifier(model="v2.5").fit(X_train, y_train)
proba = clf.predict_proba(X_test)

This version

v2.5 is the default model of TabOmics from this release on; v2 stays available as model="v2". The architecture is the same (12 layers, embedding size 192, 7.3M parameters); only the training changed: a longer base stage (300 epochs instead of 100, without v2's stage on 2048 rows), then the same widening to 3000 columns and the same stage at a low learning rate. It is better than v2 on the HDLSS, MLOmics and Shamir benchmarks. The recipe is in configs/v2.5 of the repository.

Code, documentation and licence: https://github.com/lbotecur/tabomics (GNU General Public License v3.0 or later).