lbotecur/tabomics-v2.5

TabOmics v2.5

TabOmics is a tabular foundation model for wide omics data. It is pretrained only on synthetic tables and uses in-context learning: given labeled training rows and unlabeled 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), processes every feature jointly, also works on ordinary tables, and provides sample embeddings and feature importance from the same forward pass.

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

Model details

• Classifier of up to 10 classes.
• A Transformer of 12 blocks, each with attention between features, attention between rows and an MLP; embedding size 192, 6 attention heads, 7.3M parameters.
• Pretrained only on synthetic tables; the recipes of the published models are in configs/ of the repository. Standardized features are clipped at 10 at inference.
• n_estimators averages ensemble members: forward passes on copies of the table with the columns reordered and the categories and classes renumbered at random (1 by default).

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