tabular_deep/

Three supervised architectures, each with a classifier and a regressor. They are TuiML's own implementations: torch is used only as a tensor and autograd library, and every model is trained from scratch on your data. There is no pretrained checkpoint and nothing is downloaded.

• FT-Transformer (FTTransformerClassifier,

FTTransformerRegressor) -- a feature tokenizer turns every column into its own embedding, and a Transformer attends across those feature tokens.

• SAINT (SAINTClassifier, SAINTRegressor) -- the same,

plus intersample attention: a second attention stage that runs across the rows of the batch, so a sample's representation depends on its neighbours.

• NODE (NODEClassifier, NODERegressor) -- ensembles of

oblivious decision trees made differentiable with entmax15 and stacked with DenseNet-style connectivity.

Being native implementations, all six register in the TuiML hub under their bare class names, alongside every other native algorithm.

PyTorch is an optional dependency:

.. code-block:: bash

pip install 'tuiml[torch]'

Importing this package never imports torch, and constructing a model never imports torch either -- the classes are registered and introspectable on any install. The dependency is required only by fit, which raises a clear ImportError naming the install command. See torch_backend.

>>> from tuiml.algorithms.tabular_deep import FTTransformerClassifier
>>> model = FTTransformerClassifier(d_token=8, n_blocks=1, random_state=0)
>>> model.n_blocks
1