CatBoost and PyTorch

CatBoost is a maintained boosted-tree system; PyTorch is a differentiable tensor and neural-network framework. A DataLoader does not add capabilities to CatBoost, and categorical strings should not be accidentally coerced into anonymous numeric tensors.

Why the previous example was removed

The former page used PyTorch only to batch and normalize an externally downloaded numeric dataset before converting it back to NumPy. It hid CatBoost's defining native categorical interface and added shuffling and schema risks without a learned PyTorch component.

A defensible hybrid

Use PyTorch to learn an encoder for images, text, audio, or another modality. Choose the encoder using training-only validation, then freeze it, call eval(), and extract embeddings under torch.no_grad(). Join embeddings to remaining tabular fields by stable row identifiers and keep the original categorical columns as strings for CatBoost.

Pass explicit categorical names or indices through cat_features. Keep train, validation, and test entities separate; group-aware or time-aware boundaries must survive feature extraction. If the encoder is tuned, fit it inside every outer fold.

Stacking and serving

For stacking, use out-of-fold predictions to train the combiner. Training on CatBoost's in-sample probabilities leaks target-fitting behavior. At serving time, version the PyTorch weights and preprocessing together with the embedding schema, CatBoost feature names and categorical specification, model file, and exact package versions.

Takeaway

Direct CatBoost is the executable reference for ordinary mixed-type tables. PyTorch belongs only when it supplies a genuine learned representation or separately evaluated model.