Models

A model subclasses BaseModel and implements forward(batch: Batch) -> ModelOutput. ModelOutput.preds is the main logits or regression tensor; aux can carry embeddings, per-modality predictions, or features used by callbacks and feature evaluation.

The built-in fusion factories are:

Registry key	Behavior
feature_fusion_model	Encodes available modalities, concatenates embeddings, applies optional dropout, and calls a head.
prediction_fusion_model	Runs per-modality submodels and combines logits with mean, sum, or weighted averaging.
gated_fusion_model	Computes modality gates over embeddings; masks suppress unavailable modalities.
gated_prediction_fusion_model	Computes gates over per-modality logits and requires num_classes.

A fusion model skips a modality absent from inputs and raises ValueError when no configured modality is present. gated_fusion_model requires every projected embedding to have shared_dim; use set_projection when encoders have different dimensions.

For a plugin model, register a factory under MODELS and keep the boundary stable. The API reference contains core interfaces.