core/nn/modalities/multi_modal_classifier library

Multimodal classifier — fuses per-modality features via mean pooling + concatenation + a small MLP head.

Late-fusion recipe (each modality is encoded independently and then combined at the feature level):

  1. AudioTransformer(audio) → mean-pool → [1, embedA]
  2. VideoTransformer(video) → mean-pool → [1, embedV]
  3. optional TextTransformer(tokens) → mean-pool → [1, embedT]
  4. concat on axis 1 → [1, embedA + embedV + (embedT?)]
  5. Linear(..., numClasses) → logits [1, numClasses]

The three encoders don't need to share embedDim — the fusion head sizes itself from their totals.