Skip to content

Data and loaders

A datamodule factory returns an object with train_dataloader, val_dataloader, and test_dataloader. The built-in DataModule creates PyTorch DataLoader instances from a single dataset, a mapping of named datasets, or a sequence of datasets.

The default MaskingCollate collects tensor fields into a Batch, pads variable-length sequences when configured, and produces flat mask keys such as audio_mask and sequence_mask. A plugin can register a custom collate under the collates registry.

Multiple loaders

A loader may be a single DataLoader, a mapping, or a list/tuple. Names are normalized for stable metric keys. Training supports:

  • single: consume the first loader;
  • round_robin: alternate loaders in declaration order;
  • weighted: sample loaders with train_loader_weights.

train_stop_on: min stops when the first active loader is exhausted; max continues until all active loaders are exhausted. For weighted, a finite loader can be removed in max mode. Mapping split names are sanitized to alphanumerics, dot, underscore, and dash.

Batch contract

A dataset sample should provide input tensors and may provide targets, masks, and metadata.

Batch(
    inputs={"audio": audio_tensor, "video": video_tensor},
    targets=target_tensor,
    masks={"audio_mask": audio_mask},
    meta={"id": sample_id},
)

Batch.get_masks accepts no argument for all masks or a specific flat key. For compatibility it also reads legacy nested masks from masks["mask"] and meta["masks"], flattening modality names to modality_mask.