Unified Dataset

class standard_e2e.UnifiedE2EDataset(index_data, processed_data_path, regime, feature_loaders=None, label_loaders=None, feature_loaders_config=None, label_loaders_config=None, augmentations=None, index_filters=None, modality_defaults=None)[source]

Bases: Dataset

UnifiedE2EDataset wraps feature and label frame loaders, applies optional index filters and augmentations, and provides PyTorch-compatible access to transformed frames for end-to-end training or evaluation.

Parameters:

  • index_data (DataFrame) – Index dataframe describing available frames and metadata used by frame loaders.

  • processed_data_path (str) – Root path where processed feature/label frames are stored.

  • regime (str) – Augmentation regime; must be one of FrameAugmentation.ALLOWED_REGIMES.

  • feature_loaders (list[FrameLoader] | None) – Pre-instantiated feature frame loaders; mutually exclusive with feature_loaders_config.

  • label_loaders (list[FrameLoader] | None) – Pre-instantiated label frame loaders; mutually exclusive with label_loaders_config.

  • feature_loaders_config (dict | list[dict] | None) – Configuration(s) to build feature frame loaders when feature_loaders is None.

  • label_loaders_config (dict | list[dict] | None) – Configuration(s) to build label frame loaders when label_loaders is None.

  • augmentations (list[FrameAugmentation] | None) – Ordered list of frame augmentations to apply; defaults to IdentityFrameAugmentation.

  • index_filters (list[IndexFilter] | None) – Filters applied to index_data to select usable rows.

  • modality_defaults (dict[Modality, ModalityDefaults] | None) – Default modality-specific settings passed to frame loaders.

classmethod collate_fn(batch, device=None)[source]

Collate a list of {key: TransformedFrameData} into {key: TransformedFrameDataBatch}. TransformedFrameDataBatch internally uses PyTorch’s collate with a Trajectory override.

Return type:

dict[str, TransformedFrameDataBatch]

Parameters: