Class SequenceBase

DataLoader/SequenceBase.py:18–75

class SequenceBase(IterableDataset[T_Data], ABC, ConfigTestableSubclass):
    @classmethod
    def name(cls) -> str:
        """
        Assign a short name for the dataset class. By default will be the class name.
        Overwrite this function if you want to create a more readable name used in `name` field in config.
        """
        return cls.__name__

    @abstractmethod
    def __getitem__(self, local_index: int) -> T_Data: ...

    # No need to read further ### Implementation details below ############
    def __init__(self, length: int) -> None:
        super().__init__()
        self.origin_length: Final[int] = length
        self.indices = np.arange(0, length, 1)
    
    def get_index(self, local_index: int) -> int:
        """
        SequenceBase class supports masking / sampling of sequences.
        The 'actual index' refer to the index in the original sequence,
        In contrast with the 'logical index' (index after mask is applied) used by the user.
        """
        return self.indices[local_index].item()

    @final
    def clip(self, start_idx: int | None = None, end_idx: int | None = None, step: int | None = None) -> Self:
        self.indices = self.indices[start_idx:end_idx:step]
        return self

    def preload(self) -> "PreloadedSequence[T_Data]":
        return PreloadedSequence(self)
    
    def transform(self, actions: list[Callable[[T_Data,], T_Data]] | Callable[[T_Data,], T_Data]) -> "TransformSequence[T_Data] | Self":
        if isinstance(actions, list) and len(actions) == 0: return self
        return TransformSequence(self, actions)

    def __len__(self) -> int:
        return self.indices.size
    
    def __iter__(self) -> Generator[T_Data, None, None]:
        for idx in range(len(self)): yield self[idx]

    def __repr__(self) -> str:
        return f"{self.name()}(orig_len={self.origin_length}, clip_len={len(self)})"
    
    @staticmethod
    def collate_fn(batch: list[T_Data]) -> T_Data:
        """
        Collate function for DataLoader.
        """
        return batch[0].collate(batch)
    
    @staticmethod
    def config_dict2ns(cfg: SimpleNamespace | dict[str, Any]) -> SimpleNamespace:
        if isinstance(cfg, SimpleNamespace): return cfg
        return build_dynamic_config(cfg)[0]