sbss.common.datasets.WavHDF5Dataset

class sbss.common.datasets.WavHDF5Dataset(dataset_path: Path | str, duration: int | None = None, sr: int | None = None, randperm_mic: bool = True, grp_list: Path | str | list[str] | None = None)[source]

Dataset class for loading and preprocessing multi-channel audio data stored in HDF5 format.

This dataset wraps an HDF5Dataset using a caching layer for faster access. It supports random microphone permutation for data augmentation and optional random cropping of the waveform based on the specified duration and sampling rate.

Parameters:

  • dataset_path (Path | str) – Path to the HDF5 dataset directory or file.

  • duration (int | None, optional) – Target duration of audio clips in seconds. If None, the full audio is used.

  • sr (int | None, optional) – Sampling rate used to calculate the crop length when duration is specified.

  • randperm_mic (bool, optional) – Whether to randomly permute microphone channels. Defaults to True.

  • grp_list (Path | str | list[str] | None, optional) – List or path specifying group names to load from the HDF5 dataset. If None, all groups are used.

Returns:

A PyTorch dataset yielding preprocessed multi-channel waveforms.

Return type:

torch.utils.data.Dataset

__init__(dataset_path: Path | str, duration: int | None = None, sr: int | None = None, randperm_mic: bool = True, grp_list: Path | str | list[str] | None = None) → None[source]

Methods

__init__(dataset_path[, duration, sr, ...])