espnet2.samplers.build_batch_sampler.build_batch_sampler

source

espnet2.samplers.build_batch_sampler.build_batch_sampler(type: str, batch_size: int, batch_bins: int, shape_files: Tuple[str, ...] | List[str], sort_in_batch: str = 'descending', sort_batch: str = 'ascending', drop_last: bool = False, min_batch_size: int = 1, fold_lengths: Sequence[int] = (), padding: bool = True, utt2category_file: str | None = None) → AbsSampler

Helper function to instantiate various types of BatchSampler.

This function creates and returns an instance of a specified batch sampler based on the provided parameters. It supports multiple sampler types, each designed for different batching strategies.

espnet2.samplers.build_batch_sampler.type

The mini-batch type. Options include “unsorted”, “sorted”, “folded”, “numel”, “length”, or “catbel”.

• Type: str

espnet2.samplers.build_batch_sampler.batch_size

The mini-batch size. Used for “unsorted”, “sorted”, “folded”, and “catbel” modes.

• Type: int

espnet2.samplers.build_batch_sampler.batch_bins

Used for “numel” mode.

• Type: int

espnet2.samplers.build_batch_sampler.shape_files

Text files describing the length and dimension of each feature, e.g., “uttA 1330,80”.

• Type: Union[Tuple[str, …], List[str]]

espnet2.samplers.build_batch_sampler.sort_in_batch

Sorting order for samples within each batch.

• Type: str

espnet2.samplers.build_batch_sampler.sort_batch

Sorting order for batches.

• Type: str

espnet2.samplers.build_batch_sampler.drop_last

Whether to drop the last incomplete batch.

• Type: bool

espnet2.samplers.build_batch_sampler.min_batch_size

Minimum batch size used for “numel” or “folded” mode.

• Type: int

espnet2.samplers.build_batch_sampler.fold_lengths

Used for “folded” mode to specify fold lengths.

• Type: Sequence[int]

espnet2.samplers.build_batch_sampler.padding

Whether sequences are input as a padded tensor (used for “numel” mode).

• Type: bool

espnet2.samplers.build_batch_sampler.utt2category_file

Optional file to categorize utterances.

• Type: Optional[str]

• Parameters:

  • type – Mini-batch type. Options: “unsorted”, “sorted”, “folded”, “numel”, “length”, or “catbel”.
  • batch_size – The mini-batch size.
  • batch_bins – Number of bins used for “numel” mode.
  • shape_files – Text files describing feature dimensions.
  • sort_in_batch – Sorting order for samples in each batch.
  • sort_batch – Sorting order for batches.
  • drop_last – Whether to drop the last incomplete batch.
  • min_batch_size – Minimum batch size for “numel” or “folded” mode.
  • fold_lengths – Lengths for folding (used in “folded” mode).
  • padding – Whether to pad sequences for “numel” mode.
  • utt2category_file – Optional file for categorizing utterances.

• Returns: An instance of a batch sampler.

• Return type:AbsSampler

• Raises:ValueError – If no shape files are provided or if the number of fold_lengths does not match the number of shape_files.

Examples

Create an unsorted batch sampler

sampler = build_batch_sampler(

type=”unsorted”, batch_size=32, batch_bins=0, shape_files=[“shapes.txt”],

)

Create a sorted batch sampler

sampler = build_batch_sampler(

type=”sorted”, batch_size=32, batch_bins=0, shape_files=[“shapes.txt”], sort_in_batch=”ascending”, sort_batch=”descending”,

)

Create a folded batch sampler

sampler = build_batch_sampler(

type=”folded”, batch_size=64, batch_bins=0, shape_files=[“shapes1.txt”, “shapes2.txt”], fold_lengths=[10, 20],

)