espnet2.samplers.category_balanced_sampler.CategoryBalancedSampler

source

class espnet2.samplers.category_balanced_sampler.CategoryBalancedSampler(batch_size: int, min_batch_size: int = 1, drop_last: bool = False, category2utt_file: str | None = None, epoch: int = 1, **kwargs)

Bases: AbsSampler

Sampler that maintains an equal distribution of categories (i.e., classes)

within each minibatch. If the batch size is smaller than the number of classes, all samples in the minibatch will belong to different classes.

The key_file is a text file that describes each sample name. It should be formatted as follows:

utterance_id_a utterance_id_b utterance_id_c

The first column is referred to, so a ‘shape file’ can also be used, which has the following format:

utterance_id_a 100,80 utterance_id_b 400,80 utterance_id_c 512,80

batch_size

The size of each batch.

• Type: int

min_batch_size

The minimum size of each batch. Default is 1.

• Type: int

drop_last

Whether to drop the last batch if it’s smaller than batch_size. Default is False.

• Type: bool

category2utt_file

Path to the file mapping categories to utterances.

• Type: Optional[str]

epoch

The epoch number for seeding the random number generator.

• Type: int

• Parameters:

  • batch_size (int) – The size of the minibatch.
  • min_batch_size (int , optional) – Minimum size of the minibatch. Default is 1.
  • drop_last (bool , optional) – If True, drop the last batch if it’s smaller than batch_size. Default is False.
  • category2utt_file (str , optional) – Path to the file mapping categories to utterances. Must be provided.
  • epoch (int , optional) – The epoch number for random seed. Default is 1.
  • **kwargs – Additional keyword arguments for customization.

• Returns: An iterator that yields batches of utterances.

• Return type: Iterator[Tuple[str, …]]

Examples

>>> sampler = CategoryBalancedSampler(batch_size=4, category2utt_file='path/to/file.txt')
>>> for batch in sampler:
...     print(batch)

NOTE

The random seed is initialized based on the provided epoch number to ensure reproducibility across different runs.