espnet2.samplers.sorted_batch_sampler.SortedBatchSampler

source

class espnet2.samplers.sorted_batch_sampler.SortedBatchSampler(batch_size: int, shape_file: str, sort_in_batch: str = 'descending', sort_batch: str = 'ascending', drop_last: bool = False)

Bases: AbsSampler

BatchSampler with sorted samples by length.

This sampler is designed to create batches of samples sorted by their lengths. It can sort samples either in ascending or descending order within a batch and allows for sorting of batches as well.

batch_size

The size of each batch.

• Type: int

shape_file

Path to the file containing the shape information for each sample.

• Type: str

sort_in_batch

Defines the sorting order for samples within each batch. Can be ‘descending’, ‘ascending’, or None.

• Type: str

sort_batch

Defines the sorting order for the batches. Can be ‘ascending’ or ‘descending’.

• Type: str

drop_last

If True, drop the last batch if it is smaller than the batch size.

• Type: bool

• Parameters:

  • batch_size (int) – The size of each batch. Must be greater than 0.
  • shape_file (str) – Path to the file containing the shape information for each sample.
  • sort_in_batch (str) – ‘descending’, ‘ascending’ or None for sorting samples within each batch.
  • sort_batch (str) – ‘ascending’ or ‘descending’ for sorting the batches.
  • drop_last (bool) – If True, the last batch will be dropped if it has fewer than batch_size samples.

• Raises:

  • ValueError – If sort_in_batch or sort_batch is not one of the expected values.
  • RuntimeError – If no samples or batches are found.

Examples

>>> sampler = SortedBatchSampler(batch_size=32, shape_file='shapes.csv',
...                               sort_in_batch='ascending',
...                               sort_batch='descending')
>>> for batch in sampler:
...     print(batch)

NOTE

The shape file should be a CSV file where each line corresponds to a sample and contains its length.