espnet2.iterators.sequence_iter_factory.SequenceIterFactory

source

class espnet2.iterators.sequence_iter_factory.SequenceIterFactory(dataset, batches: AbsSampler | Sequence[Sequence[Any]], num_iters_per_epoch: int | None = None, seed: int = 0, shuffle: bool = False, shuffle_within_batch: bool = False, num_workers: int = 0, collate_fn=None, pin_memory: bool = False)

Bases: AbsIterFactory

Build iterator for each epoch.

This class creates a PyTorch DataLoader with additional features for reproducibility and control over the number of samples per epoch. The following points are noteworthy:

• The random seed is determined based on the epoch number, ensuring reproducibility when resuming training.
• It allows restriction on the number of samples for each epoch, facilitating control over the interval between training and evaluation.

dataset

The dataset to be used for the DataLoader.

• Type: Any

sampler

An instance of a sampler that defines the order in which data is drawn.

• Type:AbsSampler

num_iters_per_epoch

The number of iterations to perform in one epoch.

• Type: Optional[int]

seed

The seed for random number generation.

• Type: int

shuffle

Whether to shuffle the data at the beginning of each epoch.

• Type: bool

shuffle

_within_batch

Whether to shuffle the data within each batch.

• Type: bool

num_workers

The number of worker processes for data loading.

• Type: int

collate_fn

Function to merge a list of samples into a batch.

• Type: Optional[callable]

pin_memory

If True, the data loader will copy Tensors into pinned memory before returning them.

• Type: bool

• Parameters:

  • dataset (Any) – The dataset from which to load the data.
  • batches (Union [AbsSampler , Sequence *[*Sequence *[*Any ] ] ]) – The batches of data to be sampled.
  • num_iters_per_epoch (Optional *[*int ]) – The number of iterations per epoch (default is None).
  • seed (int) – Random seed (default is 0).
  • shuffle (bool) – Whether to shuffle the data (default is False).
  • shuffle_within_batch (bool) – Whether to shuffle data within each batch (default is False).
  • num_workers (int) – Number of subprocesses to use for data loading (default is 0).
  • collate_fn (Optional *[*callable ]) – Function to merge a list of samples into a batch (default is None).
  • pin_memory (bool) – If True, the data loader will copy Tensors into pinned memory (default is False).

• Returns: A PyTorch DataLoader instance configured with the specified parameters.

• Return type: DataLoader

####### Examples

>>> factory = SequenceIterFactory(dataset, batches, num_iters_per_epoch=10)
>>> data_loader = factory.build_iter(epoch=1)

NOTE

The build_iter method can be called multiple times with different epoch values to generate iterators for different training phases.

build_iter(epoch: int, shuffle: bool | None = None) → DataLoader

Build iterator for each epoch.

This class simply creates a PyTorch DataLoader with the following features:

• The random seed is determined based on the epoch number, ensuring

reproducibility when resuming from a training session.

• Allows restriction on the number of samples per epoch, controlling the interval between training and evaluation.

dataset

The dataset to load data from.

• Type: Any

sampler

The sampler used to sample batches of data.

• Type:AbsSampler

num_iters_per_epoch

Maximum number of iterations per epoch.

• Type: Optional[int]

shuffle

Whether to shuffle the data.

• Type: bool

shuffle

_within_batch

Whether to shuffle data within each batch.

• Type: bool

seed

Seed for random number generation.

• Type: int

num_workers

Number of subprocesses to use for data loading.

• Type: int

collate_fn

Function to merge a list of samples.

• Type: Optional[callable]

pin_memory

If True, the data loader will copy Tensors into CUDA pinned memory.

• Type: bool

• Parameters:

  • epoch (int) – The current epoch number.
  • shuffle (Optional *[*bool ]) – If provided, overrides the default shuffle setting.

• Returns: A PyTorch DataLoader object configured for the current epoch.

• Return type: DataLoader

• Raises:

  • AssertionError – If the length of batches does not match
  • num_iters_per_epoch when it is specified. –

####### Examples

>>> factory = SequenceIterFactory(dataset=my_dataset, batches=my_batches)
>>> data_loader = factory.build_iter(epoch=1, shuffle=True)

NOTE

Ensure that the dataset and batches are compatible with the DataLoader’s requirements.