llm.datasets.bert

/home/runner/work/llm-pytorch/llm-pytorch/llm/datasets/bert.py

NVIDIA BERT dataset provider.

Source: https://github.com/hpcaitech/ColossalAI-Examples/blob/e0830ccc1bbc57f9c50bb1c00f3e23239bf1e231/language/roberta/pretraining/nvidia_bert_dataset_provider.py

Batch

Bases: NamedTuple

BERT pretraining batch.

input_ids instance-attribute

input_ids: LongTensor

Input sequence token IDs ((batch_size, seq_len)).

attention_mask instance-attribute

attention_mask: LongTensor

Input sequence attention mask ((batch_size, seq_len)).

Indicates which tokens in input_ids should be attended to (i.e., are not padding tokens). Also known as the input mask.

token_type_ids instance-attribute

token_type_ids: LongTensor

Token IDs indicating the segment labels ((batch_size, seq_len)).

E.g. if input_ids is composed of two distinct segments, the first segment will have token IDs set to 0 and the second to 1. Also known as the segment ids.

masked_labels instance-attribute

masked_labels: LongTensor

True token ID of masked tokens in input_ids ((batch_size, seq_len)).

Indices corresponding to non-masked tokens in input_ids are typically set to -100 to avoid contributing to the MLM loss.

next_sentence_labels instance-attribute

next_sentence_labels: LongTensor | None

Boolean tensor indicating the next sentence label ((batch_size,)).

A true (1) value indicates the next sentence/segment logically follows the first. If there is only one segment in input_ids, this can be set to None.

Sample

Bases: NamedTuple

input_ids instance-attribute

input_ids: LongTensor

Input sequence token IDs ((seq_len,)).

attention_mask instance-attribute

attention_mask: LongTensor

Input sequence attention mask ((seq_len,)).

Indicates which tokens in input_ids should be attended to (i.e., are not padding tokens). Also known as the input mask.

token_type_ids instance-attribute

token_type_ids: LongTensor

Token IDs indicating the segment labels ((seq_len,)).

E.g. if input_ids is composed of two distinct segments, the first segment will have token IDs set to 0 and the second to 1. Also known as the segment ids.

masked_labels instance-attribute

masked_labels: LongTensor

True token ID of masked tokens in input_ids ((seq_len,)).

Indices corresponding to non-masked tokens in input_ids are typically set to -100 to avoid contributing to the MLM loss.

next_sentence_label instance-attribute

next_sentence_label: LongTensor | None

Boolean scalar tensor indicating the next sentence label.

A true (1) value indicates the next sentence/segment logically follows the first. If there is only one segment in input_ids, this can be set to None.

NvidiaBertDataset

NvidiaBertDataset(input_file: str)

Bases: Dataset[Sample]

NVIDIA BERT dataset.

Like the PyTorch Dataset, this dataset is indexable returning a Sample.

Example

>>> from llm.datasets.bert import NvidiaBertDataset
>>> dataset = NvidiaBertDataset('/path/to/shard')
>>> dataset[5]
Sample(...)

Parameters:

• input_file (str) –

  HDF5 file to load.

Source code in llm/datasets/bert.py

def __init__(self, input_file: str) -> None:
    self.input_file = input_file

    # We just inspect the file enough to find the size of the dataset
    # then defer loading until we actually use the dataset. This is
    # particularly helpful for the DistributedShardedDataset which needs
    # to get the length of each shard
    self.loaded = False
    with h5py.File(self.input_file, 'r') as f:
        self.samples = len(f['next_sentence_labels'])

get_masked_labels()

get_masked_labels(
    seq_len: int,
    masked_lm_positions: list[int],
    masked_lm_ids: list[int],
) -> np.ndarray[Any, Any]

Create masked labels array.

Parameters:

• seq_len (int) –

  Sequence length.

• masked_lm_positions (list[int]) –

  Index in sequence of masked tokens

• masked_lm_ids (list[int]) –

  True token value for each position in masked_lm_position.

Returns:

• ndarray[Any, Any] –

  List with length seq_len with the true value for each corresponding masked token in input_ids and -100 for all tokens in input_ids which are not masked.

Source code in llm/datasets/bert.py

def get_masked_labels(
    seq_len: int,
    masked_lm_positions: list[int],
    masked_lm_ids: list[int],
) -> np.ndarray[Any, Any]:
    """Create masked labels array.

    Args:
        seq_len: Sequence length.
        masked_lm_positions: Index in sequence of masked tokens
        masked_lm_ids: True token value for each position in
            `masked_lm_position`.

    Returns:
        List with length `seq_len` with the true value for each corresponding \
        masked token in `input_ids` and -100 for all tokens in input_ids \
        which are not masked.
    """
    masked_lm_labels = np.ones(seq_len, dtype=np.int64) * -100
    if len(masked_lm_positions) > 0:
        # store number of  masked tokens in index
        (padded_mask_indices,) = np.nonzero(np.array(masked_lm_positions) == 0)
        if len(padded_mask_indices) != 0:
            index = padded_mask_indices[0]
        else:
            index = len(masked_lm_positions)
        masked_lm_labels[masked_lm_positions[:index]] = masked_lm_ids[:index]
    return masked_lm_labels

get_dataloader_from_nvidia_bert_shard()

get_dataloader_from_nvidia_bert_shard(
    input_file: str,
    batch_size: int,
    *,
    num_replicas: int | None = None,
    rank: int | None = None,
    seed: int = 0,
    num_workers: int = 4
) -> DataLoader[Batch]

Create a dataloader from a dataset shard.

Parameters:

• input_file (str) –

  HDF5 file to load.

• batch_size (int) –

  Size of batches yielded by the dataloader.

• num_replicas (int | None, default: None ) –

  Number of processes participating in distributed training.

• rank (int | None, default: None ) –

  Rank of the current process within num_replicas.

• seed (int, default: 0 ) –

  Random seed used to shuffle the sampler.

• num_workers (int, default: 4 ) –

  Number of subprocesses to use for data loading.

Returns:

• DataLoader[Batch] –

  Dataloader which can be iterated over to yield Batch.

Source code in llm/datasets/bert.py

def get_dataloader_from_nvidia_bert_shard(
    input_file: str,
    batch_size: int,
    *,
    num_replicas: int | None = None,
    rank: int | None = None,
    seed: int = 0,
    num_workers: int = 4,
) -> DataLoader[Batch]:
    """Create a dataloader from a dataset shard.

    Args:
        input_file: HDF5 file to load.
        batch_size: Size of batches yielded by the dataloader.
        num_replicas: Number of processes participating in distributed
            training.
        rank: Rank of the current process within `num_replicas`.
        seed: Random seed used to shuffle the sampler.
        num_workers: Number of subprocesses to use for data loading.

    Returns:
        Dataloader which can be iterated over to yield \
        [`Batch`][llm.datasets.bert.Batch].
    """
    dataset = NvidiaBertDataset(input_file)
    sampler: DistributedSampler[int] = DistributedSampler(
        dataset,
        num_replicas=num_replicas,
        rank=rank,
        shuffle=True,
        seed=seed,
    )
    dataloader = DataLoader(
        dataset,
        sampler=sampler,
        batch_size=batch_size,
        num_workers=4,
        worker_init_fn=WorkerInitObj(seed),
        pin_memory=True,
    )

    return dataloader

sharded_nvidia_bert_dataset()

sharded_nvidia_bert_dataset(
    input_dir: str,
    batch_size: int,
    *,
    num_replicas: int | None = None,
    rank: int | None = None,
    seed: int = 0,
    num_workers: int = 4
) -> Generator[Batch, None, None]

Simple generator which yields pretraining batches.

Parameters:

• input_dir (str) –

  Directory of HDF5 shards to load samples from.

• batch_size (int) –

  Size of batches yielded.

• num_replicas (int | None, default: None ) –

  Number of processes participating in distributed training.

• rank (int | None, default: None ) –

  Rank of the current process within num_replicas.

• seed (int, default: 0 ) –

  Random seed used to shuffle the sampler.

• num_workers (int, default: 4 ) –

  Number of subprocesses to use for data loading.

Yields:

• Batch –

  Batches of pretraining data.

Source code in llm/datasets/bert.py

def sharded_nvidia_bert_dataset(
    input_dir: str,
    batch_size: int,
    *,
    num_replicas: int | None = None,
    rank: int | None = None,
    seed: int = 0,
    num_workers: int = 4,
) -> Generator[Batch, None, None]:
    """Simple generator which yields pretraining batches.

    Args:
        input_dir: Directory of HDF5 shards to load samples from.
        batch_size: Size of batches yielded.
        num_replicas: Number of processes participating in distributed
            training.
        rank: Rank of the current process within `num_replicas`.
        seed: Random seed used to shuffle the sampler.
        num_workers: Number of subprocesses to use for data loading.

    Yields:
        Batches of pretraining data.
    """
    shard_filepaths = get_filepaths(
        input_dir,
        extensions=['.h5', '.hdf5'],
        recursive=True,
    )
    shard_filepaths.sort()

    for shard_filepath in shard_filepaths:
        dataloader = get_dataloader_from_nvidia_bert_shard(
            shard_filepath,
            batch_size,
            num_replicas=num_replicas,
            rank=rank,
            seed=seed,
        )
        for batch in dataloader:
            yield Batch(*batch)