mask_from_seq_lengths.py

import torch


def mask_from_seq_lengths(
    sequence_lengths: torch.Tensor, 
    max_length: int
) -> torch.BoolTensor:
    """
    our input was `[2, 2, 3]`, with a `max_length` of 4, we'd return
    `[[1, 1, 0, 0], [1, 1, 0, 0], [1, 1, 1, 0]]`.
    """
    # (batch_size, max_length)
    ones = sequence_lengths.new_ones(sequence_lengths.size(0), max_length)
    range_tensor = ones.cumsum(dim=1)
    return sequence_lengths.unsqueeze(1) >= range_tensor