Method pad_sequence

modelscope/preprocessors/templates/base.py:867–893

Pad sequence by some side Args: sequences: The input sequences in tensor. padding_value: The padding value padding_side: The padding side Returns: A tensor after padding

(sequences: List[torch.Tensor],
                     padding_value: float = 0.,
                     padding_side: Literal['right', 'left'] = 'right')

    @staticmethod
    def pad_sequence(sequences: List[torch.Tensor],
                     padding_value: float = 0.,
                     padding_side: Literal['right', 'left'] = 'right') -> torch.Tensor:
        """Pad sequence by some side

        Args:
            sequences: The input sequences in tensor.
            padding_value: The padding value
            padding_side: The padding side

        Returns:
            A tensor after padding
        """
        padding_right = padding_side == 'right'
        if padding_right:
            return pad_sequence(sequences, batch_first=True, padding_value=padding_value)

        max_len = max([s.size(0) for s in sequences])

        padded_sequences = []
        for seq in sequences:
            pad_length = max_len - seq.size(0)
            pad_tuple = [0] * ((seq.dim() - 1) * 2) + [pad_length, 0]
            padded_seq = F.pad(seq, tuple(pad_tuple), 'constant', padding_value)
            padded_sequences.append(padded_seq)

        return torch.stack(padded_sequences)

    def data_collator(self, batch: List[Dict[str, Any]], padding_to: Optional[int] = None) -> Dict[str, Any]:
        """