espnet2.speechlm.utils.data.pad_list

source

espnet2.speechlm.utils.data.pad_list(sequences: List[ndarray | Tensor], pad_value: float = 0.0) → Tuple[Tensor, Tensor]

Pad a list of sequences to the same length and stack them.

Uses right padding (padding at the end of sequences). Assumes the FIRST dimension is the time/sequence length.

• Parameters:
  • sequences – List of sequences to pad and stack. Each can be of any shape [seq_len, …] where seq_len is variable
  • pad_value – Value to use for padding (default: 0.0)
• Returns:
  • Padded and stacked tensor of shape [batch, max_seq_len, …]
  • Length tensor of shape [batch] with original sequence lengths (dtype=long)
• Return type: Tuple of
• Raises:
  • ValueError – If sequence list is empty or sequences have inconsistent shapes
  • TypeError – If sequences contain non-tensor/non-array types