nn.utils.rnn.pack_padded_sequence¶

lucid.nn.utils.rnn.pack_padded_sequence(input_: Tensor, lengths: Sequence[int] | Tensor, batch_first: bool = False, enforce_sorted: bool = True) → PackedSequence¶

Function Signature¶

def pack_padded_sequence(
    input_: Tensor,
    lengths: Sequence[int] | Tensor,
    batch_first: bool = False,
    enforce_sorted: bool = True,
) -> PackedSequence

Parameters¶

input_ (Tensor): A padded tensor of shape (seq_len, batch, feature) or (batch, seq_len, feature) when batch_first=True.
lengths (Sequence[int] | Tensor): Sequence lengths for each batch element. Must be 1D and match the batch size.
batch_first (bool, optional): If True, input_ is interpreted as (batch, seq_len, feature). Default is False.
enforce_sorted (bool, optional): If True, lengths must be sorted in decreasing order. If False, the function will sort internally and populate sorted_indices/unsorted_indices. Default is True.

Return Value¶

PackedSequence: A packed representation containing data, batch_sizes, and optional sort/unsort indices.

Behavior¶

The maximum sequence length is inferred from lengths.
The returned batch_sizes is a 1D tensor of length max_len.
When enforce_sorted=False, the input is sorted by length internally.

Examples¶

import lucid
import lucid.nn as nn

x = lucid.randn(5, 3, 4)  # (seq_len=5, batch=3, feature=4)
lengths = [5, 3, 2]

packed = nn.utils.rnn.pack_padded_sequence(x, lengths)
out, h_n = nn.RNN(4, 6)(packed)