Pack Padded Sequence Loss

Overview

In image captioning and seq2seq tasks, target sequences have different lengths and are padded. Computing cross-entropy on padding tokens wastes compute and leaks gradient noise. pack_padded_sequence strips out padding positions before loss computation, giving a clean loss signal from real tokens only.

Quick Start

from torch.nn.utils.rnn import pack_padded_sequence

# Sort batch by decreasing sequence length (required by pack_padded_sequence)
caption_lengths, sort_ind = caption_lengths.squeeze(1).sort(dim=0, descending=True)
predictions = predictions[sort_ind]
targets = targets[sort_ind]
decode_lengths = (caption_lengths - 1).tolist()  # exclude <sos>

# Pack to strip padding, then compute loss on real tokens only
packed_preds = pack_padded_sequence(predictions, decode_lengths, batch_first=True).data
packed_targets = pack_padded_sequence(targets, decode_lengths, batch_first=True).data

loss = nn.CrossEntropyLoss()(packed_preds, packed_targets)

Workflow

1. Sort batch by sequence length (descending)
2. Compute decode_lengths = actual length of each target sequence
3. Pack both predictions and targets with pack_padded_sequence
4. Compute cross-entropy on .data attribute (flattened real tokens)
5. Backpropagate — gradients only flow through non-padding positions

Key Decisions

• Sort requirement: pack_padded_sequence requires descending sort; re-sort encoder outputs too
• Alternative: Use ignore_index in CrossEntropyLoss for padding token ID — simpler but slightly less efficient
• Combine with: Teacher forcing during training, greedy/beam search during inference
• Metrics: Compute Levenshtein distance or BLEU on unpacked decoded sequences

References

• InChI / Resnet + LSTM with attention / starter