PackedSequenceCollateFunction<T>

Models & Types · AiDotNet.Data.Collation

Packs variable-length sequences into a contiguous tensor without padding, along with sequence lengths for reconstruction.

Implements: ICollateFunction<Tensor<T>, PackedSequenceBatch<T>>

How It Works

Packed sequences are more memory-efficient than padded batches because they don’t store padding tokens. The output contains a flat data tensor of all sequences concatenated, plus a lengths tensor indicating how many elements belong to each sequence. This is equivalent to PyTorch’s pack_padded_sequence.

Sequences are sorted by length (descending) for efficient RNN processing.

Constructors

Constructor Summary
PackedSequenceCollateFunction(bool sortByLength) Creates a new packed sequence collate function.

Methods

Method Summary
PackedSequenceBatch<T> Collate(IReadOnlyList<Tensor<T>> samples)