Lodestar.Embeddings 0.7.0. This page is frozen at that release. Read the current documentation for what main says now. A link to a decision or a migration page follows main, and leaves the archive.

The masked mean of every sequence in a padded batch.

public static float[][] MeanPoolBatch(ReadOnlySpan<float> tokenEmbeddings, int batchSize, int seqLen, int dim, ReadOnlySpan<long> attentionMask)

Parameters — tokenEmbeddings is row-major [batchSize × seqLen × dim] and attentionMask is row-major [batchSize × seqLen]. batchSize is the number of sequences, seqLen the padded length of every one, dim the embedding dimension.

Returns — float[][], one dim-length vector per sequence, in batch order. Not normalized.

Exceptions — ArgumentException when the spans do not match the shape the other arguments declare.

Example — two sequences of different real lengths, padded to the same width.

using Lodestar.Embeddings.Pooling;

float[] tokens = { 1f, 1f, 3f, 3f,    5f, 5f, 99f, 99f };   // 2 sequences, 2 positions, dim 2
long[] mask = { 1L, 1L,   1L, 0L };                          // the second is half padding

float[][] pooled = Pooler.MeanPoolBatch(tokens, batchSize: 2, seqLen: 2, dim: 2, mask);
float bothTokens = pooled[0][0];  // => 2
float oneToken = pooled[1][0];  // => 5

Remarks — Each sequence is pooled against its own slice of the mask, so the padding a shorter sequence carries cannot reach its vector — and, just as importantly, cannot reach its neighbour's. The first sequence averages 1 and 3 to 2; the second has one real token and comes back as 5 rather than 52.

This is the shape an encoder actually returns for a batch, so it is the call that avoids slicing the tensor by hand. Doing that slicing yourself is where off-by-one errors in seqLen produce vectors that are wrong without being obviously wrong.

The batched counterpart of MeanPool, and usually a step on the way to MeanPoolAndNormalizeBatch.

Applies to — net10.0, netstandard2.0.

See also — Pooler.MeanPoolAndNormalizeBatch, Pooler.MeanPool, Pooler.