Embeddings pooler

Pooler

Turns per-token model outputs into a single sentence embedding.

public static class Pooler

Example — two token positions, the second one padding, so only the first reaches the result.

using Lodestar.Embeddings.Pooling;

float[] tokens = { 3f, 4f, 99f, 99f };   // two positions, dim 2
long[] mask = { 1L, 0L };                 // the second is padding

float[] pooled = Pooler.MeanPoolAndNormalize(tokens, seqLen: 2, dim: 2, mask);
float x = pooled[0];  // => 0.6
float y = pooled[1];  // => 0.8

Remarks — 99f never reaches the answer, which is what the mask is for: the pooled vector is (3, 4) scaled to unit length, and the padding position is not averaged in.

The recipe is masked mean then L2 normalization — what sentence-transformers' encode does to one forward pass. The two AndNormalize members do both steps and are what most callers want; the others exist so a caller can do one half.

Every member is static and allocates only its result, so all five are safe to call from any number of threads.

Applies to — net10.0, netstandard2.0.

See also — the pooling index, the embeddings guide.

Members

Member	What it does
Pooler.L2Normalize	Scales a vector in place to unit length.
Pooler.MeanPool	The masked mean of one sequence.
Pooler.MeanPoolAndNormalize	The full recipe, one sequence.
Pooler.MeanPoolAndNormalizeBatch	The full recipe, a padded batch.
Pooler.MeanPoolBatch	The masked mean of every sequence in a batch.