Development build. This page describes main, not a released package. The latest published Lodestar.Embeddings is 0.8.0 — read its documentation.

Home › Embeddings › Pooling

Scales a vector in place to unit L2 norm.

public static void L2Normalize(Span<float> vector)

Parameters — vector is the vector to scale, modified in place. A float[] converts implicitly.

Returns — nothing — the argument is the result.

Exceptions — none. A zero vector stays zero rather than producing NaN.

Example — the 3-4-5 triangle, scaled to unit length.

using Lodestar.Embeddings.Pooling;

float[] v = { 3f, 4f };
Pooler.L2Normalize(v);

float x = v[0];  // => 0.6
float y = v[1];  // => 0.8

Remarks — This is the one place in the package that refuses to vectorize on purpose. The sum of squares is accumulated in double and computed by a scalar loop: a Vector<float> accumulator would lose precision, and it would make the answer depend on the SIMD width of the machine that ran it. The scaling pass afterwards is exact whichever way it is done, so that half is vectorized.

The consequence is worth stating plainly: a pooled vector is bit-identical across net10.0, netstandard2.0, and machines with different vector widths. That is the opposite trade from VectorMath.Dot, where the SIMD accumulation is the point and the last bits may differ.

The divisor is never below 1e-12, as in torch.nn.functional.normalize(v, p=2, dim=1), which this matches: a vector whose norm is under that floor is scaled by 1e12 rather than to unit length, and a zero vector stays zero, not NaN. That matters because an all-padding sequence pools to zero, and a NaN there would spread into every score it ever touched.

Applies to — net10.0, netstandard2.0.

See also — Pooler.MeanPoolAndNormalize, Pooler, the pooling index.