nexuml_library.layers.feature.l2_normalize¶
L2-normalization pipeline layer.
A trivial stateless postproc: applies L2 normalization to the last
dimension of the input. Has no trainable state and no fit()/is_fitted
contract — it is an ordinary pipeline composition primitive, not a
train-fitted postprocessor.
This is the canonical "normalize embeddings" step the DCASE 2026 P3
target pipeline uses. It is intentionally a separate layer from
:class:FeaturePostproc so the target scenario does not need a
train-fitted postproc / pre-materialization bridge contract.
Bases: LayerDefinition
L2-normalize the last dimension of the input.
Stateless — no trainable parameters and no fitted state. Equivalent
to torch.nn.functional.normalize(x, p=2, dim=-1, eps=eps).
Attributes:
| Name | Type | Description |
|---|---|---|
eps |
float
|
Small epsilon added to the denominator to avoid division by
zero. Defaults to |