`nexuml_library.layers.feature.pooling`¶

nexuml_library.layers.feature.pooling ¶

Pooling layers for token/sequence aggregation.

Pool tokens along a specified dimension.

Attributes:

Name	Type	Description
`dim`	`int`	Dimension to pool over (negative indexing supported, e.g. -2 for time).
`pooling_type`	`Literal['mean', 'max', 'std', 'mean_std', 'cls', 'multilayer']`	"mean": average over the pool dim "max": max over the pool dim "std": standard deviation over the pool dim "mean_std" (alias: "mean+std", "mean_plus_std"): concatenate mean and std along the last axis "cls": select the first index along the pool dim (CLS-style) "multilayer": when the input is multi-layer hidden states (e.g. `[B, L, N, D]` from an encoder that exposes all hidden states), concatenate per-layer mean+std along the last axis, preserving the per-layer ordering.
`remove_dim`	`bool`	Whether to squeeze the pooled dimension.
`skip_first`	`bool`	Skip the first token (e.g. CLS token) before pooling.

Learned attention pooling using a query token.

Attributes:

Name	Type	Description
`dim`	`int`	Token dimension to pool over.
`remove_dim`	`bool`	Whether to remove the pooled dimension.
`n_heads`	`int`	Number of attention heads.

Attentive Statistics Pooling (Okabe et al., 2018).

Computes attention-weighted mean and std, then projects to input dim.

Attributes:

Name	Type	Description
`dim`	`int`	Dimension to pool over.
`remove_dim`	`bool`	Whether to remove the pooled dimension.
`hidden_dim`	`int`	Hidden dimension for the attention MLP.