`nexuml_library.layers.feature.projector`¶

nexuml_library.layers.feature.projector ¶

Linear projection layers.

Multi-layer linear projector with optional activation and normalization.

Attributes:

Name	Type	Description
`target_dim`	`int \| None`	Output dimension. If None, uses output_sizes.
`hidden_dim`	`int`	Hidden layer size (for n_layers > 1).
`hidden_dims`	`list[int] \| None`	Explicit hidden-layer sizes. When provided, overrides hidden_dim/n_layers for layer construction.
`n_layers`	`int`	Number of linear layers.
`linear_bias`	`bool`	Whether to use bias.
`activation`	`str \| None`	Dotted class path, e.g. "torch.nn.GELU".
`normalization`	`str \| None`	Dotted class path, e.g. "torch.nn.BatchNorm1d".
`skip_last_activation`	`bool`	Skip activation after the last layer.
`flatten_dims`	`tuple[int, int] \| None`	Optional (start, end) dims to flatten before projection.

Pointwise Conv1d projector for sequence inputs (B, T, C).

Equivalent to Linear but uses Conv1d for better memory access patterns.

Attributes:

Name	Type	Description
`target_dim`	`int \| None`	Output channel dimension. If None, uses output_sizes.
`hidden_dim`	`int`	Hidden size for multi-layer projectors.
`n_layers`	`int`	Number of projection layers.
`bias`	`bool`	Whether to use bias.
`activation`	`str \| None`	Dotted class path for activation.
`normalization`	`str \| None`	Dotted class path for normalization.

Low-rank factorized projection W ≈ U @ V.

Reduces parameters from O(CD) to O((C+D)r) when r << min(C, D).

Attributes:

Name	Type	Description
`target_dim`	`int \| None`	Output dimension. If None, uses output_sizes.
`rank`	`int`	Bottleneck rank.
`bias`	`bool`	Whether to use bias.
`activation`	`str \| None`	Optional activation between U and V.