Struct MLPFeatureExtractor 

pub struct MLPFeatureExtractor { /* private fields */ }

Multi-layer perceptron feature extractor.

The network is a sequence of DenseLayers applied in order; the output of layer i is the input of layer i + 1. The layer stack must be non-empty and layer shapes must match transitively.

Parameters are stored twice on purpose:

• as structured layers: Vec<DenseLayer> — used by forward.
• as flat parameters: Vec<f64> — exposed to optimisers.

The two views are kept in sync: MLPFeatureExtractor::parameters_mut returns a borrow into the flat buffer and MLPFeatureExtractor::sync_from_flat pushes the flat buffer back into the layer weights. Mutating the flat buffer directly requires a subsequent sync_from_flat call before the next forward.

Implementations

impl MLPFeatureExtractor

pub fn from_layers(layers: Vec<DenseLayer>) -> Result<Self>

Wrap an existing Vec<DenseLayer> as an MLP feature extractor.

Fails when the layer stack is empty or when consecutive layer shapes do not match.

pub fn xavier_init( widths: &[usize], activations: &[Activation], seed: u64, ) -> Result<Self>

Build an MLP from a list of layer widths and a parallel list of activations (one per weight matrix — i.e. widths.len() - 1 entries). Weights are Xavier/Glorot-normal initialised via SciRS2-Core’s seeded RNG; biases are zero.

pub fn layers(&self) -> &[DenseLayer]

Immutable view of the layer stack.

pub fn num_layers(&self) -> usize

Number of layers.

pub fn forward_with_cache(&self, input: &[f64]) -> Result<ForwardCache>

Forward pass with per-layer caches of (pre_activation, post_activation) tensors. Used by the analytical gradient path in crate::deep_kernel::gradient.

pub fn sync_from_flat(&mut self) -> Result<()>

Push the flat parameter buffer back into the per-layer weights / biases. Call this after mutating the flat buffer returned from Self::parameters_mut but before the next forward pass.

Trait Implementations

impl Clone for MLPFeatureExtractor

fn clone(&self) -> MLPFeatureExtractor

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for MLPFeatureExtractor

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl NeuralFeatureMap for MLPFeatureExtractor

fn forward(&self, input: &[f64]) -> Result<Vec<f64>>

Map an input vector to feature space.

fn parameters_mut(&mut self) -> &mut [f64]

Mutable view of the flat parameter vector (weights + biases, in layer order). Exposed as &mut [f64] so that optimisers can apply updates in place without owning the extractor.

fn parameters(&self) -> &[f64]

Immutable view of the flat parameter vector.

fn parameter_count(&self) -> usize

Number of trainable scalar parameters.

fn input_dim(&self) -> usize

Input dimension expected by forward.

fn output_dim(&self) -> usize

Output dimension produced by forward.