Struct DifferentiableDenseLayer 

pub struct DifferentiableDenseLayer {
    pub layer: DenseLayer,
    pub surrogate: SurrogateType,
    input_cache: Vec<f64>,
    output_cache: Vec<f64>,
}

Dense SC layer with surrogate gradient backward pass.

Fields

layer: DenseLayer

surrogate: SurrogateType

input_cache: Vec<f64>

output_cache: Vec<f64>

Implementations

impl DifferentiableDenseLayer

pub fn new( n_inputs: usize, n_neurons: usize, length: usize, seed: u64, surrogate: SurrogateType, ) -> Self

Construct a differentiable dense SC layer.

pub fn forward( &mut self, input_values: &[f64], seed: u64, ) -> Result<Vec<f64>, String>

Forward pass and cache activations for backward pass.

pub fn backward( &self, grad_output: &[f64], ) -> Result<(Vec<f64>, Vec<Vec<f64>>), String>

Backward pass producing input and weight gradients.

pub fn update_weights(&mut self, weight_grads: &[Vec<f64>], lr: f64)

Apply gradient descent update and clamp weights to [0, 1].

pub fn clear_cache(&mut self)

Clear cached forward tensors.