Struct NabLayer

pub struct NabLayer {
    pub weights: Option<NDArray>,
    pub biases: Option<NDArray>,
    pub weight_gradients: Option<NDArray>,
    pub bias_gradients: Option<NDArray>,
    pub node_index: Option<usize>,
    /* private fields */
}

Represents a layer’s configuration and state

Fields

weights: Option<NDArray>

Layer weights (if any)

biases: Option<NDArray>

Layer biases (if any)

weight_gradients: Option<NDArray>

Weight gradients for optimization

bias_gradients: Option<NDArray>

Bias gradients for optimization

node_index: Option<usize>

Implementations

impl NabLayer

pub fn input(shape: Vec<usize>, name: Option<&str>) -> Self

Creates a new Input layer

Arguments

• shape - Shape of the input (excluding batch dimension)
• name - Optional name for the layer

Example

use nabla_ml::nab_layers::NabLayer;

let input_layer = NabLayer::input(vec![784], Some("mnist_input"));
assert_eq!(input_layer.get_output_shape(), &[784]);

pub fn dense( input_dim: usize, units: usize, activation: Option<&str>, name: Option<&str>, ) -> Self

Creates a new Dense (fully connected) layer

Arguments

• input_dim - Number of input features
• units - Number of output units
• activation - Optional activation function (“relu”, “sigmoid”, “tanh”, etc.)
• name - Optional name for the layer

Example

use nabla_ml::nab_layers::NabLayer;

// Dense layer with ReLU activation
let dense = NabLayer::dense(784, 128, Some("relu"), Some("hidden_1"));

pub fn activation( activation_type: &str, input_shape: Vec<usize>, name: Option<&str>, ) -> Self

Creates a new Activation layer

Arguments

• activation_type - Type of activation (“relu”, “sigmoid”, “tanh”, etc.)
• input_shape - Shape of the input (excluding batch dimension)
• name - Optional name for the layer

Example

use nabla_ml::nab_layers::NabLayer;

let relu = NabLayer::activation("relu", vec![128], Some("relu_1"));
assert_eq!(relu.get_output_shape(), &[128]);

pub fn flatten(input_shape: Vec<usize>, name: Option<&str>) -> Self

Creates a new Flatten layer

Flattens the input while keeping the batch size. For example: (batch_size, height, width, channels) -> (batch_size, height * width * channels)

Arguments

• input_shape - Shape of the input (excluding batch dimension)
• name - Optional name for the layer

Example

use nabla_ml::nab_layers::NabLayer;

// Flatten a 28x28x1 image to 784 features
let flatten = NabLayer::flatten(vec![28, 28, 1], Some("flatten_1"));
assert_eq!(flatten.get_output_shape(), &[784]);

pub fn dropout(input_shape: Vec<usize>, rate: f64, name: Option<&str>) -> Self

Creates a new Dropout layer

Randomly sets input units to 0 with a probability of rate during training. During inference (training=false), the layer behaves like an identity function.

Arguments

• input_shape - Shape of the input (excluding batch dimension)
• rate - Dropout rate between 0 and 1 (e.g., 0.5 means 50% of units are dropped)
• name - Optional name for the layer

Example

use nabla_ml::nab_layers::NabLayer;

// Dropout with 50% rate
let dropout = NabLayer::dropout(vec![128], 0.5, Some("dropout_1"));
assert_eq!(dropout.get_output_shape(), &[128]);

pub fn batch_norm( input_shape: Vec<usize>, epsilon: Option<f64>, momentum: Option<f64>, name: Option<&str>, ) -> Self

Creates a new BatchNormalization layer

Normalizes the activations of the previous layer for each batch. During training, uses batch statistics. During inference, uses running statistics.

Arguments

• input_shape - Shape of the input (excluding batch dimension)
• epsilon - Small constant for numerical stability (default: 1e-5)
• momentum - Momentum for running statistics (default: 0.99)
• name - Optional name for the layer

Example

use nabla_ml::nab_layers::NabLayer;

let bn = NabLayer::batch_norm(vec![128], None, None, Some("bn_1"));
assert_eq!(bn.get_output_shape(), &[128]);

pub fn forward(&mut self, input: &NDArray, training: bool) -> NDArray

Forward pass through the layer

pub fn backward(&mut self, gradient: &NDArray) -> NDArray

Backward pass for the layer

pub fn get_output_shape(&self) -> &[usize]

Returns the output shape of the layer

pub fn get_name(&self) -> &str

Returns the name of the layer

pub fn is_trainable(&self) -> bool

Returns whether the layer is trainable

pub fn compute_output_shape(&self, input_shape: &[usize]) -> Vec<usize>

Computes output shape for a given input shape

pub fn set_node_index(&mut self, index: usize)

pub fn set_inputs(&mut self, inputs: Vec<usize>)

Trait Implementations

impl Clone for NabLayer

fn clone(&self) -> NabLayer

Returns a copy of the value.

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

Performs copy-assignment from source.