Struct Gradients 

pub struct Gradients { /* private fields */ }

Parameter gradients for an Mlp (overwrite semantics).

Allocate once via Mlp::gradients() and reuse across training steps.

Implementations

impl Gradients

pub fn new(mlp: &Mlp) -> Self

Allocate gradient buffers suitable for mlp.

pub fn d_output_mut(&mut self) -> &mut [f32]

Mutable view of the upstream gradient buffer for the final model output.

Typical training flow:

• mlp.forward(input, &mut scratch)
• loss writes dL/d(output) into grads.d_output_mut()
• mlp.backward(input, &scratch, &mut grads)

pub fn d_output(&self) -> &[f32]

Immutable view of the final upstream gradient buffer.

pub fn d_input(&self) -> &[f32]

Returns dL/d(input) computed by the most recent backward call.

pub fn d_weights(&self, layer_idx: usize) -> &[f32]

Returns the weight gradient for the given layer (row-major (out_dim, in_dim)).

pub fn d_weights_mut(&mut self, layer_idx: usize) -> &mut [f32]

Mutable weight gradient for the given layer.

pub fn d_biases(&self, layer_idx: usize) -> &[f32]

Returns the bias gradient for the given layer (length out_dim).

pub fn d_biases_mut(&mut self, layer_idx: usize) -> &mut [f32]

Mutable bias gradient for the given layer.

pub fn zero_params(&mut self)

Zero the parameter gradient buffers (d_weights and d_biases).

pub fn scale_params(&mut self, scale: f32)

Scale parameter gradients (d_weights and d_biases) in place.

pub fn global_l2_norm_params(&self) -> f32

Compute the global L2 norm of parameter gradients.

pub fn clip_global_norm_params(&mut self, max_norm: f32) -> f32

Clip parameter gradients by global norm.

Returns the pre-clip norm.

Trait Implementations

impl Clone for Gradients

fn clone(&self) -> Gradients

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for Gradients

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

Formats the value using the given formatter.