Struct SparseSgd 

pub struct SparseSgd { /* private fields */ }

Sparse SGD optimizer over a 1-D dense parameter vector.

With momentum == 0 and weight_decay == 0 this is the canonical sparse update params[i] -= lr * grad[i], touching only the non-zero coordinates and leaving every other coordinate bit-identical.

Implementations

impl SparseSgd

pub fn new(config: SparseSgdConfig) -> Self

Create a new optimizer. State (if momentum is used) is allocated lazily.

pub fn config(&self) -> &SparseSgdConfig

Immutable view of the configuration.

pub fn step( &mut self, params: &mut Array1<f64>, grad: &CooGradient, ) -> Result<(), GpuOptimError>

Perform one sparse SGD update in place, touching only the coordinates present in grad.

Errors

Returns GpuOptimError::DimensionMismatch on dimension disagreement and propagates gradient validation errors.

Trait Implementations

impl Clone for SparseSgd

fn clone(&self) -> SparseSgd

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for SparseSgd

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

Formats the value using the given formatter.