Struct SparseAdam 

pub struct SparseAdam { /* private fields */ }

Sparse / lazy Adam optimizer over a 1-D dense parameter vector.

Per-coordinate moment state (m, v) and last_step are stored densely (Adam always needs one moment pair per parameter) but are only read and written for coordinates that carry a gradient on a given step, which is what makes the step cost proportional to the number of non-zero coordinates.

Implementations

impl SparseAdam

pub fn new(config: SparseAdamConfig) -> Self

Create a new optimizer. State is allocated lazily on the first SparseAdam::step from the length of the parameter vector.

pub fn global_step(&self) -> usize

The number of optimizer steps performed so far (the global t).

pub fn config(&self) -> &SparseAdamConfig

Immutable view of the configuration.

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

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

Errors

Returns GpuOptimError::DimensionMismatch when the parameter length and the gradient dimension disagree (or differ from a previous step), and propagates validation errors from a malformed gradient.

Trait Implementations

impl Clone for SparseAdam

fn clone(&self) -> SparseAdam

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for SparseAdam

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

Formats the value using the given formatter.