Struct AdaptiveOptimizer 

pub struct AdaptiveOptimizer {
    pub algorithm: OptimizerAlgorithm,
    pub states: HashMap<String, OptimizerState>,
    pub global_step: u64,
    /* private fields */
}

Adaptive gradient optimizer supporting Adam, AdaGrad, RMSProp, and AdamW.

Example

use ipfrs_tensorlogic::adaptive_optimizer::{
    AdaptiveOptimizer, OptimizerAlgorithm, ParameterGroup,
};

let mut opt = AdaptiveOptimizer::new(OptimizerAlgorithm::adam_default());
let mut groups = vec![
    ParameterGroup::with_grad("w", vec![0.5, -0.3], vec![0.1, -0.2]),
];
let norm = opt.step(&mut groups).expect("example: should succeed in docs");
assert!(norm > 0.0);

Fields

algorithm: OptimizerAlgorithm

The algorithm (and its hyper-parameters) used for all steps.

states: HashMap<String, OptimizerState>

Per-group optimizer states; keyed by ParameterGroup::name.

global_step: u64

Global step counter (incremented once per call to step).

Implementations

impl AdaptiveOptimizer

pub fn new(algorithm: OptimizerAlgorithm) -> Self

Create a new optimizer wrapping the given algorithm.

pub fn step( &mut self, groups: &mut [ParameterGroup], ) -> Result<f64, OptimizerError>

Perform one optimizer step across all groups.

Returns the global L2 gradient norm computed before any update.

Errors

Returns OptimizerError::DimensionMismatch if params.len() != grad.len() for any group, or OptimizerError::EmptyGroup if a group has zero parameters.

pub fn step_group( &mut self, group: &mut ParameterGroup, ) -> Result<(), OptimizerError>

Perform one optimizer step for a single parameter group.

This is the low-level primitive used by step.

Errors

Returns OptimizerError::DimensionMismatch or OptimizerError::EmptyGroup.

pub fn zero_grad(groups: &mut [ParameterGroup])

Zero-out all gradients in groups.

pub fn global_grad_norm(groups: &[ParameterGroup]) -> f64

Compute the global L2 norm of all gradients across groups.

pub fn clip_grad_norm(groups: &mut [ParameterGroup], max_norm: f64)

Scale all gradients so the global norm does not exceed max_norm.

If the current global norm is ≤ max_norm or is not finite (e.g. NaN/inf), gradients are left unchanged.

pub fn reset_state(&mut self, group_name: &str)

Clear the optimizer state for a specific group (by name).

pub fn reset_all(&mut self)

Clear all optimizer states and reset the global step counter.

pub fn stats(&self, groups: &[ParameterGroup]) -> OptimizerStats

Return a statistics snapshot.

Trait Implementations

impl Clone for AdaptiveOptimizer

fn clone(&self) -> AdaptiveOptimizer

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for AdaptiveOptimizer

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

Formats the value using the given formatter.