Struct GradientChecker 

pub struct GradientChecker {
    pub epsilon: f64,
    pub rtol: f64,
    pub atol: f64,
}

Gradient checker for numerical verification

Compares analytical gradients (from parameter-shift rule or adjoint method) with numerical gradients (finite differences) to verify correctness.

Fields

epsilon: f64

Epsilon for finite differences

rtol: f64

Relative tolerance for comparison

atol: f64

Absolute tolerance for comparison

Implementations

impl GradientChecker

pub fn new() -> Self

Create new gradient checker with default tolerances

pub fn with_epsilon(epsilon: f64) -> Self

Create with custom epsilon

pub fn with_tolerances(epsilon: f64, rtol: f64, atol: f64) -> Self

Create with custom tolerances

pub fn numerical_gradient<F>( &self, param: &mut TQParameter, param_idx: usize, loss_fn: &mut F, ) -> Result<f64>

where F: FnMut() -> Result<f64>,

Compute numerical gradient using finite differences

For function f and parameter θ: ∂f/∂θ ≈ [f(θ + ε) - f(θ - ε)] / (2ε)

pub fn check_gradient( &self, analytical: f64, numerical: f64, ) -> GradientCheckResult

Check if analytical and numerical gradients match

Trait Implementations

impl Default for GradientChecker

fn default() -> Self

Returns the “default value” for a type.