Struct ReverseAD 

pub struct ReverseAD<F: IntegrateFloat> { /* private fields */ }

Reverse mode automatic differentiation engine

Implementations

impl<F: IntegrateFloat> ReverseAD<F>

pub fn new(nvars: usize) -> Self

Create a new reverse AD engine

pub fn with_checkpoint_strategy(self, strategy: CheckpointStrategy) -> Self

Set checkpointing strategy

pub fn gradient<Func>( &mut self, f: Func, x: ArrayView1<'_, F>, ) -> IntegrateResult<Array1<F>>

where Func: Fn(&mut Tape<F>, &[usize]) -> usize,

Compute gradient using reverse mode AD

pub fn jacobian<Func>( &mut self, f: Func, x: ArrayView1<'_, F>, ) -> IntegrateResult<Array2<F>>

where Func: Fn(&mut Tape<F>, &[usize]) -> Vec<usize>,

Compute Jacobian using reverse mode AD

pub fn hessian<Func>( &mut self, f: Func, x: ArrayView1<'_, F>, ) -> IntegrateResult<Array2<F>>

where Func: Fn(&mut Tape<F>, &[usize]) -> usize + Clone,

Compute Hessian (second derivatives) using reverse-over-forward AD

pub fn batch_gradient<Func>( &mut self, f: Func, x_batch: &[Array1<F>], ) -> IntegrateResult<Vec<Array1<F>>>

where Func: Fn(&mut Tape<F>, &[usize]) -> usize + Clone,

Compute gradients for multiple inputs in batch

pub fn jvp<Func>( &mut self, f: Func, x: ArrayView1<'_, F>, v: ArrayView1<'_, F>, ) -> IntegrateResult<Array1<F>>

where Func: Fn(&mut Tape<F>, &[usize]) -> Vec<usize>,

Compute Jacobian-vector product efficiently without forming full Jacobian

pub fn vjp<Func>( &mut self, f: Func, x: ArrayView1<'_, F>, v: ArrayView1<'_, F>, ) -> IntegrateResult<Array1<F>>

where Func: Fn(&mut Tape<F>, &[usize]) -> Vec<usize>,

Compute vector-Jacobian product (useful for backpropagation)