Stdlib

std.aad

Helpers for tape-based reverse-mode automatic differentiation.

(import std.aad)

Gradients

Symbol	Description
`(grad f xs)`	Gradient of scalar-valued `f` at point `xs`.
`(check-grad f xs [tol])`	Numerical vs. analytical gradient check; returns boolean.
`(check-grad-report f xs [tol])`	Same as above, returns a diagnostic record.
`(with-checkpoint thunk)`	Run `thunk` inside a tape checkpoint.

These work on AAD nodes and forward to plain numbers when possible.

Symbol	Description
`(ad-abs x)`	Subgradient absolute value.
`(ad-max x y)`	Subgradient maximum.
`(ad-min x y)`	Subgradient minimum.
`(ad-relu x)`	ReLU activation.
`(ad-clamp x lo hi)`	Clamp with subgradients.
`(softplus x)`	Smooth approximation of ReLU.
`(smooth-abs x epsilon)`	Smooth absolute value.
`(smooth-clamp x lo hi epsilon)`	Smooth clamp.