valueAndGradFns

valueAndGradFns(f, options?): object

Split an objective into the SEPARATE value and gradient functions that an API taking a (fn, gradFn) pair expects — @tangent.to/mc’s model.potential(name, fn, gradFn) is the case this exists for.

The two share one evaluation: calling .value(p) then .gradient(p) on the same parameters runs the tape once, not twice. That matters because a sampler’s value-and-gradient path calls both in turn, and the forward pass is a full sweep over the data.

The cache holds exactly one entry and compares parameters structurally against a defensive copy, so mutating a parameter array in place invalidates it correctly rather than returning a stale gradient. A call that passes inputs bypasses the cache: the same parameters on a different batch are a different evaluation, and copying a batch to compare it would cost what the cache saves.

Parameters

f

(x) => Var

objective built from this package’s ops

options?

compile?

boolean

build the tape once and replay it, via compile. Worth an order of magnitude on a sampler, which calls this thousands of times at the same shapes; read compile’s constraint before turning it on. Off by default: a static graph is an assumption about your objective, and one this package cannot check for you.

Returns

object

With compile: true, compiled is the underlying compile closure, so its toJSON() is reachable: what lets a model send its likelihood to a worker as data.

compiled?

optional compiled?: Function

gradient

gradient: (x) => any

Parameters

x

any

Returns

any

value

value: (x) => number

Parameters

x

any

Returns

number

Example

const { value, gradient } = valueAndGradFns((p) => logLik(p), { compile: true });
model.potential('y', value, gradient);