Downhill.Linear.BackGrad

Documentation

newtype BackGrad a v

Linear expression, made for backpropagation. It is similar to Expr BackFun, but has a more flexible form.

Constructors

BackGrad (forall x. (x -> VecBuilder v) -> Term a x)

Instances

(BasicVector v, AdditiveGroup v) => AdditiveGroup (BackGrad r v)

Defined in Downhill.Linear.BackGrad

Methods

zeroV :: BackGrad r v

(^+^) :: BackGrad r v -> BackGrad r v -> BackGrad r v

negateV :: BackGrad r v -> BackGrad r v

(^-^) :: BackGrad r v -> BackGrad r v -> BackGrad r v

(BasicVector v, VectorSpace v) => VectorSpace (BackGrad r v)

Defined in Downhill.Linear.BackGrad

Associated Types

type Scalar (BackGrad r v)

Methods

(*^) :: Scalar (BackGrad r v) -> BackGrad r v -> BackGrad r v

type Scalar (BackGrad r v)

Defined in Downhill.Linear.BackGrad

type Scalar (BackGrad r v) = Scalar v

realNode :: Expr a v -> BackGrad a v

Creates a BackGrad that is backed by a real node. Gradient of type v will be computed and stored in a graph for this node.

inlineNode :: forall r u v. (VecBuilder v -> VecBuilder u) -> BackGrad r u -> BackGrad r v

inlineNode f x will apply function f to variable x without creating a node. All of the gradients coming to this expression will be forwarded to the parents of x. However, if this expression is used more than once, f will be evaluated multiple times, too. It is intended to be used for newtype wrappers. inlineNode f x also doesn't prevent compiler to inline and optimize x

sparseNode :: forall r a z. BasicVector z => (VecBuilder z -> VecBuilder a) -> BackGrad r a -> BackGrad r z

castBackGrad :: forall r v z. VecBuilder z ~ VecBuilder v => BackGrad r v -> BackGrad r z

BackGrad doesn't track the type of the node. Type of BackGrad can be changed freely as long as VecBuilder stays the same.