#include <grad_def.hpp>

Inheritance diagram for teq::iGradientBuilder:

Public Member Functions
virtual	~iGradientBuilder (void)=default

virtual TensptrT	local_derivative (FuncptrT op, size_t arg_idx) const =0

virtual TensptrT	chain_rule (FuncptrT op, const TensptrT &local_der, TensptrT supcomp_grad, size_t arg_idx) const =0

virtual TensptrT	get_const_one (Shape shape) const =0
	Return tensor representing 1 constant. More...

virtual TensptrT	get_const_zero (Shape shape) const =0
	Return tensor representing 0 constant. More...

virtual TensptrT	add (TensptrT &lhs, TensptrT &rhs) const =0
	Return functor representing lhs + rhs. More...

TensptrT	derive (TensptrT root, TensptrT target) const
	Return derivative of root with respect to target. More...

Detailed Description

Define manditory definitions required for tensor differentiation For some graph F(G(x)), chain rule for calculating dF/dx is defined in the following order:

calcualte dF/dG => F local derivative and derivative of super composition (supcomp_grad for G)
calculate dG/dx => G local derivative
chain dF/dG (supcomp_grad) and dG/dx (local_der) This top-down approach updates tensor shape information such that output derivative dF/dx has the shape of x

Constructor & Destructor Documentation

◆ ~iGradientBuilder()

virtual teq::iGradientBuilder::~iGradientBuilder ( void )

virtualdefault

Member Function Documentation

◆ add()

virtual TensptrT teq::iGradientBuilder::add	(	TensptrT &	lhs,
		TensptrT &	rhs
	)		const

pure virtual

Return functor representing lhs + rhs.

Implemented in eteq::GradientBuilder< T >.

◆ chain_rule()

virtual TensptrT teq::iGradientBuilder::chain_rule	(	FuncptrT	op,
		const TensptrT &	local_der,
		TensptrT	supcomp_grad,
		size_t	arg_idx
	)		const

pure virtual

Let op be functor F with arguments args, and local_der is derivative of F wrt one of args (say x) Let supcomp_grad be defined as dG/dF where G is some super-functor using F Return derivative G wrt to arg x by applying chain rule

Implemented in eteq::GradientBuilder< T >.