Popularity
48 Stars
Updated Last
1 Year Ago
Started In
November 2020

ReversePropagation.jl

A Julia package for reverse propagation along a syntax tree, using source-to-source transformation via Symbolics.jl.

Basic usage: Reverse-mode automatic differentiation

The gradient function calculates the gradient of an expression or function with respect to given variables:

julia> using Symbolics, ReversePropagation

julia> f( (x, y) ) = x + (x * y);

julia> vars = @variables x, y;

julia> ∇f = ReversePropagation.gradient(f, vars);

julia> ∇f( (1, 2) )
(3, (3, 1))

The gradient function returns both the value of the function and the gradient.

Basic usage: Forward–backward contractor (interval constraint propagation)

The forward–backward contractor corresponding to an expression takes a box and tries to exclude parts of the box that do not satisfy a constraint.

The contractor is constructed from a symbolic version of the constraint expression:

julia> vars = @variables x, y 
    
julia> ex = x^2 + y^2
    
julia> C = forward_backward_contractor(ex, vars)  # construct the contractor

julia> using IntervalArithmetic

julia> constraint = 0..1

julia> X = IntervalBox(-10..10, 2)

julia> C(X, constraint)

Here the contractor corresponds to the constraint expression x^2 + y^2.

The result of the final call tries to exclude regions of the input box X that do not satisfy x^2 + y^2 ∈ 0..1, where 0..1 denotes the interval [0, 1]. This call returns the contracted box, as well as the value of the original function over the input box.

Parameters may be included in the expression; their symbolic expressions must be passed in when constructing the contractor, and their numerical values when executing the contraction:

julia> @variables a

julia> ex = x^2 + a * y^2
    
julia> C = forward_backward_contractor(ex, vars, [a])

julia> aa = 1..1  # value of the variable `a` to use

julia> C(X, constraint, aa) == ( (-1..1, -1..1), 0..200 )

Tracing and transformations

The package works by tracing an input Julia function into a Symbolics.jl expression. It then transforms that expression into a static single-assignment (SSA) form, before finally emitting Julia code.

The unexported gradient_code function can be used to inspect this process:

julia> ex = f(vars);  #  x + (x * y)

julia> code, final, gradient_vars = ReversePropagation.gradient_code(ex, vars);

julia> code
7-element Vector{Assignment}:
 Assignment(_a, x*y)
 Assignment(_b, _a + x)
 Assignment(_b̄, 1)
 Assignment(_ā, _b̄)
 Assignment(x̄, _b̄)
 Assignment(x̄, x̄ + _ā*y)
 Assignment(ȳ, _ā*x)

License

The code is licensed under the MIT license.

Copyright: David P. Sanders, 2021