Popularity
10 Stars
Updated Last
5 Months Ago
Started In
October 2015

ReverseDiffTape

Build Status

Welcome to the ReverseDiffTape.jl, a Julia package for reverse mode differentiation on a tape.

#Overview This julia package implements reverse mode automatic/algorithmic differentiation algorithm for computing gradient and Hessian of a scalar function. The Hessian evaluation and pattern finding use a state-of-the-art reverse Hessian algorithm, named Edge_Pushing.

#Interface with JuMP throught MathProgBase

  • Delegate the AD evaluation routines to use ReverseDiffTape implementation
include("interface_tape.jl")               
using TapeInterface                            
m = Model(solver=TapeSolver(IpoptSolver())) #creating TapeSolver instance
  • Rest proceeds as the same of a normal JuMP model
@variable(m, -100<=x[1:10]<=100)
@NLobjective(m, Min, sum{ cos(sum{ x[i+j], j=1:3}) + sum{ (x[i] + x[j])^2, j=1:3} , i=1:7} )
status = solve(m)

Benchmarking results

The bencharmking results are presented in The 7th International Conference on Algorithmic Differentation, titled "On efficient Hessian computation using the edge pushing algorithm in Julia".

An improved performance results are submitted in the post-conference publication in Optimization Methods and Software. The results can be reproduced using scripts at benchmark directory of this repository.

#References: R.M. Gower and M.P. Mello. "A new framework for the computation of Hessians", Optimization Methods and Software 27-2, pp. 251–273, 2012. paper