Skip to content
/ apg Public

A lightweight accelerated proximal-gradient package for matlab

License

Notifications You must be signed in to change notification settings

bodono/apg

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

22 Commits
 
 
 
 
 
 
 
 
 
 

Repository files navigation

apg

(@author bodonoghue) MATLAB script
Implements an Accelerated Proximal Gradient method (Nesterov 2007, Beck and Teboulle 2009)

solves:

minimize f(x) + h(x)
over x \in R^dim_x

where f is smooth, convex - user supplies function to evaluate gradient of f
h is convex - user supplies function to evaluate the proximal operator of h

call as:

x = apg( grad_f, prox_h, dim_x, opts )

this takes in two function handles:

grad_f(v, opts) = df(v)/dv 
(gradient of f at v)

prox_h(v, t, opts) = argmin_x ( t * h(x) + 1/2 * norm(x-v)^2 )
where t is the (scalar, positive) step size at that iteration

if h = 0, then use prox_h = [] or prox_h = @(x,t,opts)(x)

put all necessary function data in opts struct

each iteration of apg requires one gradient evaluation of f and one prox step with h

quits when:

norm( y(k) - x(k+1) ) < EPS * max( 1,norm( x(k+1) )

apg implements something similar to TFOCS step-size adaptation (Becker, Candes and Grant 2010)
and gradient-scheme adaptive restarting (O'Donoghue and Candes 2013)

see examples/ for usage instances

optional opts fields defined are below (with defaults)
to use defaults simply call apg with opts = []

X_INIT = zeros(dim_x,1); % initial starting point

USE_RESTART = true; % use adaptive restart scheme

MAX_ITERS = 2000; % maximum iterations before termination

EPS = 1e-6; % tolerance for termination

ALPHA = 1.01; % step-size growth factor

BETA = 0.5; % step-size shrinkage factor

QUIET = false; % if false writes out information every 100 iters

GEN_PLOTS = true; % if true generates plots of proximal gradient

USE_GRA = false; % if true uses unaccelerated proximal gradient descent

Example of usage:

function x = apg_lasso(A, b, rho, opts)
    % uses apg to solve a lasso problem
    %
    % min_x (1/2)*sum_square(A*x - b) + rho * norm(x,1) 
    %
    % rho is the L1-regularization weight

    opts.A = A;
    opts.b = b;
    opts.rho = rho;

    x = apg(@quad_grad, @soft_thresh, size(A,2), opts);

end

function g = quad_grad(x, opts)
    g = opts.A'*(opts.A*x - opts.b);
end

function v = soft_thresh(x, t, opts)
    v = sign(x) .* max(abs(x) - t*opts.rho,0);
end

About

A lightweight accelerated proximal-gradient package for matlab

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages