Awesome

SparseGDLibrary : Sparse Gradient Descent Library in MATLAB

Authors: Hiroyuki Kasai

Last page update: Oct. 15, 2018

Latest library version: 1.0.3 (see Release notes for more info)

Introduction

The SparseGDLibrary is a pure-Matlab library of a collection of unconstrained optimization algorithms for sparse modeling.

List of sparse gradient algorithms available in SparseGDLibrary

• APG (Accelerated gradient descent, i.e., Nesterov AGD)
• ISTA (Iterative shrinkage-thresholding algorithm)
• FISTA (Fast iterative shrinkage-thresholding algorithm)
• CD (Coodinate descent) for Lasso and Elastic Net
• ADMM (The alternating direction method of multipliers) for Lasso

List of line-search algorithms available in SparseGDLibrary

• Backtracking line search (a.k.a Armijo condition)
• Strong wolfe line search
• Exact line search
  • Only for quadratic problem.
• TFOCS-style line search

Supported problems

• Lasso (Least absolute shrinkage and selection operator) problem
• Elastic Net problem
• Group lasso problem
• Matrix completion problem with trace norm minimization
• L1-norm logistic regression
• L1-norm linear support vector machine (SVM)
• Robust principal component analysis (PCA) problem
  • Video background subtraction application
  • Image inpainting application
• Trace norm convex clustering problem

Folders and files

First to do

Run run_me_first for path configurations.

%% First run the setup script
run_me_first; 

Usage example 1 (Lasso problem)

Now, just execute demo for demonstration of this package.

%% Execute the demonstration script
demo; 

The "demo.m" file contains below.

% set number of dimensions, cardinality and noise level.
n = 1280; 
d = 100;  
k = 15; 
noise_level = 0.01;

% generate dataset
[A, b, x0, lambda] = generate_lasso_data(n, d, k, noise_level);

% define problem
problem = lasso_problem(A, b, lambda);

% perform algorithms (FISTA and ADMM)
options.w_init = zeros(n,1); 
[w_fista, info_fista] = fista(problem, options);
[w_admm, info_admm] = admm_lasso(problem, options); 

% plot all
% display iter vs. cost
display_graph('iter','cost', {'FISTA', 'ADMM'}, {w_fista, w_admm}, {info_fista, info_admm});
% display iter vs. l1-norm
display_graph('iter','sol_optimality_gap', {'FISTA', 'ADMM'}, {w_fista, w_admm}, {info_fista, info_admm});  

• Output results

Usage example 1: more plots

The decrease of the l1-norm of the solution according to iterations is illustrated.

display_graph('iter','l1-norm', {'FISTA', 'ADMM'}, {w_fista, w_admm}, {info_fista, info_admm}, 'linear');

The final coefficients in each position of the solution are displayed in comparisn with the ogirinal input sparse signal.

display_graph('coeff_pos','coeff_amp', {'FISTA','ADMM','Original (x0)'}, {w_fista, w_admm, x0}, {w_fista, w_admm, x0}, 'linear', 'line-with-mark', 1);           

• Output results

Usage example 2 (Lasso problem with cross-validation)

Execute demo_lasso_cv for the lasso problem with cross-validation.

% Execute the demonstration script
demo_lass_cv; 

The "demo_lass_cv.m" file contains below.

function demo_lasso_cv()

% set number of dimensions, cardinality and noise level.
n = 1280; 
d = 100;  
k = 15; 
noise_level = 0.01;

% generate dataset
[A, b, ~, ~, lambda_max] = generate_lasso_data(n, d, k, noise_level);

% set algorithms and solver (e.g., FISTA)
algorithm = {'FISTA'};

% initialize
% define parameters for cross-validation
num_cv = 10;
lambda_unit = lambda_max/num_cv;
lambda_array = 0+lambda_unit:lambda_unit:lambda_max;

% prepare arrays for solutions
W = zeros(n, num_cv);
l1_norm = zeros(num_cv,1);    
aprox_err = zeros(num_cv,1);  

% set options
options.w_init = zeros(n,1); 

% perform cross-validations
for i=1:length(lambda_array)
    lambda = lambda_array(i);
    problem = lasso_problem(A, b, lambda);

    [W(:,i), infos] = fista(problem, options);
    l1_norm(i) = infos.reg(end);
    aprox_err(i) = infos.cost(end);
end

% plot all
% display l1-norm vs. coefficient
display_graph('l1-norm','coeffs', algorithm, l1_norm, {W}, 'linear');
% display lambda vs. coefficient
display_graph('lambda','coeffs', algorithm, lambda_array, {W}, 'linear');
% display l1-norm vs. approximation error
display_graph('l1-norm','aprox_err', algorithm, l1_norm, {aprox_err}, 'linear');

end  

• Output results

License

The SparseGDLibrary is free and open source for academic/research purposes (non-commercial).

Problems or questions

If you have any problems or questions, please contact the author: Hiroyuki Kasai (email: kasai at is dot uec dot ac dot jp)

Release Notes

• Version 1.0.3 (Oct. 13, 2018)
  • ISTA solver is added.
• Version 1.0.2 (July 07, 2017)
  • Robust PCA problem and trace norm convex clustering problem, and those related solvers and test files are added.
• Version 1.0.1 (April 27, 2017)
  • Group lasso problem is added.
• Version 1.0.0 (April 24, 2017)
  • Initial version.