comparison toolboxes/alps/ALPSdemo.m @ 154:0de08f68256b ivand_dev

ALPS toolbox - Algebraic Pursuit added to smallbox
author Ivan Damnjanovic lnx <ivan.damnjanovic@eecs.qmul.ac.uk>
date Fri, 12 Aug 2011 11:17:47 +0100
parents
children
comparison
equal deleted inserted replaced
153:af307f247ac7 154:0de08f68256b
1 % =========================================================================
2 % Algebraic Pursuit algorithms - Demo
3 % =========================================================================
4 % Algebraic Pursuit (ALPS) algorithms are accelerated hard thresholding methods
5 % on sparse recovery of linear inverse systems. In particular, let
6 % Phi : M x N real matrix (M < N)
7 % x* : N x 1 K-sparse data vector
8 % n : N x 1 additive noise vector
9 % then, y = Phi x* + n is the undersampled M x 1 measurement vector. ALPS
10 % solve the following minimization problem
11 % minimize ||y - Phi x||^2 subject to x is K-sparse vector.
12 %
13 % Detailed discussion on the algorithm can be found in
14 % [1] "On Accelerated Hard Thresholding Methods for Sparse Approximation", written
15 % by Volkan Cevher, Technical Report, 2011.
16 % =========================================================================
17 % 01/04/2011, by Anastasios Kyrillidis. anastasios.kyrillidis@epfl.ch, EPFL.
18 % =========================================================================
19
20 clear all
21 close all
22 clc
23
24 addpath ALPS/;
25
26 % General parameters
27 N = 1000; % Size of sparse vector x
28 M = 300; % Number of measurements
29 K = ceil(M*(0.25)); % Sparsity of x*
30 sigma = 0.000; % Additive noise variance
31 verbose = 1; % Print information
32
33 % ALPS parameters
34 its = 600; % Maximum number of iterations
35 Ttol= 1e-6; % Convergence tolerance
36 mode = 1; % Binary representation of (SolveNewtonb, GradientDescentx, SolveNewtox)
37 memory = 1; % Memory of IHT method
38 tau = 0; % Momentum term
39 useCG = 0; % Usage of Conjugate gradients method
40 mu = 0; % Step size selection
41
42 % Please refer to generate_matrix.m and generate_vector.m files
43 m_ensemble = 'Gaussian'; % Measurement matrix ensemble type
44 x_ensemble = 'Gaussian'; % Sparse vector ensemble type
45
46 % reset(RandStream.getDefaultStream);
47
48 %% Generate data
49 x = generate_vector(N, K, x_ensemble);
50 Phi = generate_matrix(M, N, m_ensemble);
51 noise = sigma*randn(M,1);
52 y = Phi*x + noise;
53
54 %% ALPS reconstruction
55 disp('=====================================================================');
56 [x_hat, numiter, time, x_path] = AlgebraicPursuit(y, Phi, K, 'memory', memory, 'mode', mode, 'tolerance', Ttol, 'ALPSiterations', its, 'verbose', verbose, 'tau', tau, 'useCG', useCG, 'mu', mu);
57 if (numiter ~= -1)
58 str = sprintf('Error norm: %f ', norm(x-x_hat));
59 disp(str);
60 end;
61
62 %% Plot error per iteration
63 if (numiter ~= -1)
64 error = sqrt(sum(abs(x*ones(1,size(x_path,2))-x_path).^2));
65
66 set(0, 'DefaultAxesFontSize', 14);
67 figure;
68 semilogy(error,'LineWidth',3); hold on;
69 grid on;
70 semilogy((norm(noise))*ones(1,its),'m-','LineWidth',1);
71
72 xlabel('[i]:iteration number')
73 ylabel('||x-x_i||')
74 axis([1 numiter + 5 10^-5 2*10^0])
75
76 legend(strcat('Time = [',num2str(time),']'), 'Noise Level');
77 title(strcat('N = ', num2str(N), ', M = ', num2str(M), ', K = ', num2str(K)));
78 shg;
79 end;