view toolboxes/AudioInpaintingToolbox/Experiments/MissingSampleTopologyExperiment/MissingSampleTopologyExperiment.m @ 153:af307f247ac7 ivand_dev

Example scripts for Two Step Dictionary Learning - Image Denoising experiments.

function MissingSampleTopologyExperiment(expParam)
% For a total number of missing samples C in a frame, create several
% configuration of B holes with length A, where A*B=C (i.e. the total 
% number of missing samples is constant). Test several values of C, several
% solvers. For each C, test all possible combination of (A,B) such that
% A*B=C.
% Note that for each combination (A,B), a number of frames are tested at
% random and SNR results are then averaged.
%
% Usage: MissingSampleTopologyExperiment(expParam)
%
%
% Inputs:
%          - expParam is an optional structure where the user can define
%          the experiment parameters.
%          - expParam.soundDir: path to sound directory. All the .wav files
%          in this directory will be tested at random.
%          - expParam.destDir: path to store the results.
%          - expParam.N: frame length
%          - expParam.NFramesPerHoleSize: number of frames to use for each
%          testing configuration (A,B). Results are then averaged.
%          - expParam.totalMissSamplesList: list of all tested values C for
%          the total number of missing samples in a frame
%          - expParam.solvers: list of solvers with their parameters
%
% -------------------
%
% Audio Inpainting toolbox
% Date: June 28, 2011
% By Valentin Emiya, Amir Adler, Maria Jafari
% This code is distributed under the terms of the GNU Public License version 3 (http://www.gnu.org/licenses/gpl.txt).

if ~isdeployed
    addpath('../../Problems/');
    addpath('../../Solvers/');
    addpath('../../Utils/');
    addpath('../../Utils/dictionaries/');
    addpath('../../Utils/evaluation/');
%     addpath('../../Utils/TCPIP_SocketCom/');
%     javaaddpath('../../Utils/TCPIP_SocketCom');
    dbstop if error
    close all
end

%% Set parameters
if nargin<1
    expParam = [];
end
% Path to audio files
if ~isfield(expParam,'soundDir'),
    expParam.soundDir = '../../Data/testSpeech8kHz_from16kHz/';
end
if ~isfield(expParam,'destDir'),
    expParam.destDir = '../../tmp/missSampTopoExp/';
end
if ~exist(expParam.destDir,'dir')
    mkdir(expParam.destDir)
end


% frame length
if ~isfield(expParam,'N'),
    expParam.N = 512;
    expParam.N = 256;
    warning('AITB:N','Frame length=256 is used to have faster computations. Recommended frame length is 512 at 8kHz.');
end

% Number of random frames to test
if ~isfield(expParam,'NFramesPerHoleSize'),
    expParam.NFramesPerHoleSize = 20;
    warning('AITB:NFrames','expParam.NFramesPerHoleSize = 20 is used to have faster computations. Recommended value: several hundreds.');
end

% Number of missing samples: which numbers to test?
if ~isfield(expParam,'totalMissSamplesList')
    expParam.totalMissSamplesList = [12,36,60,120,180,240];
    expParam.totalMissSamplesList = [12,36];
    warning('AITB:Miss','expParam.totalMissSamplesList = [12,36] is used to have faster computations. Recommended list: expParam.totalMissSamplesList = [12,36,60,120,180,240].');
end

% Choose the solver methods you would like to test: OMP, L1, Janssen
if ~isfield(expParam,'solvers'),
    nSolver = 0;
    nSolver = nSolver+1;
    expParam.solvers(nSolver).name = 'OMP-G';
    expParam.solvers(nSolver).inpaintFrame = @inpaintFrame_OMP_Gabor; % solver function
    expParam.solvers(nSolver).param.N = expParam.N; % frame length
    expParam.solvers(nSolver).param.OMPerr = 0.001;
    expParam.solvers(nSolver).param.sparsityDegree = expParam.solvers(nSolver).param.N/4;
    expParam.solvers(nSolver).param.D_fun = @Gabor_Dictionary; % Dictionary (function handle)
    expParam.solvers(nSolver).param.redundancyFactor = 2; % Dictionary redundancy
    expParam.solvers(nSolver).param.wa = @wRect; % Analysis window
    
    nSolver = nSolver+1;
    expParam.solvers(nSolver).name = 'Janssen';
    expParam.solvers(nSolver).inpaintFrame = @inpaintFrame_janssenInterpolation; % solver function
    expParam.solvers(nSolver).param.N = expParam.N; % frame length
end




%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
soundDir = expParam.soundDir;
wavFiles = dir([soundDir '*.wav']);
wavFiles = arrayfun(@(x)[soundDir x.name],wavFiles,'UniformOutput',false);

%% Draw a list of random frames
% Choose an audio file at random
frameParam.kFrameFile = randi(length(wavFiles),expParam.NFramesPerHoleSize);

% For each audio file, find maximum mean energy among all frames
[dum fs] = wavread([soundDir wavFiles{1}],'size');
Ne = round(512/16000*fs);
E2m = zeros(length(wavFiles),1);
for kf = 1:length(wavFiles)
    x=wavread(wavFiles{kf});
    xm = filter(ones(Ne,1)/Ne,1,abs(x.^2));
    E2m(kf) = 10*log10(max(xm));
end

% Choose the location of a frame at random, with a minimum energy
maxDiffE2m = 10;
frameParam.kFrameBegin = NaN(expParam.NFramesPerHoleSize,1);
for kf = 1:expParam.NFramesPerHoleSize
    siz = wavread(wavFiles{frameParam.kFrameFile(kf)},'size');
    while true
        frameParam.kFrameBegin(kf) = randi(siz(1)-expParam.N+1);
        x = wavread(wavFiles{frameParam.kFrameFile(kf)},[0,expParam.N-1]+frameParam.kFrameBegin(kf));
        E2m0 = 10*log10(mean(abs(x.^2)));
        if E2m(frameParam.kFrameFile(kf))-E2m0 <= maxDiffE2m
            break
        end
    end
end

%% Test each number of missing samples
PerfRes = cell(length(expParam.totalMissSamplesList),length(expParam.solvers));
factorsToTest = cell(length(expParam.totalMissSamplesList),length(expParam.solvers));
outputFile = [expParam.destDir 'missSampTopoExp.mat'];
for kSolver = 1:length(expParam.solvers)
    fprintf('\n ------ Solver: %s ------\n\n',...
        expParam.solvers(kSolver).name);
    for kMiss = 1:length(expParam.totalMissSamplesList)
        NMissSamples = expParam.totalMissSamplesList(kMiss);
        factorsToTest{kMiss} = allFactors(NMissSamples);
        PerfRes{kMiss,kSolver} = zeros([length(factorsToTest{kMiss}),expParam.NFramesPerHoleSize]);
        for kFactor = 1:length(factorsToTest{kMiss})
            holeSize = factorsToTest{kMiss}(kFactor);
            NHoles = NMissSamples/holeSize;
            fprintf('%d %d-length holes (%d missing samples = %.1f%%)\n',...
                NHoles,holeSize,NMissSamples,NMissSamples/expParam.N*100)
            problemParameters.holeSize = holeSize;
            problemParameters.NHoles = NHoles;
            for kFrame = 1:expParam.NFramesPerHoleSize
                %% load audio frame
                xFrame = wavread(...
                    wavFiles{frameParam.kFrameFile(kFrame)},...
                    frameParam.kFrameBegin(kFrame)+[0,expParam.N-1]);
                
                %% generate problem
                [problemData, solutionData] = ...
                    generateMissingGroupsProblem(xFrame,problemParameters);
                
                %% solve problem
                xEst = ...
                    expParam.solvers(kSolver).inpaintFrame(...
                    problemData,...
                    expParam.solvers(kSolver).param);
                
                %% compute and store performance
                [SNRAll, SNRmiss] = ...
                    SNRInpaintingPerformance(...
                    solutionData.xClean,...
                    problemData.x,...
                    xEst,...
                    problemData.IMiss);
                PerfRes{kMiss,kSolver}(kFactor,kFrame) = SNRmiss(2);
                
            end
        end
        save(outputFile,'PerfRes','expParam');
    end
end

figure
Nrows = floor(sqrt(length(expParam.solvers)));
Ncols = ceil(sqrt(length(expParam.solvers))/Nrows);
cmap = lines;
for kSolver = 1:length(expParam.solvers)
    subplot(Nrows,Ncols,kSolver)
    hold on,grid on
    for kMiss = 1:length(expParam.totalMissSamplesList)
        plot(factorsToTest{kMiss},mean(PerfRes{kMiss,kSolver},2),...
            'color',cmap(kMiss,:));
    end
    title(expParam.solvers(kSolver).name)
end
return

function m = allFactors(n)
% Find the list of all factors (not only prime factors)

primeFactors = factor(n);

degrees = zeros(size(primeFactors));

for k=1:length(degrees)
    degrees(k) = sum(primeFactors==primeFactors(k));
end

[primeFactors, I] = unique(primeFactors);
degrees = degrees(I);

D = (0:degrees(1)).';
for k=2:length(degrees)
    Dk = ones(size(D,1),1)*(0:degrees(k));
    D = [repmat(D,degrees(k)+1,1),Dk(:)];
end

m = unique(sort(prod((ones(size(D,1),1)*primeFactors).^D,2)));

return