test_srng.m

最详尽的神经网络源码

%% First example 

clear, clc, close all hidden
disp('Running example 1 ....')

d                                     = 2;
Ntrain                                = 100;
m                                     = 2;
M0                                    = [0 ; 0];
R0                                    = [1 0 ; 0 1];
M1                                    = [2 ; 3];
R1                                    = [0.5 0.1 ; 0.2 1];
vect_test                             = (-4:0.1:8);
options.epsilonk                      = 0.005;
options.epsilonl                      = 0.001;
options.epsilonlambda                 = 10e-8;
options.sigmastart                    = 2;
options.sigmaend                      = 10e-4;
options.sigmastretch                  = 10e-3;
options.threshold                     = 10e-10;
options.xi                            = 0.1;
options.nb_iterations                 = 5000;
options.metric_method                 = 1;
options.shuffle                       = 1;
options.updatelambda                  = 1;



Xtrain                                = [M0(: , ones(1 , Ntrain/2)) + chol(R0)'*randn(d , Ntrain/2) , M1(: , ones(1 , Ntrain/2)) + chol(R1)'*randn(d , Ntrain/2)];
ytrain                                = [zeros(1 , Ntrain/2) , ones(1 , Ntrain/2)];
[X , Y]                               = meshgrid(vect_test);
Xtest                                 = [X(:)' ; Y(:)'];

Nproto_pclass                         = 4*ones(1 , length(unique(ytrain)));


[Wproto , yproto , lambda]            = ini_proto(Xtrain , ytrain , Nproto_pclass);
[Wproto_est , yproto_est , lambda_est,  E_SRNG]    = srng_model(Xtrain , ytrain , Wproto , yproto , lambda , options);
ytest_est                             = NN_predict(Xtest , Wproto_est , yproto_est,lambda_est,options);

indtrain0                             = (ytrain == 0);
indtrain1                             = (ytrain == 1);

indproto0                             = (yproto_est == 0);
indproto1                             = (yproto_est == 1);

figure(1)
imagesc(vect_test , vect_test , reshape(ytest_est , length(vect_test) , length(vect_test)) )
axis ij
hold on
plot(Xtrain(1 , indtrain0) , Xtrain(2 , indtrain0) , 'k+' , Xtrain(1 , indtrain1) , Xtrain(2 , indtrain1) , 'm+' , Wproto_est(1 , indproto0) ,  Wproto_est(2 , indproto0) , 'ko' , Wproto_est(1 , indproto1) ,  Wproto_est(2 , indproto1) , 'mo')
h = voronoi(Wproto_est(1 , :) , Wproto_est(2 , :));
set(h , 'color' , 'y' , 'linewidth' , 2)
hold off
title('E_{SRNG}(t)' , 'fontsize' , 12)
colorbar


figure(2)
plot(E_SRNG);
title('E_{SRNG}(t)' , 'fontsize' , 12)


figure(3)
stem(lambda_est);
title('\lambda' , 'fontsize' , 12)



pause

%% Second example : Roc curve on ionosphere data 

disp('Running example 2 ....')

clear, close all hidden

load ionosphere
Nproto_pclass                      = 4*ones(1 , length(unique(y)));


options.epsilonk                   = 0.005;
options.epsilonl                   = 0.001;
options.epsilonlambda              = 10e-8;
options.sigmastart                 = 2;
options.sigmaend                   = 10e-4;
options.sigmastretch               = 10e-3;
options.threshold                  = 10e-10;
options.xi                         = 2;
options.nb_iterations              = 5000;
options.metric_method              = 1;
options.shuffle                    = 1;
options.updatelambda               = 1;



options.method                     = 7;
options.holding.rho                = 0.7;
options.holding.K                  = 20;


X                                  = normalize(X);
[Itrain , Itest]                   = sampling(X , y , options);

Perftrain                          = zeros(1 , size(Itrain , 1));
Perftest                           = zeros(1 , size(Itrain , 1));
tptrain                            = zeros(size(Itrain , 1) , 100);
fptrain                            = zeros(size(Itrain , 1) , 100);
tptest                             = zeros(size(Itrain , 1) , 100);
fptest                             = zeros(size(Itrain , 1) , 100);

for i = 1 : size(Itrain , 1)

    [Xtrain , ytrain , Xtest , ytest]  = samplingset(X , y , Itrain , Itest , i);

    [Wproto , yproto , lambda]         = ini_proto(Xtrain , ytrain , Nproto_pclass);
    [Wproto_est , yproto_est , lambda_est,  E_SRNG]    = srng_model(Xtrain , ytrain , Wproto , yproto , lambda, options);

    [ytest_est , disttest]             = NN_predict(Xtest , Wproto_est , yproto_est , lambda_est , options);
    [ytrain_est , disttrain]           = NN_predict(Xtrain , Wproto_est , yproto_est , lambda_est , options);

    Perftrain(i)                       = perf_classif(ytrain , ytrain_est);
    Perftest(i)                        = perf_classif(ytest , ytest_est);

    dktrain                            = min(disttrain(yproto==0 , :));
    dltrain                            = min(disttrain(yproto~=0 , :));
    nutrain                            = (dktrain - dltrain)./(dktrain + dltrain);
    [tptrain(i , :) , fptrain(i , :)]  = basicroc(ytrain , nutrain);


    dktest                             = min(disttest(yproto==0 , :));
    dltest                             = min(disttest(yproto~=0 , :));
    nutest                             = (dktest - dltest)./(dktest + dltest);
    [tptest(i , :) , fptest(i , :)]    = basicroc(ytest , nutest);

    disp(sprintf('%d/%d' , i , options.holding.K))
    drawnow

end

disp('Performances Train/Test')
disp([mean(Perftrain,2) , mean(Perftest,2)])

figure(1)
plot(E_SRNG);
title('E_{SRNG}(t)' , 'fontsize' , 12)

figure(2)
stem(lambda_est);
title('\lambda' , 'fontsize' , 12)

figure(3)
plot(mean(fptrain , 1) , mean(tptrain , 1) , mean(fptest , 1) , mean(tptest , 1) , 'r' , 'linewidth'  , 2)
axis([-0.02 , 1.02 , -0.02 , 1.02])
xlabel('false positive rate');
ylabel('true positive rate');
title('ROC curve','fontsize' , 12);
legend(['Train'] , ['Test'])