LSGAN——最小二乘GAN(matlab源码)

1、LSGAN的原理简述

最小二乘GAN，正如它的名字所指示的，目标函数将是一个平方误差，考虑到D网络的目标是分辨两类，如果给生成样本和真实样本分别编码为a,b，那么采用平方误差作为目标函数，D的目标就是

 G的目标函数将编码a换成编码c，这个编码表示D将G生成的样本当成真实样本

 2、matlab实战（手写数据集为例）

clear all; close all; clc;
%% Least Squares Generative Adversarial Network
%% Load Data
load('mnistAll.mat')
trainX = preprocess(mnist.train_images); 
trainY = mnist.train_labels;
testX = preprocess(mnist.test_images); 
testY = mnist.test_labels;
%% Settings
settings.latent_dim = 100;
settings.batch_size = 32; settings.image_size = [28,28,1]; 
settings.lrD = 0.0002; settings.lrG = 0.0002; settings.beta1 = 0.5;
settings.beta2 = 0.999; settings.maxepochs = 50;
paramsGen.FCW1 = dlarray(...
    initializeGaussian([256,settings.latent_dim],.02));
paramsGen.FCb1 = dlarray(zeros(256,1,'single'));
paramsGen.BNo1 = dlarray(zeros(256,1,'single'));
paramsGen.BNs1 = dlarray(ones(256,1,'single'));
paramsGen.FCW2 = dlarray(initializeGaussian([512,256]));
paramsGen.FCb2 = dlarray(zeros(512,1,'single'));
paramsGen.BNo2 = dlarray(zeros(512,1,'single'));
paramsGen.BNs2 = dlarray(ones(512,1,'single'));
paramsGen.FCW3 = dlarray(initializeGaussian([1024,512]));
paramsGen.FCb3 = dlarray(zeros(1024,1,'single'));
paramsGen.BNo3 = dlarray(zeros(1024,1,'single'));
paramsGen.BNs3 = dlarray(ones(1024,1,'single'));
paramsGen.FCW4 = dlarray(initializeGaussian(...
    [prod(settings.image_size),1024]));
paramsGen.FCb4 = dlarray(zeros(prod(settings.image_size)...
    ,1,'single'));

stGen.BN1 = []; stGen.BN2 = []; stGen.BN3 = [];
%% preprocess
function x = preprocess(x)
x = double(x)/255;
x = (x-.5)/.5;
x = reshape(x,28*28,[]);
end
%% extract data
function x = gatext(x)
x = gather(extractdata(x));
end
%% gpu dl array wrapper
function dlx = gpdl(x,labels)
dlx = gpuArray(dlarray(x,labels));
end
%% Weight initialization
function parameter = initializeGaussian(parameterSize,sigma)
if nargin < 2
    sigma = 0.05;
end
parameter = randn(parameterSize, 'single') .* sigma;
end
%% Generator
function [dly,st] = Generator(dlx,params,st)
% fully connected
%1
dly = fullyconnect(dlx,params.FCW1,params.FCb1);
dly = leakyrelu(dly,0.2);
if isempty(st.BN1)
    [dly,st.BN1.mu,st.BN1.sig] = batchnorm(dly,params.BNo1,params.BNs1);
else
    [dly,st.BN1.mu,st.BN1.sig] = batchnorm(dly,params.BNo1,...
        params.BNs1,st.BN1.mu,st.BN1.sig);
end
%2
dly = fullyconnect(dly,params.FCW2,params.FCb2);
dly = leakyrelu(dly,0.2);
if isempty(st.BN2)
    [dly,st.BN2.mu,st.BN2.sig] = batchnorm(dly,params.BNo2,params.BNs2);
else
    [dly,st.BN2.mu,st.BN2.sig] = batchnorm(dly,params.BNo2,...
        params.BNs2,st.BN2.mu,st.BN2.sig);
end
%3
dly = fullyconnect(dly,params.FCW3,params.FCb3);
dly = leakyrelu(dly,0.2);
if isempty(st.BN3)
    [dly,st.BN3.mu,st.BN3.sig] = batchnorm(dly,params.BNo3,params.BNs3);
else
    [dly,st.BN3.mu,st.BN3.sig] = batchnorm(dly,params.BNo3,...
        params.BNs3,st.BN3.mu,st.BN3.sig);
end
%4
dly = fullyconnect(dly,params.FCW4,params.FCb4);
% tanh
dly = tanh(dly);
end
%% Discriminator
function [dly,st] = Discriminator(dlx,params,st)
% fully connected 
%1
dly = fullyconnect(dlx,params.FCW1,params.FCb1);
dly = leakyrelu(dly,0.2);
dly = dropout(dly);
if isempty(st.BN1)
    [dly,st.BN1.mu,st.BN1.sig] = batchnorm(dly,params.BNo1,params.BNs1);
else
    [dly,st.BN1.mu,st.BN1.sig] = batchnorm(dly,params.BNo1,...
        params.BNs1,st.BN1.mu,st.BN1.sig);
end
%2
dly = fullyconnect(dly,params.FCW2,params.FCb2);
dly = leakyrelu(dly,0.2);
dly = dropout(dly);
if isempty(st.BN2)
    [dly,st.BN2.mu,st.BN2.sig] = batchnorm(dly,params.BNo2,params.BNs2);
else
    [dly,st.BN2.mu,st.BN2.sig] = batchnorm(dly,params.BNo2,...
        params.BNs2,st.BN2.mu,st.BN2.sig);
end
%3
dly = fullyconnect(dly,params.FCW3,params.FCb3);
dly = leakyrelu(dly,0.2);
dly = dropout(dly);
%4
dly = fullyconnect(dly,params.FCW4,params.FCb4);
% sigmoid
% dly = sigmoid(dly);
end
%% modelGradients
function [GradGen,GradDis,stGen,stDis]=modelGradients(x,z,paramsGen,...
    paramsDis,stGen,stDis)
[fake_images,stGen] = Generator(z,paramsGen,stGen);
d_output_real = Discriminator(x,paramsDis,stDis);
[d_output_fake,stDis] = Discriminator(fake_images,paramsDis,stDis);

% Loss due to true or not
d_loss = .5*mean((d_output_real-1).^2+(d_output_fake).^2);
g_loss = .5*mean((d_output_fake-1).^2);

% For each network, calculate the gradients with respect to the loss.
GradGen = dlgradient(g_loss,paramsGen,'RetainData',true);
GradDis = dlgradient(d_loss,paramsDis);
end
%% progressplot
function progressplot(paramsGen,stGen,settings)
r = 5; c = 5;
noise = gpdl(randn([settings.latent_dim,r*c]),'CB');
gen_imgs = Generator(noise,paramsGen,stGen);
gen_imgs = reshape(gen_imgs,28,28,[]);

fig = gcf;
if ~isempty(fig.Children)
    delete(fig.Children)
end

I = imtile(gatext(gen_imgs));
I = rescale(I);
imagesc(I)
title("Generated Images")
colormap gray

drawnow;
end
%% dropout
function dly = dropout(dlx,p)
if nargin < 2
    p = .3;
end
n = p*10;
mask = randi([1,10],size(dlx));
mask(mask<=n)=0;
mask(mask>n)=1;
dly = dlx.*mask;

end