深度卷积对抗生成网络（DCGAN）matlab实战

 一、原理

深度卷积对抗生成网络 （DCGAN）将GAN与CNN相结合，奠定后几乎所有GAN的基本网络架构。DCGAN极大地提升了原始GAN训练的稳定性以及生成结果质量。

DCGAN网络设计中采用了当时对CNN比较流行的改进方案：

1、将空间池化层用卷积层替代，这种替代只需要将卷积的步长stride设置为大于1的数值。改进的意义是下采样过程不再是固定的抛弃某些位置的像素值，而是可以让网络自己去学习下采样方式。

2、将全连接层去除

3、采用BN层，BN的全称是Batch Normalization，是一种用于常用于卷积层后面的归一化方法，起到帮助网络的收敛等作用。作者实验中发现对所有的层都使用BN会造成采样的震荡（我也不理解什么是采样的震荡，我猜是生成图像趋于同样的模式或者生成图像质量忽高忽低）和网络不稳定。

4、在生成器中除输出层使用Tanh（Sigmoid）激活函数，其余层全部使用ReLu激活函数。

5、在判别器所有层都使用LeakyReLU激活函数，防止梯度稀。

下面是DCGAN的生成器网络架构图。

 二、代码实战

clear all; close all; clc;
%% Deep Convolutional 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.latentDim = 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;
%% Generator
paramsGen.FCW1 = dlarray(initializeGaussian([128*7*7,...
    settings.latentDim]));
paramsGen.FCb1 = dlarray(zeros(128*7*7,1,'single'));
paramsGen.TCW1 = dlarray(initializeGaussian([3,3,128,128]));
paramsGen.TCb1 = dlarray(zeros(128,1,'single'));
paramsGen.BNo1 = dlarray(zeros(128,1,'single'));
paramsGen.BNs1 = dlarray(ones(128,1,'single'));
paramsGen.TCW2 = dlarray(initializeGaussian([3,3,64,128]));
paramsGen.TCb2 = dlarray(zeros(64,1,'single'));
paramsGen.BNo2 = dlarray(zeros(64,1,'single'));
paramsGen.BNs2 = dlarray(ones(64,1,'single'));
paramsGen.CNW1 = dlarray(initializeGaussian([3,3,64,1]));
paramsGen.CNb1 = dlarray(zeros(1,1,'single'));
stGen.BN1 = []; stGen.BN2 = [];

%% Discriminator
paramsDis.CNW1 = dlarray(initializeGaussian([3,3,1,32]));
paramsDis.CNb1 = dlarray(zeros(32,1,'single'));
paramsDis.CNW2 = dlarray(initializeGaussian([3,3,32,64]));
paramsDis.CNb2 = dlarray(zeros(64,1,'single'));
paramsDis.BNo1 = dlarray(zeros(64,1,'single'));
paramsDis.BNs1 = dlarray(ones(64,1,'single'));
paramsDis.CNW3 = dlarray(initializeGaussian([3,3,64,128]));
paramsDis.CNb3 = dlarray(zeros(128,1,'single'));
paramsDis.BNo2 = dlarray(zeros(128,1,'single'));
paramsDis.BNs2 = dlarray(ones(128,1,'single'));
paramsDis.CNW4 = dlarray(initializeGaussian([3,3,128,256]));
paramsDis.CNb4 = dlarray(zeros(256,1,'single'));
paramsDis.BNo3 = dlarray(zeros(256,1,'single'));
paramsDis.BNs3 = dlarray(ones(256,1,'single'));
paramsDis.FCW1 = dlarray(initializeGaussian([1,256*4*4]));
paramsDis.FCb1 = dlarray(zeros(1,1,'single'));
stDis.BN1 = []; stDis.BN2 = []; stDis.BN3 = [];

% average Gradient and average Gradient squared holders
avgG.Dis = []; avgGS.Dis = []; avgG.Gen = []; avgGS.Gen = [];
%% Train
numIterations = floor(size(trainX,4)/settings.batch_size);
out = false; epoch = 0; global_iter = 0;


%% 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 = -mean(.9*log(d_output_real+eps)+log(1-d_output_fake+eps));
g_loss = -mean(log(d_output_fake+eps));

% 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.latentDim,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;

实验结果

epoch = 5;

epoch = 6