基于DCGAN生成手写数字--pytorch

DCGAN对GAN的改善在于使用深度卷积网络代替全连接网络

全部代码：

import torch
from torch import nn
from torchvision.datasets import MNIST
from torch.utils.data import DataLoader,Dataset
import torchvision
import os
import numpy as np
from PIL import Image
import matplotlib.pyplot as plt

class Discirmintor(nn.Module):
    def __init__(self):
        super(Discirmintor, self).__init__()
        # 28,28,1 ---> 14,14,32
        self.conv1=nn.Conv2d(in_channels=1,out_channels=32,kernel_size=3,stride=2,padding=1)
        self.bn1=nn.BatchNorm2d(num_features=32,momentum=0.8)
        # 14,14,32 ---> 7,7,64
        self.conv2=nn.Conv2d(in_channels=32,out_channels=64,kernel_size=3,stride=2,padding=1)
        self.bn2=nn.BatchNorm2d(num_features=64,momentum=0.8)
        # 7,7,64 ---> 3,3,128
        self.conv3=nn.Conv2d(in_channels=64,out_channels=128,kernel_size=3,stride=2,padding=1)
        self.bn3=nn.BatchNorm2d(num_features=128,momentum=0.8)
        # 3,3,128 ---> 1,1,128
        self.avg=nn.AvgPool2d(kernel_size=3)
        self.flatten=nn.Flatten()

        self.fc=nn.Linear(128,1)
        self.lr=nn.LeakyReLU(0.2)
        self.sigmoid=nn.Sigmoid()

    def forward(self,x):
        x=x.view(-1,1,28,28)
        x=self.lr(self.bn1(self.conv1(x)))
        x=self.lr(self.bn2(self.conv2(x)))
        x=self.lr(self.bn3(self.conv3(x)))

        x=self.avg(x)
        x=x.view(-1,128)
        x=self.fc(x)
        x=self.sigmoid(x)

        return x

class Generator(nn.Module):
    def __init__(self):
        super(Generator, self).__init__()
        self.fc=nn.Linear(noise_size,7*7*256)
        # 7,7,256 ---> 14,14,128
        self.up1=nn.UpsamplingNearest2d(scale_factor=2)
        self.conv1=nn.Conv2d(in_channels=256,out_channels=128,kernel_size=3,padding=1)
        self.bn1=nn.BatchNorm2d(num_features=128,momentum=0.8)
        # 14,14,128 ---> 28,28,64
        self.up2=nn.UpsamplingNearest2d(scale_factor=2)
        self.conv2=nn.Conv2d(in_channels=128,out_channels=64,kernel_size=3,padding=1)
        self.bn2=nn.BatchNorm2d(num_features=64,momentum=0.8)
        # 28,28,64 --->28,28,3
        self.conv3=nn.Conv2d(in_channels=64,out_channels=1,kernel_size=3,padding=1)

        self.relu=nn.ReLU()
        self.tanh=nn.Tanh()

    def forward(self,x):
        x=self.fc(x)
        x=self.relu(x)
        x=x.view(-1,256,7,7)
        x=self.relu(self.bn1(self.conv1(self.up1(x))))
        x=self.relu(self.bn2(self.conv2(self.up2(x))))

        x=self.conv3(x)
        x=self.tanh(x)
        return x

def to_img(image):
    image=0.5*(image+1)
    image=torch.clamp(image,0,1)
    image=image.view(-1,28,28,1)
    return image

def save_img(fake_image,epoch):
    r, c = 5, 5
    fig, axs = plt.subplots(r, c)
    cnt = 0
    for i in range(r):
        for j in range(c):
            axs[i, j].imshow(fake_image[cnt, :, :, 0],cmap='gray')
            axs[i, j].axis('off')
            cnt += 1
    fig.savefig("images/DCGAN-Mnist/epoch_{}.png".format(epoch + 1))
    plt.close()

def train(epochs):
    for epoch in range(epochs):
        for idx,(img,_) in enumerate(dataloader):
            img=img.to(device)
            num_img=img.size(0)

            real_img=img.view(num_img,-1)
            real_label=torch.ones(num_img,1)
            real_label=real_label.to(device)

            fake_img=torch.randn(num_img,noise_size)
            fake_img=fake_img.to(device)
            fake_label=torch.zeros(num_img,1)
            fake_label=fake_label.to(device)

            # 训练判别器
            real_out=D(real_img)
            d_loss_real=criterion(real_out,real_label)

            fake_img=G(fake_img).detach()
            fake_out=D(fake_img)
            d_loss_fake=criterion(fake_out,fake_label)

            d_loss=d_loss_real+d_loss_fake

            optimizer_D.zero_grad()
            d_loss.backward()
            optimizer_D.step()

            # 训练生成器
            fake_img=torch.randn(num_img,noise_size)
            fake_img=fake_img.to(device)

            fake_img=G(fake_img)
            fake_out=D(fake_img)
            g_loss=criterion(fake_out,real_label)

            optimizer_G.zero_grad()
            g_loss.backward()
            optimizer_G.step()

        print('epoch :{}, d_loss:{}, g_loss:{} '.format(epoch,d_loss.item(),g_loss.item()))
        fake_image=to_img(fake_img.cpu().data)
        save_img(fake_image,epoch)

        torch.save(D.state_dict(), 'models/DCGAN-Mnist/discrimintor.pth')
        torch.save(G.state_dict(),'models/DCGAN-Mnist/generator.pth')

if __name__ == '__main__':
    transformer = torchvision.transforms.Compose([
        torchvision.transforms.ToTensor(),
        torchvision.transforms.Normalize((0.5,), (0.5,))
    ])
    dataset=MNIST(root='mnist',train=True,transform=transformer,download=True)
    dataloader=DataLoader(dataset=dataset,shuffle=True,batch_size=512)
    epoch=500
    noise_size=100

    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    D = Discirmintor()
    G = Generator()
    D = D.to(device)
    G = G.to(device)
    criterion = nn.BCELoss()
    optimizer_G = torch.optim.Adam(G.parameters(), lr=0.0003)
    optimizer_D = torch.optim.Adam(D.parameters(), lr=0.0001)

    train(epoch)

训练250个epoch的结果