GAN网络原理介绍和代码

 

GAN网络的整体公式：

公式各参数介绍如下：

X是真实地图片，而对应的标签是1。

G(Z)是通过给定的噪声Z，生成图片(实际上是通过给定的Z生成一个tensor)，对应的标签是0。

D是一个二分类网络，对于给定的图片判别真假。

 

D和G的参数更新方式：

D通过输入的真假图片，通过BCE(二分类交叉熵)更新自己的参数。

D对G(Z)生成的标签L，G尽可能使L为true，也就是1，通过BCE(二分类交叉熵)更新自己的参数。

 

公式演变：

对于G来说要使D无法判别自己生成的图片是假的，故而要使G(Z)越大越好，所以就使得V(G,D)越小越好；而对于D，使G(Z)越小D(X)越大，故而使V(G,D)越大越好

 

为了便于求导，故而加了log，变为如下：

 

最后对整个batch求期望，变为如下：

 

 

基于mnist实现的GAN网络结构对应的代码 

import itertools
import math
import time

import torch
import torchvision
import torch.nn as nn
import torchvision.datasets as dsets
import torchvision.transforms as transforms
import matplotlib.pyplot as plt
from IPython import display
from torch.autograd import Variable
transform = transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
])

train_dataset = dsets.MNIST(root='./data/', train=True, download=True, transform=transform)
train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=100, shuffle=True)


class Discriminator(nn.Module):
    def __init__(self):
        super().__init__()
        self.model = nn.Sequential(
            nn.Linear(784, 1024),
            nn.LeakyReLU(0.2, inplace=True),
            nn.Dropout(0.3),
            nn.Linear(1024, 512),
            nn.LeakyReLU(0.2, inplace=True),
            nn.Dropout(0.3),
            nn.Linear(512, 256),
            nn.LeakyReLU(0.2, inplace=True),
            nn.Dropout(0.3),
            nn.Linear(256, 1),
            nn.Sigmoid()
        )

    def forward(self, x):
        out = self.model(x.view(x.size(0), 784))
        out = out.view(out.size(0), -1)
        return out


class Generator(nn.Module):
    def __init__(self):
        super().__init__()
        self.model = nn.Sequential(
            nn.Linear(100, 256),
            nn.LeakyReLU(0.2, inplace=True),
            nn.Linear(256, 512),
            nn.LeakyReLU(0.2, inplace=True),
            nn.Linear(512, 1024),
            nn.LeakyReLU(0.2, inplace=True),
            nn.Linear(1024, 784),
            nn.Tanh()
        )

    def forward(self, x):
        x = x.view(x.size(0), -1)
        out = self.model(x)
        return out


discriminator = Discriminator().cuda()
generator = Generator().cuda()
criterion = nn.BCELoss()
lr = 0.0002
d_optimizer = torch.optim.Adam(discriminator.parameters(), lr=lr)
g_optimizer = torch.optim.Adam(generator.parameters(), lr=lr)


def train_discriminator(discriminator, images, real_labels, fake_images, fake_labels):
    discriminator.zero_grad()
    outputs = discriminator(images)
    real_loss = criterion(outputs, real_labels)
    real_score = outputs

    outputs = discriminator(fake_images)
    fake_loss = criterion(outputs, fake_labels)
    fake_score = outputs

    d_loss = real_loss + fake_loss
    d_loss.backward()
    d_optimizer.step()
    return d_loss, real_score, fake_score
def train_generator(generator, discriminator_outputs, real_labels):
    generator.zero_grad()
    g_loss = criterion(discriminator_outputs, real_labels)
    g_loss.backward()
    g_optimizer.step()
    return g_loss

# draw samples from the input distribution to inspect the generation on training
num_test_samples = 16
test_noise = Variable(torch.randn(num_test_samples, 100).cuda())
# create figure for plotting
size_figure_grid = int(math.sqrt(num_test_samples))
fig, ax = plt.subplots(size_figure_grid, size_figure_grid, figsize=(6, 6))
for i, j in itertools.product(range(size_figure_grid), range(size_figure_grid)):
    ax[i, j].get_xaxis().set_visible(False)
    ax[i, j].get_yaxis().set_visible(False)

# set number of epochs and initialize figure counter
num_epochs = 200
num_batches = len(train_loader)
num_fig = 0

for epoch in range(num_epochs):
    for n, (images, _) in enumerate(train_loader):
        images = Variable(images.cuda())
        real_labels = Variable(torch.ones(images.size(0)).cuda())

        # Sample from generator
        noise = Variable(torch.randn(images.size(0), 100).cuda())
        fake_images = generator(noise)
        fake_labels = Variable(torch.zeros(images.size(0)).cuda())

        # Train the discriminator
        d_loss, real_score, fake_score = train_discriminator(discriminator, images, real_labels, fake_images,
                                                             fake_labels)

        # Sample again from the generator and get output from discriminator
        noise = Variable(torch.randn(images.size(0), 100).cuda())
        fake_images = generator(noise)
        outputs = discriminator(fake_images)

        # Train the generator
        g_loss = train_generator(generator, outputs, real_labels)

        if (n + 1) % 100 == 0:
            test_images = generator(test_noise)

            for k in range(num_test_samples):
                i = k // 4
                j = k % 4
                ax[i, j].cla()
                ax[i, j].imshow(test_images[k, :].data.cpu().numpy().reshape(28, 28), cmap='Greys')
            display.clear_output(wait=True)
            display.display(plt.gcf())

            plt.savefig('results/mnist-gan-%03d.png' % num_fig)
            num_fig += 1
            print('Epoch [%d/%d], Step[%d/%d], d_loss: %.4f, g_loss: %.4f, '
                  'D(x): %.2f, D(G(z)): %.2f'
                  % (epoch + 1, num_epochs, n + 1, num_batches, d_loss.data[0], g_loss.data[0],
                     real_score.data.mean(), fake_score.data.mean()))

fig.close()