背景知识
DCGAN
全称Deep Convolutional Generative Adversarial Network,是GAN的改进。
DCGAN的特点包括
1. 生成器使用卷积转置层,又称反卷积层。这里用于将低维的潜在向量转变为高维的图像。
2. 生成器大部分层使用ReLU激活函数,输出层使用Tanh激活函数。判别器的大部分层使用LeakyReLU激活函数。
3. DCGAN不使用全连接层,完全依赖卷积层处理数据。
实验
数据处理
batch_size = 128 # 批量大小
image_size = 64 # 训练图像空间大小
nc = 3 # 图像彩色通道数
nz = 100 # 隐向量的长度
ngf = 64 # 特征图在生成器中的大小
ndf = 64 # 特征图在判别器中的大小
num_epochs = 3 # 训练周期数
lr = 0.0002 # 学习率
beta1 = 0.5 # Adam优化器的beta1超参数
数据加载
def create_dataset_imagenet(dataset_path):
dataset = ds.ImageFolderDataset(dataset_path,
num_parallel_workers=4,
shuffle=True,
decode=True)
# 数据增强操作
transforms = [
vision.Resize(image_size),
vision.CenterCrop(image_size),
vision.HWC2CHW(),
lambda x: ((x / 255).astype("float32"))
]
# 数据映射操作
dataset = dataset.project('image')
dataset = dataset.map(transforms, 'image')
# 批量操作
dataset = dataset.batch(batch_size)
return dataset
dataset = create_dataset_imagenet('./faces')
网络构建
模型权重由mean=0,sigma=0.02的正态分布随机初始化
生成器
生成器G的功能是将隐向量z映射到数据空间。具体是通过一系列Conv2dTranspose转置卷积层来完成的,每层都使用批归一化层和ReLU,输出经过tanh使返回在[-1,1]的范围内。论文中的图像如下。
其中,我们通过输入的nz ngf nc控制生成器结构。nz是隐向量z的长度,ngf与通过生成器传播的特征图大小有关,nc是输出图像的通道数
weight_init = Normal(mean=0, sigma=0.02)
gamma_init = Normal(mean=1, sigma=0.02)
class Generator(nn.Cell):
def __init__(self):
super(Generator, self).__init__()
self.generator = nn.SequentialCell(
nn.Conv2dTranspose(nz, ngf * 8, 4, 1, 'valid', weight_init=weight_init),
nn.BatchNorm2d(ngf * 8, gamma_init=gamma_init),
nn.ReLU(),
nn.Conv2dTranspose(ngf * 8, ngf * 4, 4, 2, 'pad', 1, weight_init=weight_init),
nn.BatchNorm2d(ngf * 4, gamma_init=gamma_init),
nn.ReLU(),
nn.Conv2dTranspose(ngf * 4, ngf * 2, 4, 2, 'pad', 1, weight_init=weight_init),
nn.BatchNorm2d(ngf * 2, gamma_init=gamma_init),
nn.ReLU(),
nn.Conv2dTranspose(ngf * 2, ngf, 4, 2, 'pad', 1, weight_init=weight_init),
nn.BatchNorm2d(ngf, gamma_init=gamma_init),
nn.ReLU(),
nn.Conv2dTranspose(ngf, nc, 4, 2, 'pad', 1, weight_init=weight_init),
nn.Tanh()
)
def construct(self, x):
return self.generator(x)
generator = Generator()
判别器
通过一系列的Conv2d、BatchNorm2d和LeakyReLU层对其进行处理,最后通过Sigmoid激活函数得到最终概率。这里使用卷积而不是池化来进行下采样,它可以让网络学习自己的池化特征。
class Discriminator(nn.Cell):
def __init__(self):
super(Discriminator, self).__init__()
self.discriminator = nn.SequentialCell(
nn.Conv2d(nc, ndf, 4, 2, 'pad', 1, weight_init=weight_init),
nn.LeakyReLU(0.2),
nn.Conv2d(ndf, ndf * 2, 4, 2, 'pad', 1, weight_init=weight_init),
nn.BatchNorm2d(ngf * 2, gamma_init=gamma_init),
nn.LeakyReLU(0.2),
nn.Conv2d(ndf * 2, ndf * 4, 4, 2, 'pad', 1, weight_init=weight_init),
nn.BatchNorm2d(ngf * 4, gamma_init=gamma_init),
nn.LeakyReLU(0.2),
nn.Conv2d(ndf * 4, ndf * 8, 4, 2, 'pad', 1, weight_init=weight_init),
nn.BatchNorm2d(ngf * 8, gamma_init=gamma_init),
nn.LeakyReLU(0.2),
nn.Conv2d(ndf * 8, 1, 4, 1, 'valid', weight_init=weight_init),
)
self.adv_layer = nn.Sigmoid()
def construct(self, x):
out = self.discriminator(x)
out = out.reshape(out.shape[0], -1)
return self.adv_layer(out)
discriminator = Discriminator()
定义损失函数与优化器
adversarial_loss = nn.BCELoss(reduction='mean')
# 为生成器和判别器设置优化器
optimizer_D = nn.Adam(discriminator.trainable_params(), learning_rate=lr, beta1=beta1)
optimizer_G = nn.Adam(generator.trainable_params(), learning_rate=lr, beta1=beta1)
optimizer_G.update_parameters_name('optim_g.')
optimizer_D.update_parameters_name('optim_d.')
模型训练
在生成器和判别器的训练中,分别获取训练过程的损失,并在每个周期结束时统计。
def generator_forward(real_imgs, valid):
# 将噪声采样为发生器的输入
z = ops.standard_normal((real_imgs.shape[0], nz, 1, 1))
# 生成一批图像
gen_imgs = generator(z)
# 损失衡量发生器绕过判别器的能力
g_loss = adversarial_loss(discriminator(gen_imgs), valid)
return g_loss, gen_imgs
def discriminator_forward(real_imgs, gen_imgs, valid, fake):
# 衡量鉴别器从生成的样本中对真实样本进行分类的能力
real_loss = adversarial_loss(discriminator(real_imgs), valid)
fake_loss = adversarial_loss(discriminator(gen_imgs), fake)
d_loss = (real_loss + fake_loss) / 2
return d_loss
grad_generator_fn = ms.value_and_grad(generator_forward, None,
optimizer_G.parameters,
has_aux=True)
grad_discriminator_fn = ms.value_and_grad(discriminator_forward, None,
optimizer_D.parameters)
@ms.jit
def train_step(imgs):
valid = ops.ones((imgs.shape[0], 1), mindspore.float32)
fake = ops.zeros((imgs.shape[0], 1), mindspore.float32)
(g_loss, gen_imgs), g_grads = grad_generator_fn(imgs, valid)
optimizer_G(g_grads)
d_loss, d_grads = grad_discriminator_fn(imgs, gen_imgs, valid, fake)
optimizer_D(d_grads)
return g_loss, d_loss, gen_imgs
G_losses = []
D_losses = []
image_list = []
total = dataset.get_dataset_size()
# 循环训练网络
for epoch in range(num_epochs):
generator.set_train()
discriminator.set_train()
# 为每轮训练读入数据
for i, (imgs, ) in enumerate(dataset.create_tuple_iterator()):
g_loss, d_loss, gen_imgs = train_step(imgs)
if i % 100 == 0 or i == total - 1:
# 输出训练记录
print('[%2d/%d][%3d/%d] Loss_D:%7.4f Loss_G:%7.4f' % (
epoch + 1, num_epochs, i + 1, total, d_loss.asnumpy(), g_loss.asnumpy()))
D_losses.append(d_loss.asnumpy())
G_losses.append(g_loss.asnumpy())
# 每个epoch结束后,使用生成器生成一组图片
generator.set_train(False)
fixed_noise = ops.standard_normal((batch_size, nz, 1, 1))
img = generator(fixed_noise)
image_list.append(img.transpose(0, 2, 3, 1).asnumpy())
# 保存网络模型参数为ckpt文件
mindspore.save_checkpoint(generator, "./generator.ckpt")
mindspore.save_checkpoint(discriminator, "./discriminator.ckpt")
生成器和判别器的训练损失图如下:
模型预测
# 从文件中获取模型参数并加载到网络中
mindspore.load_checkpoint("./generator.ckpt", generator)
fixed_noise = ops.standard_normal((batch_size, nz, 1, 1))
img64 = generator(fixed_noise).transpose(0, 2, 3, 1).asnumpy()
fig = plt.figure(figsize=(8, 3), dpi=120)
images = []
for i in range(3):
images.append(np.concatenate((img64[i * 8:(i + 1) * 8]), axis=1))
img = np.clip(np.concatenate((images[:]), axis=0), 0, 1)
plt.axis("off")
plt.imshow(img)
plt.show()
总结
本章使用DCGAN模型完成了漫画图像的生成,其中DCGAN与传统GAN的区别是反卷积层和卷积层在判别器上的使用。