摘要：
记录了昇思MindSpore AI框架使用70171张动漫头像图片训练一个DCGAN神经网络生成式对抗网络，并用来生成漫画头像的过程、步骤。包括环境准备、下载数据集、加载数据和预处理、构造网络、模型训练等。
一、概念
深度卷积对抗生成网络DCGAN
Deep Convolutional Generative Adversarial Networks
扩展GAN
判别器
组成
卷积层
BatchNorm层
LeakyReLU激活层
功能
输入是3*64*64图像
输出是真图像概率
生成器
组成
转置卷积层
BatchNorm层
ReLU激活层
功能
输入是标准正态分布中提取出的隐向量z
输出是3*64*64 RGB图像。
- 环境准备
%%capture captured_output
# 实验环境已经预装了mindspore==2.2.14，如需更换mindspore版本，可更改下面mindspore的版本号
!pip uninstall mindspore -y
!pip install -i https://pypi.mirrors.ustc.edu.cn/simple mindspore==2.2.14三、数据准备与处理
1.下载数据集
下载到指定目录下并解压，代码如下：
from download import download
url = "https://download.mindspore.cn/dataset/Faces/faces.zip"
path = download(url, "./faces", kind="zip", replace=True)输出：
Downloading data from https://download-mindspore.osinfra.cn/dataset/Faces/faces.zip (274.6 MB)file_sizes: 100%|████████████████████████████| 288M/288M [00:52<00:00, 5.49MB/s]
Extracting zip file...
Successfully downloaded / unzipped to ./faces2.数据集介绍
使用的动漫头像数据集共有70,171张动漫头像图片，图片大小均为96*96。
数据集目录结构如下：
./faces/faces
├── 0.jpg
├── 1.jpg
├── 2.jpg
├── 3.jpg
├── 4.jpg...
├── 70169.jpg
└── 70170.jpg3.数据处理
(1) 执行过程参数定义：
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(2) 数据处理和增强
create_dataset_imagenet函数
import numpy as np
import mindspore.dataset as ds
import mindspore.dataset.vision as vision
def create_dataset_imagenet(dataset_path):"""数据加载"""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')(3) 查看训练数据
matplotlib模块
数据转换成字典迭代器
create_dict_iterator函数
import matplotlib.pyplot as plt
def plot_data(data):# 可视化部分训练数据plt.figure(figsize=(10, 3), dpi=140)for i, image in enumerate(data[0][:30], 1):plt.subplot(3, 10, i)plt.axis("off")plt.imshow(image.transpose(1, 2, 0))plt.show()
sample_data = next(dataset.create_tuple_iterator(output_numpy=True))
plot_data(sample_data)
四、构造网络
模型权重随机初始化
范围：mean为0，sigma为0.02的正态分布【数学不好】
1. 生成器
生成器G
隐向量z映射数据空间
数据源是图像
生成与源图像大小相同的 RGB 图像
Conv2dTranspose转置卷积层
每个层与BatchNorm2d层和ReLu激活层配对
tanh函数
输出[-1,1]范围内数据
DCGAN生成图像过程如下所示：
生成器结构参数：
nz 隐向量z的长度
ngf 有关生成器传播的特征图大小
nc 输出图像通道数
生成器代码：
import mindspore as ms
from mindspore import nn, ops
from mindspore.common.initializer import Normal
weight_init = Normal(mean=0, sigma=0.02)
gamma_init = Normal(mean=1, sigma=0.02)
class Generator(nn.Cell):"""DCGAN网络生成器"""
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()2. 判别器
判别器D
二分类网络模型
Conv2d
BatchNorm2d
LeakyReLU
Sigmoid激活函数
输出判定图像真实概率
判别器代码：
class Discriminator(nn.Cell):"""DCGAN网络判别器"""
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()五、模型训练
1. 损失函数
二进制交叉熵损失函数MindSpore.nn.BCELoss
# 定义损失函数
adversarial_loss = nn.BCELoss(reduction='mean')2. 优化器
Adam优化器
lr = 0.0002
beta1 = 0.5
# 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.')3. 训练模型
训练判别器
提高判别图像真伪的概率
Goodfellow方法：提高随机梯度更新判别器
最大化logD(x)+log(1-D(G(z)))
训练生成器
最小化log(1−D(G(z)))
产生更好的虚拟图像
两个部分分别
获取训练损失
每个周期结束统计
批量推送fixed_noise到生成器
跟踪G的训练进度
模型训练正向逻辑：
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) / 2return 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循环训练网络
迭代50次收集生成器、判别器的损失一次
绘制损失函数的图像
import mindspore
G_losses = []
D_losses = []
image_list = []
total = dataset.get_dataset_size()
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")输出：
[ 1/3][ 1/549] Loss_D: 0.2635 Loss_G: 4.8150
[ 1/3][101/549] Loss_D: 0.4023 Loss_G: 4.9807
[ 1/3][201/549] Loss_D: 0.2425 Loss_G: 1.6335
[ 1/3][301/549] Loss_D: 0.5856 Loss_G: 0.6079
[ 1/3][401/549] Loss_D: 0.1922 Loss_G: 4.3977
[ 1/3][501/549] Loss_D: 0.1065 Loss_G: 2.3724
[ 1/3][549/549] Loss_D: 0.1893 Loss_G: 1.6483
[ 2/3][ 1/549] Loss_D: 0.3370 Loss_G: 4.4347
[ 2/3][101/549] Loss_D: 0.4681 Loss_G: 0.8623
[ 2/3][201/549] Loss_D: 0.1856 Loss_G: 3.7501
[ 2/3][301/549] Loss_D: 0.1932 Loss_G: 2.6333
[ 2/3][401/549] Loss_D: 0.1310 Loss_G: 2.2524
[ 2/3][501/549] Loss_D: 0.2531 Loss_G: 1.4690
[ 2/3][549/549] Loss_D: 0.1192 Loss_G: 5.7166
[ 3/3][ 1/549] Loss_D: 0.0716 Loss_G: 2.9886
[ 3/3][101/549] Loss_D: 0.1345 Loss_G: 2.6544
[ 3/3][201/549] Loss_D: 0.1097 Loss_G: 2.8604
[ 3/3][301/549] Loss_D: 0.2066 Loss_G: 6.1513
[ 3/3][401/549] Loss_D: 0.0797 Loss_G: 3.2336
[ 3/3][501/549] Loss_D: 0.2618 Loss_G: 4.0991
[ 3/3][549/549] Loss_D: 0.5600 Loss_G:10.75094. 结果展示
描绘D和G损失与训练迭代的关系图：
plt.figure(figsize=(10, 5))
plt.title("Generator and Discriminator Loss During Training")
plt.plot(G_losses, label="G", color='blue')
plt.plot(D_losses, label="D", color='orange')
plt.xlabel("iterations")
plt.ylabel("Loss")
plt.legend()
plt.show()输出：
显示隐向量fixed_noise训练生成的图像
import matplotlib.pyplot as plt
import matplotlib.animation as animation
def showGif(image_list):show_list = []fig = plt.figure(figsize=(8, 3), dpi=120)for epoch in range(len(image_list)):images = []for i in range(3):row = np.concatenate((image_list[epoch][i * 8:(i + 1) * 8]), axis=1)images.append(row)img = np.clip(np.concatenate((images[:]), axis=0), 0, 1)plt.axis("off")show_list.append([plt.imshow(img)])
ani = animation.ArtistAnimation(fig, show_list, interval=1000, repeat_delay=1000, blit=True)ani.save('./dcgan.gif', writer='pillow', fps=1)
showGif(image_list)输出：
训练次数增多，图像质量越好
num_epochs达到50以上，生成动漫头像图片与数据集较为相似
加载生成器网络模型参数文件来生成图像代码：
# 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()输出：
