数据集
香港中文大学，CeleA数据集，http://mmlab.ie.cuhk.edu.hk/projects/CelebA.html
总共202599张面部图片。
制作HDF5压缩格式文件
分层数据格式（hierarchical data format）是一种成熟的，开源的压缩数据格式，专门用于存储非常大量的数据。
在《PyTorch 生成对抗网络编程》[英]塔克里·拉希德一书中将数据处理成HDF5格式运行代码，本来想重构一下的。然后发现如果用ImgaeFloder直接读取的话基本要重构所有代码。还是制作HDF5吧！
import os
import zipfile
import h5py
import imageio
hdf5_file = './celeba_aligned_small.h5py'
total_images = 202599
with h5py.File(hdf5_file, 'w') as hf: # 打开h5py文件，文件不存在则会创建文件
count = 0
with zipfile.ZipFile('img_align_celeba.zip', 'r') as zf:
# 这个压缩文件里是一个文件夹img_align_celeba文件夹中有200000多张图片
for i in zf.namelist(): # zf.namelist()返还压缩文件中的文件列表名
# zf.namelist()[0]是'img_align_celeba/' 即文件夹路径
# zf.namelist()[1]是'img_align_celeba/000001.jpg' 即文件夹下的文件路径
if i[-4:] == '.jpg':
ofile = zf.extract(i) # 解压单个文件至ofile中
# 默认解压在当前文件夹即在'./'路径下创建img_align_celeba文件夹，把图片(i)放入文件夹中
# ofile是解压后图片(i)的相对地址是一个字符串
img = imageio.imread(ofile)
# 使用imageio.imread读取图片，此时img打印出来是一个数组
os.remove(ofile) # 用完即弃
# 删除图片，不占存储空间
hf.create_dataset('img_align_celeba/'+str(count)+'.jpg',
data=img, compression='gzip', compression_opts=9)
# compression是压缩方式, compression_opts是压缩程度的参数
# 在celeba_aligned_small.h5py文件中生成组img_align_celeba，在组中保存img数组
count = count + 1
if count % 1000 == 0:
print('images done ...', count)
if count == total_images: # 取前所有图片
break
生成人脸
上篇博文使用的是MLP（全连接神经网络），这次生成人脸换成了CNN（卷积神经网络）。并采用GPU加速,个人认为代码写的不是很好，没有使用多线程，导致训练巨慢，1060训练一轮要4个小时左右，3090也要三个小时左右一轮。抽时间把代码重构一遍，从读取文件开始。
import h5py
import numpy as np
import matplotlib.pyplot as plt
from torch.utils.data import Dataset
import torch
import torch.nn as nn
def crop_centre(img, new_width, new_height):
height, width, _ = img.shape
startx = width // 2 - new_width // 2
starty = height // 2 - new_height // 2
return img[starty:starty + new_height, startx:startx + new_width, :]
def generate_random_image(size):
random_data = torch.rand(size)
return torch.cuda.FloatTensor(random_data)
def generate_random_seed(size):
random_data = torch.randn(size)
return random_data
class CelebADataset(Dataset):
def __init__(self, file):
self.file_object = h5py.File(file, 'r')
self.dataset = self.file_object['img_align_celeba']
def __len__(self):
return len(self.dataset)
def __getitem__(self, index):
if index >= len(self.dataset):
raise IndexError
img = np.array(self.dataset[str(index) + '.jpg'])
img = crop_centre(img, 128, 128)
return torch.cuda.FloatTensor(img).permute(2, 0, 1).view(1, 3, 128, 128) / 255.0
def plot_image(self, index):
img = np.array(self.dataset[str(index) + '.jpg'])
img = crop_centre(img, 128, 128)
plt.imshow(img, interpolation='nearest')
plt.show()
# 构建鉴别器
class Discriminator(nn.Module):
def __init__(self):
# 初始化父类
super(Discriminator, self).__init__()
self.feature = nn.Sequential(
nn.Conv2d(3, 256, kernel_size=8, stride=2),
nn.BatchNorm2d(256),
nn.GELU(),
nn.Conv2d(256, 256, kernel_size=8, stride=2),
nn.BatchNorm2d(256),
nn.GELU(),
nn.Conv2d(256, 3, kernel_size=8, stride=2),
nn.GELU(),
)
self.classifier = nn.Sequential(
nn.Linear(3 * 10 * 10, 1),
nn.Sigmoid()
)
self.loss_function = nn.BCELoss()
# 创建优化器
self.optimiser = torch.optim.Adam(self.parameters(), lr=0.01)
self.counter = 0
self.progress = []
def forward(self, inputs):
x = self.feature(inputs)
x = x.view(-1)
x = self.classifier(x)
return x
def train(self, inputs, targets):
outputs = self.forward(inputs)
loss = self.loss_function(outputs, targets)
# 每训练10此增加计数器
self.counter += 1
if self.counter % 10 == 0:
self.progress.append(loss.item())
if self.counter % 10000 == 0:
print("counter = ", self.counter)
# 清楚梯度，反向传播， 更新权重
self.optimiser.zero_grad()
loss.backward()
self.optimiser.step()
# 构建生成器
class Generator(nn.Module):
def __init__(self):
super(Generator, self).__init__()
self.linear = nn.Sequential(
nn.Linear(100, 3*11*11),
nn.GELU(),
)
self.feature = nn.Sequential(
nn.ConvTranspose2d(3, 256, kernel_size=8, stride=2),
nn.BatchNorm2d(256),
nn.GELU(),
nn.ConvTranspose2d(256, 256, kernel_size=8, stride=2),
nn.BatchNorm2d(256),
nn.GELU(),
nn.ConvTranspose2d(256, 3, kernel_size=8, stride=2, padding=1),
nn.BatchNorm2d(3),
nn.Sigmoid()
)
# 创建优化器
self.optimiser = torch.optim.Adam(self.parameters(), lr=0.01)
self.counter = 0
self.progress = []
def forward(self, x):
x = self.linear(x)
x = x.view(1, 3, 11, 11)
x = self.feature(x)
return x
def train(self, D, inputs, targets): # 用鉴别器的损失来训练生成，discriminator
g_output = self.forward(inputs) # 生成器generator的输出(g_output)
d_output = D.forward(g_output) # 分类器discriminator的输出,输入生成器generator的输出
loss = D.loss_function(d_output, targets)
self.counter += 1
if self.counter % 10 == 0:
self.progress.append(loss.item())
self.optimiser.zero_grad()
loss.backward()
self.optimiser.step()
if torch.cuda.is_available():
torch.set_default_tensor_type(torch.cuda.FloatTensor)
print('using cuda:', torch.cuda.get_device_name(0))
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# 现在浮点类型默认cuda张量
# x = generate_random_image(2)
# print(x.device)
celeba_dataset = CelebADataset('celeba_aligned_small.h5py')
# celeba_dataset.plot_image(66) # 输出一张图试试
D = Discriminator()
G = Generator()
G.to(device)
D.to(device)
epoches = 1
for epoch in range(epoches):
print('开始第', epoch+1, '轮', '*************'*4)
for image_data in celeba_dataset:
# 真实样本1，训练鉴别器
D.train(image_data, torch.cuda.FloatTensor([1.0]))
# 随机生成噪声，告诉鉴别器这是0
D.train(generate_random_image((1, 3, 128, 128)), torch.cuda.FloatTensor([0.0]))
# 训练生成器
G.train(D, generate_random_seed(100), torch.cuda.FloatTensor([1.0]))
# torch.save(G, 'face-cnn-1.pth')
生成图片
for i in range(6):
output = G.forward(generate_random_seed(100))
img = output.detach().permute(0,2,3,1).view(128,128,3).cpu().numpy()
plt.subplot(2, 3, i+1)
plt.imshow(img)
plt.show()
运行一轮我这1060跑了4个小时，设备好的炼丹师们可以多运行几轮试试。
感觉他这代码写的不咋好，而且用HDF5文件格式有17G大的数据，原始数据还不到1.5G
给他重构了一下：
import numpy as np
import matplotlib.pyplot as plt
from torch.utils.data import Dataset
import torch
import torch.nn as nn
import torch.utils.data as Data
from torchvision import transforms
from torchvision.datasets import ImageFolder
train_data_dir = r'face'
batchsize=32
train_data_transforms = transforms.Compose([
transforms.CenterCrop(128), # 128
transforms.ToTensor(),
])
train_data = ImageFolder(train_data_dir, transform=train_data_transforms)
train_data_loader = Data.DataLoader(
train_data,
batch_size=batchsize,
shuffle=True,
num_workers=32,
drop_last=True
)
# 构建鉴别器
class Discriminator(nn.Module):
def __init__(self):
# 初始化父类
super(Discriminator, self).__init__()
self.feature = nn.Sequential(
nn.Conv2d(3, 256, kernel_size=8, stride=2),
nn.BatchNorm2d(256),
nn.GELU(),
nn.Conv2d(256, 256, kernel_size=8, stride=2),
nn.BatchNorm2d(256),
nn.GELU(),
nn.Conv2d(256, 3, kernel_size=8, stride=2),
nn.GELU(),
)
self.classifier = nn.Sequential(
nn.Linear(3 * 10 * 10, 1),
nn.Sigmoid()
)
def forward(self, inputs):
x = self.feature(inputs)
x = x.view(x.size(0), -1)
x = self.classifier(x)
return x
# 构建生成器
class Generator(nn.Module):
def __init__(self):
super(Generator, self).__init__()
self.linear = nn.Sequential(
nn.Linear(100, 3*11*11),
nn.GELU(),
)
self.feature = nn.Sequential(
nn.ConvTranspose2d(3, 256, kernel_size=8, stride=2),
nn.BatchNorm2d(256),
nn.GELU(),
nn.ConvTranspose2d(256, 256, kernel_size=8, stride=2),
nn.BatchNorm2d(256),
nn.GELU(),
nn.ConvTranspose2d(256, 3, kernel_size=8, stride=2, padding=1),
nn.BatchNorm2d(3),
nn.Sigmoid()
)
def forward(self, x):
x = self.linear(x)
x = x.view(batchsize, 3, 11, 11)
x = self.feature(x)
return x
D = Discriminator()
G = Generator()
d_optimizer = torch.optim.SGD(D.parameters(), lr=0.01)
g_optimizer = torch.optim.SGD(G.parameters(), lr=0.01)
loss_func = nn.BCELoss()
if torch.cuda.is_available():
print('using cuda:', torch.cuda.get_device_name(0))
D = D.cuda()
G = G.cuda()
if __name__ == '__main__':
d_z_loss = []
epoches = 25
for epoch in range(epoches):
print('开始第', epoch + 1, '轮', '*******************' * 3)
sum_d_loss = 0
for step, (b_x, _) in enumerate(train_data_loader): # b_x shape(32,3,128,128)
# 训练辨别器,real data
real_out = D(b_x.cuda())
real_out = real_out.squeeze() # (batch_size,1) -> (batchsize)
real_label = torch.ones(batchsize).cuda()
d_loss_real = loss_func(real_out, real_label) # 标签为1,为真实样本,辨别器的loss
# fake data
fake_img = G(torch.rand(batchsize, 100).cuda()).detach()
fake_out = D(fake_img).squeeze()
fake_label = torch.zeros(batchsize).cuda()
d_loss_fake = loss_func(fake_out, fake_label)
# 更新辨别器
d_loss = (d_loss_real + d_loss_fake)
sum_d_loss += d_loss.item()
d_optimizer.zero_grad() # 在反向传播之前，先将梯度归0
d_loss.backward() # 将误差反向传播
d_optimizer.step() # 更新参数
# 训练生成器
fake_img = G(torch.rand(batchsize, 100).cuda())
output = D(fake_img).squeeze()
g_loss = loss_func(output, real_label)
g_optimizer.zero_grad() # 梯度归0
g_loss.backward() # 进行反向传播
g_optimizer.step() # .step()一般用在反向传播后面,用于更新生成网络的参数
fake_img = G(torch.rand(batchsize, 100).cuda())
output = D(fake_img).squeeze()
g_loss = loss_func(output, real_label)
g_optimizer.zero_grad() # 梯度归0
g_loss.backward() # 进行反向传播
g_optimizer.step() # .step()一般用在反向传播后面,用于更新生成网络的参数
print('{:.5%}'.format(step / len(train_data_loader)))
d_z_loss.append(sum_d_loss / len(train_data_loader))
print(d_z_loss)
if epoch % 2 == 0:
name = 'face'+str(epoch)+'.pth'
torch.save(G, name, _use_new_zipfile_serialization=False)
torch.save(D,'Generator.pth', _use_new_zipfile_serialization=False)
扫一扫
632
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
