forked from chenyuntc/pytorch-book
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathmain.py
executable file
·194 lines (157 loc) · 6.67 KB
/
main.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
# coding:utf8
import os
import ipdb
import torch as t
import torchvision as tv
import tqdm
from model import NetG, NetD
from torchnet.meter import AverageValueMeter
class Config(object):
data_path = 'data/' # 数据集存放路径
num_workers = 4 # 多进程加载数据所用的进程数
image_size = 96 # 图片尺寸
batch_size = 256
max_epoch = 200
lr1 = 2e-4 # 生成器的学习率
lr2 = 2e-4 # 判别器的学习率
beta1 = 0.5 # Adam优化器的beta1参数
gpu = True # 是否使用GPU
nz = 100 # 噪声维度
ngf = 64 # 生成器feature map数
ndf = 64 # 判别器feature map数
save_path = 'imgs/' # 生成图片保存路径
vis = True # 是否使用visdom可视化
env = 'GAN' # visdom的env
plot_every = 20 # 每间隔20 batch，visdom画图一次
debug_file = '/tmp/debuggan' # 存在该文件则进入debug模式
d_every = 1 # 每1个batch训练一次判别器
g_every = 5 # 每5个batch训练一次生成器
save_every = 10 # 没10个epoch保存一次模型
netd_path = None # 'checkpoints/netd_.pth' #预训练模型
netg_path = None # 'checkpoints/netg_211.pth'
# 只测试不训练
gen_img = 'result.png'
# 从512张生成的图片中保存最好的64张
gen_num = 64
gen_search_num = 512
gen_mean = 0 # 噪声的均值
gen_std = 1 # 噪声的方差
opt = Config()
def train(**kwargs):
for k_, v_ in kwargs.items():
setattr(opt, k_, v_)
device=t.device('cuda') if opt.gpu else t.device('cpu')
if opt.vis:
from visualize import Visualizer
vis = Visualizer(opt.env)
# 数据
transforms = tv.transforms.Compose([
tv.transforms.Resize(opt.image_size),
tv.transforms.CenterCrop(opt.image_size),
tv.transforms.ToTensor(),
tv.transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
dataset = tv.datasets.ImageFolder(opt.data_path, transform=transforms)
dataloader = t.utils.data.DataLoader(dataset,
batch_size=opt.batch_size,
shuffle=True,
num_workers=opt.num_workers,
drop_last=True
)
# 网络
netg, netd = NetG(opt), NetD(opt)
map_location = lambda storage, loc: storage
if opt.netd_path:
netd.load_state_dict(t.load(opt.netd_path, map_location=map_location))
if opt.netg_path:
netg.load_state_dict(t.load(opt.netg_path, map_location=map_location))
netd.to(device)
netg.to(device)
# 定义优化器和损失
optimizer_g = t.optim.Adam(netg.parameters(), opt.lr1, betas=(opt.beta1, 0.999))
optimizer_d = t.optim.Adam(netd.parameters(), opt.lr2, betas=(opt.beta1, 0.999))
criterion = t.nn.BCELoss().to(device)
# 真图片label为1，假图片label为0
# noises为生成网络的输入
true_labels = t.ones(opt.batch_size).to(device)
fake_labels = t.zeros(opt.batch_size).to(device)
fix_noises = t.randn(opt.batch_size, opt.nz, 1, 1).to(device)
noises = t.randn(opt.batch_size, opt.nz, 1, 1).to(device)
errord_meter = AverageValueMeter()
errorg_meter = AverageValueMeter()
epochs = range(opt.max_epoch)
for epoch in iter(epochs):
for ii, (img, _) in tqdm.tqdm(enumerate(dataloader)):
real_img = img.to(device)
if ii % opt.d_every == 0:
# 训练判别器
optimizer_d.zero_grad()
## 尽可能的把真图片判别为正确
output = netd(real_img)
error_d_real = criterion(output, true_labels)
error_d_real.backward()
## 尽可能把假图片判别为错误
noises.data.copy_(t.randn(opt.batch_size, opt.nz, 1, 1))
fake_img = netg(noises).detach() # 根据噪声生成假图
output = netd(fake_img)
error_d_fake = criterion(output, fake_labels)
error_d_fake.backward()
optimizer_d.step()
error_d = error_d_fake + error_d_real
errord_meter.add(error_d.item())
if ii % opt.g_every == 0:
# 训练生成器
optimizer_g.zero_grad()
noises.data.copy_(t.randn(opt.batch_size, opt.nz, 1, 1))
fake_img = netg(noises)
output = netd(fake_img)
error_g = criterion(output, true_labels)
error_g.backward()
optimizer_g.step()
errorg_meter.add(error_g.item())
if opt.vis and ii % opt.plot_every == opt.plot_every - 1:
## 可视化
if os.path.exists(opt.debug_file):
ipdb.set_trace()
fix_fake_imgs = netg(fix_noises)
vis.images(fix_fake_imgs.detach().cpu().numpy()[:64] * 0.5 + 0.5, win='fixfake')
vis.images(real_img.data.cpu().numpy()[:64] * 0.5 + 0.5, win='real')
vis.plot('errord', errord_meter.value()[0])
vis.plot('errorg', errorg_meter.value()[0])
if (epoch+1) % opt.save_every == 0:
# 保存模型、图片
tv.utils.save_image(fix_fake_imgs.data[:64], '%s/%s.png' % (opt.save_path, epoch), normalize=True,
range=(-1, 1))
t.save(netd.state_dict(), 'checkpoints/netd_%s.pth' % epoch)
t.save(netg.state_dict(), 'checkpoints/netg_%s.pth' % epoch)
errord_meter.reset()
errorg_meter.reset()
@t.no_grad()
def generate(**kwargs):
"""
随机生成动漫头像，并根据netd的分数选择较好的
"""
for k_, v_ in kwargs.items():
setattr(opt, k_, v_)
device=t.device('cuda') if opt.gpu else t.device('cpu')
netg, netd = NetG(opt).eval(), NetD(opt).eval()
noises = t.randn(opt.gen_search_num, opt.nz, 1, 1).normal_(opt.gen_mean, opt.gen_std)
noises = noises.to(device)
map_location = lambda storage, loc: storage
netd.load_state_dict(t.load(opt.netd_path, map_location=map_location))
netg.load_state_dict(t.load(opt.netg_path, map_location=map_location))
netd.to(device)
netg.to(device)
# 生成图片，并计算图片在判别器的分数
fake_img = netg(noises)
scores = netd(fake_img).detach()
# 挑选最好的某几张
indexs = scores.topk(opt.gen_num)[1]
result = []
for ii in indexs:
result.append(fake_img.data[ii])
# 保存图片
tv.utils.save_image(t.stack(result), opt.gen_img, normalize=True, range=(-1, 1))
if __name__ == '__main__':
import fire
fire.Fire()