-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
177 lines (138 loc) · 5.67 KB
/
train.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
import os
import torch
from torch import nn
# from torchvision.datasets import CIFAR10, STL10
from model_pro import CVAE, Discriminator
from logger import Logger
from dataset import get_dataloader
from matplotlib import pyplot as plt
from inference import denormalize, to_uint8
from PIL import Image
def test_reverse(x, y, args):
denorm_images = denormalize(x[0:3], mean=args.tran_mean, std=args.tran_std)
uint8_images = to_uint8(denorm_images)
print(uint8_images[0])
plt.figure()
for i, img in enumerate(uint8_images):
# 如果生成的是 1xHxW，调整为 HxWxC
# if img.shape[0] == 3: # 检查是否为 (3, H, W)
# img = np.transpose(img, (1, 2, 0)) # 转换为 (H, W, C)
#
# img = (img * 255).astype(np.uint8) # 转换为 [0, 255]
img = img.permute(1, 2, 0).cpu().numpy() # 转换为 (H, W, C)
img = Image.fromarray(img)
plt.imshow(img)
plt.axis('off')
plt.savefig(f'#{y[i]}.png')
plt.close() # 保存后再关闭
print(f"generated images saved to {args.recon_dir}")
# loss function for cvae
def loss_G(x_recon, x, mean, log, fake_validity, real_labels, args):
recon_loss = nn.MSELoss()(x_recon, x)
kl_div = -0.5 * torch.sum(1 + log - mean.pow(2) - log.exp())
#test
# print(f"fake_validity shape: {fake_validity.shape}, real_labels shape: {real_labels.shape}")
adv_loss = nn.BCELoss()(fake_validity, real_labels)
recon_loss = recon_loss * args.w_recon
kl_div = kl_div * args.w_kl
adv_loss = adv_loss * args.w_adv
# weight to be fixed...
print(f"recon: {recon_loss:.6f}, kl: {kl_div:.6f}, adv: {adv_loss:.6f}, x0: {x.size(0)}")
return (recon_loss + kl_div + adv_loss) / x.size(0)
def loss_D(real_validity, real_labels, fake_validity, fake_labels):
adversarial_loss = nn.BCELoss()
d_real_loss = adversarial_loss(real_validity, real_labels)
d_fake_loss = adversarial_loss(fake_validity, fake_labels)
d_loss = (d_real_loss + d_fake_loss) * 0.5
print(f"d_loss: {d_loss:.6f}")
return d_loss
def train(args):
# logger setup
global real_labels
if args.log_dir is not None and not os.path.exists(args.log_dir):
os.makedirs(args.log_dir)
logger = Logger(os.path.join(args.log_dir, args.log_path))
# keep the best model parameters according to avg_loss
# tracker = {"epoch" : None, "criterion" : None}
# device setup
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
logger.write(f"we're using :: {device}\n\n")
# data preparations
# transform = transforms.Compose([
# transforms.ToTensor(),
# transforms.Normalize(mean=args.tran_mean, std=args.tran_std),
# ])
# cifar = CIFAR10(args.data_dir, train=True, transform=transform, download=True)
# dataset = DataLoader(dataset=cifar, batch_size=args.batch_size, shuffle=True, drop_last=True)
dataset, class_idx = get_dataloader(args)
# model setup
cvae = CVAE(input_channel=args.input_channel, condition_dim=args.num_classes, latent_dim=args.latent_size).to(device)
discriminator = Discriminator(in_channel=args.input_channel, condition_dim=args.num_classes).to(device)
optim_G = torch.optim.Adam(cvae.parameters(), lr=args.g_lr, betas=(0.5, 0.999))
optim_D = torch.optim.Adam(discriminator.parameters(), lr=args.d_lr, betas=(0.5, 0.999))
if args.preTrain == True:
cvae.load_state_dict(torch.load("cvae_model.pth"))
# 打印模型的一个参数，检查加载是否成功
for name, param in cvae.named_parameters():
print(f"{name}: {len(param.data)}")
break
logger.write(f"pretrained model: {args.model_path}\n\n")
real_label = 1.0 # fixed...
fake_label = 0.0
# Training
cvae.train()
discriminator.train()
for epoch in range(args.epochs):
g_epoch_loss = 0
d_epoch_loss = 0
# step_counter = 0
for x, y in dataset:
x, y = x.to(device), y.to(device)
c = nn.functional.one_hot(y, num_classes=args.num_classes).float().to(device) # one-hot encoding
real_labels = torch.full((x.size(0),), real_label, dtype=torch.float, device=device).unsqueeze(-1)
fake_labels = torch.full((x.size(0),), fake_label, dtype=torch.float, device=device).unsqueeze(-1)
## test
# print(x[0])
# print("")
# test_reverse(x, y, args)
# exit(0)
# if step_counter == 0:
optim_D.zero_grad()
optim_G.zero_grad()
""" 数据准备1 """
# 真实图像进行判别
vd_r = discriminator(x, c)
""" 数据准备2 """
# 生成潜在向量
z = torch.randn(x.size(0), args.latent_size, device=device)
# 潜在向量z生成图像，并判别
x_p = cvae.inference(z, c)
vd_p = discriminator(x_p.detach(), c) # fixed...
""" update Discriminator """
d_loss = loss_D(vd_r, real_labels, vd_p, fake_labels)
d_loss.backward(retain_graph=True)
optim_D.step()
d_epoch_loss += d_loss.item()
""" update generator """
for _ in range(args.gd_ratio):
""" 数据准备3 """
x_f, m, log = cvae(x, c)
vd_f = discriminator(x_f, c) # fixed...
g_loss = loss_G(x_f, x, m, log, vd_f, real_labels, args)
g_loss.backward()
optim_G.step()
g_epoch_loss += g_loss.item()
# update counter
# step_counter += 1
# keep the best model parameters according to avg_loss
g_avg_loss = g_epoch_loss / (len(dataset) * args.gd_ratio)
d_avg_loss = d_epoch_loss / len(dataset)
torch.save(cvae.state_dict(), args.model_path)
torch.save(discriminator.state_dict(), args.D_path)
logger.write(f"Epoch {epoch + 1}/{args.epochs}, D_loss: {d_avg_loss:.6f}, G_Loss: {g_avg_loss:.6f}\n")
# end Training
logger.write("\n\nTraining completed\n\n")
if os.path.exists(args.model_path):
logger.write(f"model with the best performance saved to {args.model_path}.\n")
# close
logger.close()