-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathgan_example.py
150 lines (105 loc) · 3.86 KB
/
gan_example.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
import torch
import torch.nn as nn
from torch import optim
from torch.utils.data import DataLoader
from torchvision import transforms
from torchvision import utils
import torchvision.datasets as dsets
import cv2
import numpy as np
from matplotlib import pyplot as plt
from tqdm import tqdm
import models
'''
Functions
'''
def sample_z(batch_size, d_noise=100):
return torch.rand(batch_size, d_noise, device=device)
def imshow_grid(img):
img = utils.make_grid(img.cpu().detach())
img = (img+1)/2
npimg = img.numpy()
plt.imshow(np.transpose(npimg, (1,2,0)))
plt.show()
def run_epoch(generator, discriminator, g_optimizer, d_optimizer):
generator.train()
discriminator.train()
criterion = nn.BCELoss()
for img_batch, label_batch in train_data_loader:
img_batch, label_batch = img_batch.to(device), label_batch.to(device)
img_batch = img_batch.reshape(-1, 28 * 28)
d_optimizer.zero_grad()
p_real = discriminator(img_batch)
p_fake = discriminator(generator(sample_z(batch_size)))
# loss_real = (-1) * torch.log(p_real)
# loss_fake = (-1) * torch.log(1 - p_fake)
# loss_d = (loss_real + loss_fake).mean()
loss_d = criterion(p_real, torch.ones(p_real.size()).to(device)) + criterion(
p_fake, torch.zeros(p_fake.size()).to(device)
)
loss_d.backward()
d_optimizer.step()
g_optimizer.zero_grad()
p_fake = discriminator(generator(sample_z(batch_size)))
loss_g = criterion(p_fake, torch.ones(p_fake.size()).to(device))
loss_g.backward()
g_optimizer.step()
def evaluate_model(generator, discriminator):
p_real, p_fake = 0.0, 0.0
generator.eval()
discriminator.eval()
for img_batch, label_batch in test_data_loader:
img_batch, label_batch = img_batch.to(device), label_batch.to(device)
with torch.autograd.no_grad():
p_real += (torch.sum(discriminator(img_batch.view(-1, 28 * 28))).item())/10000.0
p_fake += (torch.sum(discriminator(generator(sample_z(batch_size)))).item())/10000.0
return p_real, p_fake
def init_params(model):
for p in model.parameters():
if p.dim() > 1:
nn.init.xavier_normal_(p)
else:
nn.init.uniform_(p, 0.1, 0.2)
batch_size = 200
dropout_p = 0.1
d_noise = 100
d_hidden = 256
epoch_num = 200
device = torch.device("cuda:2")
normalize = transforms.Normalize(mean=[0.5, ], std=[0.5, ])
preprocess = transforms.Compose([
transforms.ToTensor(),
normalize
])
# MNIST dataset
train_data = dsets.MNIST(root="data/", train=True, transform=preprocess, download=True)
test_data = dsets.MNIST(root="data/", train=False, transform=preprocess, download=True)
train_data_loader = DataLoader(train_data, batch_size, shuffle=True)
test_data_loader = DataLoader(test_data, batch_size, shuffle=False)
G = models.G().to(device)
D = models.D().to(device)
init_params(G)
init_params(D)
optimizer_g = optim.Adam(G.parameters(), lr=0.0002)
optimizer_d = optim.Adam(D.parameters(), lr=0.0002)
p_real_trace = []
p_fake_trace = []
print("Train Started")
for epoch in tqdm(range(epoch_num)):
run_epoch(G, D, optimizer_g, optimizer_d)
p_real, p_fake = evaluate_model(G, D)
p_real_trace.append(p_real)
p_fake_trace.append(p_fake)
if (epoch + 1) % 50 == 0:
print("(epoch %i/200) p_real: %f, p_g: %f" % (epoch + 1, p_real, p_fake))
cv2.imshow_grid(G(sample_z(16)).view(-1, 1, 28, 28))
# plot loss
plt.plot(p_fake_trace, label='D(x_generated)')
plt.plot(p_real_trace, label='D(x_real)')
plt.legend(bbox_to_anchor=(1.05, 1), loc=2, borderaxespad=0.)
plt.show()
# print test image
vis_loader = torch.utils.data.DataLoader(test_data, 16, True)
img_vis, label_vis = next(iter(vis_loader))
imshow_grid(img_vis)
imshow_grid(G(sample_z(16,100)).view(-1, 1, 28, 28))