-
Notifications
You must be signed in to change notification settings - Fork 4
/
5gaussian.py
191 lines (158 loc) · 5.81 KB
/
5gaussian.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
import torch
from torch.utils.data import Dataset, DataLoader, TensorDataset
import os
from torchvision import transforms, datasets
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
from matplotlib.pyplot import plot,savefig, clf, scatter, legend, xlim, ylim, hist
import numpy as np
import torch.nn as nn
import torch.nn.functional as F
import utils as myutil
import datasets
seed = 0
torch.manual_seed(seed)
use_cuda = torch.cuda.is_available()
device = torch.device('cuda' if use_cuda else 'cpu')
class generator(torch.nn.Module):
def __init__(self, d):
super(generator, self).__init__()
self.d = d
self.fc1 = torch.nn.Linear(self.d, g_dim)
self.fc2 = torch.nn.Linear(g_dim, g_dim)
self.fc3 = torch.nn.Linear(g_dim, 2)
def forward(self, z):
x = self.fc1(z)
x = torch.nn.functional.relu(x)
x = self.fc2(x)
x = torch.nn.functional.relu(x)
x = self.fc3(x)
return x
class discriminator(torch.nn.Module):
# initializers
def __init__(self, dim):
super(discriminator, self).__init__()
self.dim = dim
l = [torch.nn.Linear(self.dim, d_dim), nn.LeakyReLU(0.2)]
for _ in range(Dlayer_num-2):
l.append(torch.nn.Linear(d_dim, d_dim))
l.append(nn.LeakyReLU(0.2))
l.append(torch.nn.Linear(d_dim, 1, bias=False))
self.net = nn.Sequential(*l)
def forward(self, input):
input = input.float()
return self.net(input)
def get_rsgan_gloss(dis_fake, dis_real):
scalar = torch.FloatTensor([0]).to(device)
z = dis_real - dis_fake
z_star = torch.max(z, scalar.expand_as(z))
return (z_star + torch.log(torch.exp(z - z_star) + torch.exp(0 - z_star))).mean()
def get_rsgan_dloss(dis_fake, dis_real):
scalar = torch.FloatTensor([0]).to(device)
z = dis_fake - dis_real
z_star = torch.max(z, scalar.expand_as(z))
return (z_star + torch.log(torch.exp(z - z_star) + torch.exp(0 - z_star))).mean()
def get_gloss(dis_fake, dis_real):
if model_type == 'rsgan':
return get_rsgan_gloss(dis_fake, dis_real)
elif model_type == 'vanilla':
return (F.softplus(-dis_fake)).mean()
def get_dloss(dis_fake, dis_real):
if model_type == 'rsgan':
return get_rsgan_dloss(dis_fake, dis_real)
elif model_type == 'vanilla':
return (F.softplus(-dis_real)).mean() + (F.softplus(dis_fake)).mean()
def train():
dir_name = '{}_{}gaussian_dlr{}_glr{}_ddim{}_gdim{}_gfreq{}_dfreq{}_seed{}/'.format(
model_type, gaussian_num, d_lr, g_lr, d_dim, g_dim, g_freq, d_freq, seed)
if not os.path.exists(dir_name):
os.makedirs(dir_name)
save_path = dir_name + 'models/'
if not os.path.exists(save_path):
os.makedirs(save_path)
img_path = dir_name + 'images/'
if not os.path.exists(img_path):
os.makedirs(img_path)
d = discriminator(dim=2)
g = generator(input_dim)
d.to(device)
g.to(device)
d_optimizer = torch.optim.Adam(d.parameters(), lr=d_lr, betas=(0, 0.9))
g_optimizer = torch.optim.Adam(g.parameters(), lr=g_lr, betas=(0, 0.9))
g_losses = []
d_losses = []
grad_normD, grad_normG = [], []
loader = datasets.toy_DataLoder(n=data_num//gaussian_num, batch_size=batch_size, gaussian_num=gaussian_num)
data_iter = iter(loader)
for i in range(1, num_iters + 1):
try:
x = next(data_iter)[0].to(device)
z = torch.randn(batch_size, 2).to(device)
except StopIteration:
data_iter = iter(loader)
x = next(data_iter)[0].to(device)
z = torch.randn(batch_size, 2).to(device)
for _ in range(d_freq):
d_optimizer.zero_grad()
x_hat = g(z).detach()
y_hat = d(x_hat)
y = d(x)
d_loss = get_dloss(y_hat, y)
d_losses.append(d_loss.item())
d_loss.backward()
d_optimizer.step()
if model_type == 'rsgan':
list(d.children())[-1][-1].weight.data = torch.nn.functional.normalize(
list(d.children())[-1][-1].weight.data, dim=1)
grad_normD.append(myutil.getGradNorm(d))
for _ in range(g_freq):
g_optimizer.zero_grad()
x_hat = g(z)
y_hat = d(x_hat)
y = d(x)
g_loss = get_gloss(y_hat, y)
g_losses.append(g_loss.item())
g_loss.backward()
g_optimizer.step()
grad_normG.append(myutil.getGradNorm(g))
if i % print_freq == 0:
print('Iteration: {}; G-Loss: {}; D-Loss: {};'.format(i, g_loss, d_loss))
if i % save_freq == 0:
torch.save(g.state_dict(), save_path + 'G_varseed{}_epoch{}.pth'.format(seed, i))
torch.save(d.state_dict(), save_path + 'D_varseed{}_epoch{}.pth'.format(seed, i))
if i == 1 or i % plot_freq == 0:
fake = g(z)
gen_data = fake.data.cpu().numpy()
x = x.detach().cpu().numpy()
plt.scatter(x[:, 0], x[:, 1], c='r')
plt.scatter(gen_data[:, 0], gen_data[:, 1], c='y')
plt.xlim([-5, 5])
plt.ylim([-5, 5])
plt.xticks(fontsize=10 * scale)
plt.yticks(fontsize=10 * scale)
savefig(img_path + '/gaussian_plot%05d.jpg'%(i/plot_freq))
clf()
myutil.saveproj(y.cpu(), y_hat.cpu(), i, save_path)
if __name__ == '__main__':
g_dim = 128
d_dim = 128
Dlayer_num = 3
data_num = 10000
batch_size = 128
g_lr = 1e-4
d_lr = 1e-4
input_dim = 2
num_iters = 20000
print_freq = 500
plot_freq = 100
save_freq = 100
g_freq = 1
d_freq = 1
model_type = 'rsgan'
# model_type = 'vanilla'
loss_type = 'log'
gaussian_num = 5
img_size = 6
scale = img_size / 3
train()