-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathutils.py
146 lines (123 loc) · 4.18 KB
/
utils.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
import torch.nn as nn
import tqdm
from torch.autograd import Variable as V
import torch.nn.functional as F
import torch
import random
def reset_grad(M):
if not isinstance(M,list):
M=[M]
for m in M:
m.zero_grad()
def train_DCGAN_discriminator(D,G,data,D_optimizer,args,writer):
y_real_ = V(torch.ones(data['real_batch'].size()[0]))
y_fake_ = V(torch.zeros(data['noise'].size()[0]))
BCE_loss=nn.BCELoss()
if args.is_cuda:
y_real_,y_fake_=y_real_.cuda(),y_fake_.cuda()
D_real_loss=BCE_loss(D(data['real_batch']).squeeze(),y_real_)
G_result=G(data['noise'])
# print ("D_real_loss",D_real_loss)
D_fake_loss=BCE_loss(D(G_result).squeeze(),y_fake_)
# print ("D_fake_loss",D_fake_loss)
D_loss=D_real_loss+D_fake_loss
D_loss.backward()
D_optimizer.step()
reset_grad([D,G])
writer.add_scalar("Discriminator/Error",D_loss,args.epoch)
writer.export_scalars_to_json("./all_scalars.json")
def train_DCGAN_generator(D,G,data,G_optimizer,args,writer):
BCE_loss=nn.BCELoss()
y_real_ = V(torch.ones(data['noise'].size()[0]))
if args.is_cuda:
y_real_=y_real_.cuda()
G_result=G(data['noise'])
G_loss=BCE_loss(D(G_result).squeeze(),y_real_)
# print ("G_real_loss",G_loss)
G_loss.backward()
G_optimizer.step()
reset_grad([D,G])
writer.add_scalar("Generator/Error",G_loss,args.epoch)
writer.export_scalars_to_json("./all_scalars.json")
def train_WGAN_discriminator(D,G,data,D_optimizer,args,writer):
for critic_repetitions in range(args.n_critic):
real_embed=D(data['real_batch']).mean()
G_result=G(data['noise'])
fake_embed=D(G_result).mean()
if args.is_cuda:
Penalty=V(torch.zeros(1)).cuda()
else:
Penalty=V(torch.zeros(1))
if args.is_GP:
alpha=random.random()
x_hat = (alpha*G_result+(1-alpha)*data['real_batch']).detach()
x_hat.requires_grad = True
# loss_D = D(x_hat).sum()
loss_D = D(x_hat).mean()
loss_D.backward()
x_hat.grad.volatile = False
Penalty=(((x_hat.grad -1)**2 ).mean())* args.LAMBDA
D_loss=-(real_embed-fake_embed).mean()+Penalty
D_loss.backward()
D_optimizer.step()
if not args.is_GP:
for p in D.parameters():
p.data.clamp_(args.clamp_lower,args.clamp_upper)
reset_grad([D,G])
writer.add_scalar("Discriminator/Error",D_loss,args.epoch)
writer.export_scalars_to_json("./all_scalars.json")
def train_WGAN_generator(D,G,data,G_optimizer,args,writer):
G_result=G(data['noise'])
fake_embed=D(G_result)
G_loss=-fake_embed.mean()
G_loss.backward()
G_optimizer.step()
reset_grad([D,G])
writer.add_scalar("Generator/Error",G_loss,args.epoch)
writer.export_scalars_to_json("./all_scalars.json")
def train_LSGAN_discriminator(D,G,data,D_optimizer,args,writer):
real_embed=D(data['real_batch'])
G_result=G(data['noise'])
fake_embed=D(G_result)
D_loss=((real_embed-1)**2).mean()+args.LAMBDA*((fake_embed)**2).mean()
D_loss.backward()
D_optimizer.step()
reset_grad([D,G])
writer.add_scalar("Discriminator/Error",D_loss,args.epoch)
writer.export_scalars_to_json("./all_scalars.json")
def train_LSGAN_generator(D,G,data,G_optimizer,args,writer):
G_result=G(data['noise'])
fake_embed=D(G_result)
G_loss=args.LAMBDA*((fake_embed-1)**2).mean()
G_loss.backward()
G_optimizer.step()
reset_grad([D,G])
writer.add_scalar("Generator/Error",G_loss,args.epoch)
writer.export_scalars_to_json("./all_scalars.json")
def train_discriminator(D,G,data,D_optimizer,args,writer,type="DCGAN"):
if type=='DCGAN':
train_DCGAN_discriminator(D,G,data,D_optimizer,args,writer)
elif type=='WGAN':
train_WGAN_discriminator(D,G,data,D_optimizer,args,writer)
elif type=='LSGAN':
train_LSGAN_discriminator(D,G,data,D_optimizer,args,writer)
else:
print("Implementation Remaining")
pass
def train_generator(D,G,data,G_optimizer,args,writer,type="DCGAN"):
if type=='DCGAN':
train_DCGAN_generator(D,G,data,G_optimizer,args,writer)
elif type=='WGAN':
train_WGAN_generator(D,G,data,G_optimizer,args,writer)
elif type=='LSGAN':
train_LSGAN_generator(D,G,data,G_optimizer,args,writer)
else:
print("Implementation Remaining")
pass
def validate(G,writer,epoch,args,type='DCGAN'):
z_ = torch.randn((5, 100)).view(-1, 100, 1, 1)
if args.is_cuda:
G_result=G(V(z_).cuda())
else:
G_result=G(V(z_))
writer.add_image("GeneratedImage/"+type,G_result,epoch)