-
Notifications
You must be signed in to change notification settings - Fork 0
/
wgan.py
165 lines (144 loc) · 5.7 KB
/
wgan.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
#!/usr/bin/env python
#encoding=utf-8
from torch.autograd import Variable
import torch
from argparse import ArgumentParser
import os
import json
import sys
from model import datest
from torch.utils.data import DataLoader
from model import vocab
from model import networkG,networkD
from torch import optim
from torch.autograd import grad
def criterion_G(D_fake):
D_fake = D_fake.type(torch.FloatTensor)
return -torch.mean(D_fake)
def criterion_D(true_input,fake_input):
return -(torch.mean(true_input)-torch.mean(fake_input))
def gen_random():
return torch.randn((1,3,32,32)).cuda()
def set_fixed_parameter(module):
for i in module.parameters():
i.requires_grad = False
def set_unfixed_parameter(module):
for i in module.parameters():
i.requires_grad = True
#print module
def changeonehot(tensor,vocab_size):
list_ = []
#print tensor,tensor.shape
for i in tensor[0,:]:
index = long(i)
new_tensor = torch.zeros(vocab_size)
new_tensor[index] = 1
list_.append(new_tensor)
return torch.unsqueeze(torch.stack(list_),0)
def cal_penalty(net,true_input,fake_input):
alpha = torch.rand(1,1,1)
alpha = Variable(alpha.expand(true_input.size()).cuda())
interpolates = alpha*true_input + (1-alpha)*fake_input
disc_interpolates = net(interpolates)
gradients = grad(outputs = disc_interpolates,inputs = interpolates,grad_outputs = torch.ones(disc_interpolates.size()).cuda(),create_graph = True,retain_graph = True,only_inputs = True)[0]
gradient_penalty = ((gradients.norm(2,dim = 1)-1)**2).mean()*10
return gradient_penalty
def train(FLAGS):
poem_datest = datest(FLAGS)
poem_data = DataLoader(
poem_datest,
batch_size = FLAGS.batchsize,
#batch_size = 1,
shuffle = True,
num_workers = 1)
netG = networkG(FLAGS)
netG.cuda()
netD = networkD(FLAGS)
netD.cuda()
if FLAGS.type == "test":
if FLAGS.netG_param:
netG.load_state_dict(torch.load(FLAGS.netG_param))
if FLAGS.netD_param:
netD.load_state_dict(torch.load(FLAGS.netD_param))
loss_G = criterion_G
loss_D = criterion_D
optimizer_G = optim.Adam(netG.parameters(),lr = FLAGS.learning_rate_G)
optimizer_D = optim.Adam(netD.parameters(),lr = FLAGS.learning_rate_D)
for epoch in range(1,FLAGS.epochs+1):
for idx,data in enumerate(poem_data):
####################################
#1.get true data and fake data
####################################
netD.zero_grad()
step = data.shape[1]
true_input = Variable(changeonehot(data,FLAGS.vocab_size).cuda())
#print true_input
#true_input = Variable(data.cuda())
#print data.shape
noise = Variable(gen_random())
fake_input = netG(noise,step)
#print fake_input.data.shape
#print true_output.data.shape,fake_output.data.shape
#TODO
#1.without penalty train
#################################
#2.train D
#################################
set_fixed_parameter(netG)
set_unfixed_parameter(netD)
#print true_input.data.shape,fake_input.data.shape
true_output = netD(true_input)
fake_output = netD(fake_input)
loss_d = loss_D(true_output,fake_output)
gradient_penalty = cal_penalty(netD,true_input,fake_input)
loss_total_d = loss_d + gradient_penalty
loss_total_d = loss_d
loss_total_d.backward()
optimizer_D.step()
#########################
#3.train_G
#########################
set_fixed_parameter(netD)
set_unfixed_parameter(netG)
netG.zero_grad()
noise_input = Variable(gen_random(),requires_grad=True)
step = 24
fake_output = netD(netG(noise_input,step))
loss_g = loss_G(fake_output)
loss_g.backward()
optimizer_G.step()
if idx % 200 == 199:
print "111"
print "%d epoch %d run: lossd = %.3f,lossg = %.3f" % (epoch,idx+1,loss_d.data[0],loss_g.data[0])
tmp_output = netG(Variable(gen_random()),24)
_,softmax = torch.max(tmp_output,2)
softmax = softmax.data
sentence = []
for i in softmax[0,:]:
index = int(i)
sentence.append(FLAGS.vocab[index])
print "".join(sentence).encode('utf-8')
netGsavename = os.path.join(FLAGS.save_dir,"netG%d-%d.pth" %(epoch,idx+1))
netDsavename = os.path.join(FLAGS.save_dir,"netD%d-%d.pth" %(epoch,idx+1))
torch.save(netG.state_dict(),netGsavename)
torch.save(netD.state_dict(),netDsavename)
if __name__ == "__main__":
parser = ArgumentParser()
parser.add_argument("--type",type = str)
parser.add_argument("--batchsize",type = int)
parser.add_argument("--input",type = str)
parser.add_argument("--vocab_dir",type = str)
parser.add_argument("--epochs",type = int)
parser.add_argument("--learning_rate_G",type = float)
parser.add_argument("--learning_rate_D", type = float)
parser.add_argument('--save_dir',type = str)
parser.add_argument('--netG_param',type = str,default = None)
parser.add_argument('--netD_param',type = str,default = None)
FLAGS = parser.parse_args()
FLAGS.vocab = vocab(FLAGS.vocab_dir)
FLAGS.vocab_size = len(FLAGS.vocab)
print FLAGS
if FLAGS.type == "train":
train(FLAGS)
if FLAGS.type == "predict":
predict(FLAGS)