/
train_gdan.py
237 lines (186 loc) · 8.38 KB
/
train_gdan.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
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
import os
import time
import yaml
import pprint
from pathlib import Path
from tqdm import tqdm
import numpy as np
from sklearn.metrics import accuracy_score
from sklearn.metrics.pairwise import cosine_distances
import torch
import torch.nn as nn
import torch.optim as optim
from torch import autograd
from torch.autograd import Variable
from models.gdan import CVAE, Discriminator, Regressor
from utils.config_gdan import parser
from utils.data_factory import DataManager
from utils.utils import load_data, update_values, get_negative_samples
from utils.logger import Logger, log_args
args = parser.parse_args()
# if yaml config exists, load and override default ones
if args.config is not None:
with open(args.config, 'r') as fin:
options_yaml = yaml.load(fin)
update_values(options_yaml, vars(args))
data_dir = Path(args.data_root)
cub_dir = data_dir / Path(args.data_name)
att_path = cub_dir / Path('att_splits.mat')
res_path = cub_dir / Path('res101.mat')
save_dir = Path(args.ckpt_dir)
if not save_dir.is_dir():
save_dir.mkdir(parents=True)
result_dir = Path(args.result)
if not result_dir.is_dir():
result_dir.mkdir(parents=True)
result_path = save_dir / Path('gdan_loss.txt')
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu
pprint.pprint(vars(args))
log_path = save_dir / Path('gdan_log.txt')
print('log file:', log_path)
logMaster = Logger(str(log_path))
log_args(log_path, args)
def main():
logger = logMaster.get_logger('main')
logger.info('loading data...')
att_feats, train_data, val_data, test_data, test_s_data, classes = load_data(att_path=att_path, res_path=res_path)
logger.info('building model...')
gen = CVAE(x_dim=args.x_dim, s_dim=args.s_dim, z_dim=args.z_dim, enc_layers=args.enc, dec_layers=args.dec)
gen.train()
states = torch.load(args.vae_ckpt)
gen.load_state_dict(states['model'])
dis = Discriminator(x_dim=args.x_dim, s_dim=args.s_dim, layers=args.dis)
reg = Regressor(x_dim=args.x_dim, s_dim=args.s_dim, layers=args.reg)
gen.cuda()
dis.cuda()
reg.cuda()
mse_loss = nn.MSELoss()
l1_loss = nn.L1Loss()
adam_betas = (0.8, 0.999)
gen_opt = optim.Adam(gen.parameters(), lr=args.learning_rate, weight_decay=0.01, betas=adam_betas)
dis_opt = optim.Adam(dis.parameters(), lr=args.learning_rate, weight_decay=0.01, betas=adam_betas)
reg_opt = optim.Adam(reg.parameters(), lr=args.learning_rate, weight_decay=0.01, betas=adam_betas)
train_manager = DataManager(train_data, args.epoch, args.batch, infinite=True)
ones = Variable(torch.ones([args.batch, 1]), requires_grad=False).float().cuda()
zeros = Variable(torch.zeros([args.batch, 1]), requires_grad=False).float().cuda()
loss_history = []
logger.info('start training...')
for epoch in range(args.epoch):
running_loss = 0
t1 = time.time()
d_total_loss = 0.0
g_total_loss = 0.0
cyc_total_loss = 0.0
r_total_loss = 0.0
rd_total_loss = 0.0
vae_total_loss = 0.0
g_scores = 0.0
if args.steps == -1:
steps = train_manager.num_batch
else:
steps = args.steps
for batch in tqdm(range(steps), leave=False, ncols=70, unit='b'):
for i in range(args.d_iter):
dis.zero_grad()
# get true data
data = train_manager.get_batch()
X = Variable(torch.from_numpy(np.asarray([item[0] for item in data]))).float().cuda()
Y = [item[1] for item in data]
S = Variable(torch.from_numpy(att_feats[Y])).float().cuda()
Yc = get_negative_samples(Y, classes['train'])
Sc = Variable(torch.from_numpy(att_feats[Yc])).float().cuda()
# get fake data
Xp, _, _ = gen.forward(X, S)
Xp = Xp.detach() # fix the generator
Xpp = gen.sample(S).detach()
Sp = reg.forward(X).detach() # fix the regressor
# get scores
true_scores = dis.forward(X, S)
fake_scores = dis.forward(Xp, S)
fake_scores2 = dis.forward(Xpp, S)
reg_scores = dis.forward(X, Sp)
ctrl_scores = dis.forward(X, Sc)
# calculate loss
d_loss = mse_loss(true_scores, ones) + mse_loss(fake_scores, zeros) + args.theta3 * mse_loss(reg_scores, zeros) \
+ mse_loss(ctrl_scores, zeros)
d_loss.backward()
dis_opt.step()
d_total_loss += d_loss.cpu().data.numpy()
for i in range(args.g_iter):
gen.zero_grad()
reg.zero_grad()
# get true data
data = train_manager.get_batch()
X = Variable(torch.from_numpy(np.asarray([item[0] for item in data]))).float().cuda()
Y = [item[1] for item in data]
S = Variable(torch.from_numpy(att_feats[Y])).float().cuda()
# get fake data
Xp, mu, log_sigma = gen.forward(X, S)
Xp2 = gen.sample(S)
Sp = reg.forward(X)
Spp = reg.forward(Xp)
Xpp, _, _ = gen.forward(X, Sp)
# get scores
fake_scores = dis.forward(Xp, S)
fake_scores2 = dis.forward(Xp2, S)
reg_scores = dis.forward(X, Sp)
# calculate loss
vae_loss = gen.vae_loss(X=X, Xp=Xp, mu=mu, log_sigma=log_sigma)
cyc_loss = mse_loss(Spp, S) + mse_loss(Xpp, X)
g_loss = mse_loss(fake_scores, ones)
r_loss = mse_loss(Sp, S)
rd_loss = mse_loss(reg_scores, ones)
total_loss = vae_loss + g_loss + args.theta1 * cyc_loss + args.theta2 * r_loss + args.theta3 * rd_loss
total_loss.backward()
gen_opt.step()
reg_opt.step()
vae_total_loss += vae_loss.cpu().data.numpy()
g_total_loss += g_loss.cpu().data.numpy()
cyc_total_loss += cyc_loss.cpu().data.numpy()
r_total_loss += r_loss.cpu().data.numpy()
rd_total_loss += rd_loss.cpu().data.numpy()
g_scores += np.mean(fake_scores.cpu().data.numpy())
g_total_steps = steps * args.g_iter
d_total_steps = steps * args.d_iter
vae_avg_loss = vae_total_loss / g_total_steps
g_avg_loss = g_total_loss / g_total_steps
cyc_avg_loss = cyc_total_loss / g_total_steps
r_avg_loss = r_total_loss / g_total_steps
rd_avg_loss = rd_total_loss / g_total_steps
d_avg_loss = d_total_loss / d_total_steps
g_avg_score = g_scores / g_total_steps
loss_history.append(f'{g_avg_loss:.4}\t{d_avg_loss:.4}\t{cyc_avg_loss:.4}\t{r_avg_loss:.4}\t'
f'{rd_avg_loss:.4}\t{g_avg_score:.4}\t{vae_avg_loss:.4}\n')
elapsed = (time.time() - t1)/60.0
if (epoch+1) % 10 == 0 or epoch == 0:
filename = 'gdan_' + str(epoch + 1) + '.pkl'
save_path = save_dir / Path(filename)
states = dict()
states['epoch'] = epoch + 1
states['gen'] = gen.state_dict()
states['dis'] = dis.state_dict()
states['reg'] = reg.state_dict()
states['enc_layers'] = args.enc
states['dec_layers'] = args.dec
states['reg_layers'] = args.reg
states['dis_layers'] = args.dis
states['z_dim'] = args.z_dim
states['x_dim'] = args.x_dim
states['s_dim'] = args.s_dim
states['gen_opt'] = gen_opt.state_dict()
states['dis_opt'] = dis_opt.state_dict()
states['reg_opt'] = reg_opt.state_dict()
states['theta1'] = args.theta1
states['theta2'] = args.theta2
states['theta3'] = args.theta3
torch.save(states, str(save_path))
logger.info(f'epoch: {epoch+1:4}, g_loss: {g_avg_loss: .4}, d_loss: {d_avg_loss: .4}, \n'
f'cyc_loss: {cyc_avg_loss: .4}, r_loss: {r_avg_loss: .4}, rd_loss: {rd_avg_loss: .4}, '
f'g_score: {g_avg_score:.4}, vae loss: {vae_avg_loss:.4}')
with result_path.open('w') as fout:
for s in loss_history:
fout.write(s)
logger.info('program finished')
if __name__ == '__main__':
main()