-
Notifications
You must be signed in to change notification settings - Fork 12
/
Copy pathtrainer_warp.py
149 lines (120 loc) · 6.39 KB
/
trainer_warp.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
import os
import glob
import time
import torch
import torch.utils
import torch.nn as nn
import torchvision
from torch.autograd import Variable
from torch.utils.data import DataLoader
from dataset import VaganDataset, LRWdataset
from model_warp import Generator, Discriminator
from tensorboard_logger import configure, log_value
class Trainer():
def __init__(self, config):
self.generator = Generator()
self.discriminator = Discriminator()
print(self.generator)
print(self.discriminator)
self.bce_loss_fn = nn.BCELoss()
self.mse_loss_fn = nn.MSELoss()
self.opt_g = torch.optim.Adam(filter(lambda p: p.requires_grad, self.generator.parameters()),
lr=config.lr, betas=(config.beta1, config.beta2))
self.opt_d = torch.optim.Adam(self.discriminator.parameters(),
lr=config.lr, betas=(config.beta1, config.beta2))
if config.dataset == 'grid':
self.dataset = VaganDataset(config.dataset_dir, train=config.is_train)
elif config.dataset == 'lrw':
self.dataset = LRWdataset(config.dataset_dir, train=config.is_train)
self.data_loader = DataLoader(self.dataset,
batch_size=config.batch_size,
num_workers=config.num_thread,
shuffle=True, drop_last=True)
data_iter = iter(self.data_loader)
data_iter.next()
self.ones = Variable(torch.ones(config.batch_size), requires_grad=False)
self.zeros = Variable(torch.zeros(config.batch_size), requires_grad=False)
if config.cuda:
device_ids = [int(i) for i in config.device_ids.split(',')]
self.generator = nn.DataParallel(self.generator.cuda(), device_ids=device_ids)
self.discriminator = nn.DataParallel(self.discriminator.cuda(), device_ids=device_ids)
self.bce_loss_fn = self.bce_loss_fn.cuda()
self.mse_loss_fn = self.mse_loss_fn.cuda()
self.ones = self.ones.cuda()
self.zeros = self.zeros.cuda()
self.config = config
self.start_epoch = 0
def fit(self):
config = self.config
configure("{}".format(config.log_dir), flush_secs=5)
num_steps_per_epoch = len(self.data_loader)
cc = 0
for epoch in range(self.start_epoch, config.max_epochs):
for step, (example, real_im, landmarks, right_audio, wrong_audio) in enumerate(self.data_loader):
t1 = time.time()
if config.cuda:
example = Variable(example).cuda()
real_im = Variable(real_im).cuda()
right_audio = Variable(right_audio).cuda()
wrong_audio = Variable(wrong_audio).cuda()
else:
example = Variable(example)
real_im = Variable(real_im)
right_audio = Variable(right_audio)
wrong_audio = Variable(wrong_audio)
fake_im = self.generator(example, right_audio)
# Train the discriminator
D_real = self.discriminator(real_im, right_audio)
D_wrong = self.discriminator(real_im, wrong_audio)
D_fake = self.discriminator(fake_im.detach(), right_audio)
loss_real = self.bce_loss_fn(D_real, self.ones)
loss_wrong = self.bce_loss_fn(D_wrong, self.zeros)
loss_fake = self.bce_loss_fn(D_fake, self.zeros)
loss_disc = loss_real + 0.5*(loss_fake + loss_wrong)
loss_disc.backward()
self.opt_d.step()
self._reset_gradients()
# Train the generator
fake_im = self.generator(example, right_audio)
D_fake = self.discriminator(fake_im, right_audio)
loss_gen = self.bce_loss_fn(D_fake, self.ones)
loss_gen.backward()
self.opt_g.step()
self._reset_gradients()
t2 = time.time()
if (step+1) % 1 == 0 or (step+1) == num_steps_per_epoch:
steps_remain = num_steps_per_epoch-step+1 + \
(config.max_epochs-epoch+1)*num_steps_per_epoch
eta = int((t2-t1)*steps_remain)
print("[{}/{}][{}/{}] Loss_D: {:.4f} Loss_G: {:.4f}, ETA: {} second"
.format(epoch+1, config.max_epochs,
step+1, num_steps_per_epoch,
loss_disc.data[0], loss_gen.data[0], eta))
log_value('discriminator_loss',loss_disc.data[0] , step + num_steps_per_epoch * epoch)
log_value('generator_loss',loss_gen.data[0] , step + num_steps_per_epoch * epoch)
if (step ) % (num_steps_per_epoch/10) == 0 :
fake_store = fake_im.data.permute(0,2,1,3,4).contiguous().view(config.batch_size*16,3,64,64)
torchvision.utils.save_image(fake_store,
"{}fake_{}.png".format(config.sample_dir, cc), nrow=16,normalize=True)
real_store = real_im.data.permute(0,2,1,3,4).contiguous().view(config.batch_size*16,3,64,64)
torchvision.utils.save_image(real_store,
"{}real_{}.png".format(config.sample_dir, cc), nrow=16,normalize=True)
cc += 1
if epoch % 1 == 0:
torch.save(self.generator.state_dict(),
"{}/generator_{}.pth"
.format(config.model_dir,epoch))
torch.save(self.discriminator.state_dict(),
"{}/discriminator_{}.pth"
.format(config.model_dir, epoch))
def load(self, directory):
paths = glob.glob(os.path.join(directory, "*.pth"))
gen_path = [path for path in paths if "generator" in path][0]
disc_path = [path for path in paths if "discriminator" in path][0]
self.generator.load_state_dict(torch.load(gen_path))
self.discriminator.load_state_dict(torch.load(disc_path))
self.start_epoch = int(gen_path.split(".")[0].split("_")[-1])
print("Load pretrained [{}, {}]".format(gen_path, disc_path))
def _reset_gradients(self):
self.generator.zero_grad()
self.discriminator.zero_grad()