-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathrunner.py
176 lines (148 loc) · 6.35 KB
/
runner.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
import os
import ipdb
import matplotlib.pyplot as plt
import torch
import torch.utils.data as data
from loader import Dataset
from model import CMC_G, CMC_D, CMCLoss
def save_ckpt(state, filename):
filename = f'{filename}.pth'
torch.save(state, filename)
def load_ckpt(model, optimizer, filename, map_location, logger=None):
if os.path.isfile(filename):
logger.info(f'>>> Load ckpt from {filename}')
checkpoint = torch.load(filename, map_location)
epoch = checkpoint.get('epoch', -1)
if model is not None and checkpoint['model_state'] is not None:
model.load_state_dict(checkpoint['model_state'])
if optimizer is not None and checkpoint['optimizer_state'] is not None:
optimizer.load_state_dict(checkpoint['optimizer_state'])
logger.info('>>> Load done !')
else:
raise FileNotFoundError
return epoch
def model_state_to_cpu(model_state):
model_state_cpu = type(model_state)()
for key, val in model_state.items():
model_state_cpu[key] = val.cpu()
return model_state_cpu
def get_ckpt_state(G_model=None, D_model=None, G_optimizer=None, D_optimizer=None, epoch=None):
G_optim_state = G_optimizer.state_dict() if G_optimizer is not None else None
D_optim_state = D_optimizer.state_dict() if D_optimizer is not None else None
if G_model is not None:
if isinstance(G_model, torch.nn.DataParallel):
G_model_state = model_state_to_cpu(G_model.module.state_dict())
else:
G_model_state = G_model.state_dict()
else:
G_model_state = None
if D_model is not None:
if isinstance(D_model, torch.nn.DataParallel):
D_model_state = model_state_to_cpu(D_model.module.state_dict())
else:
D_model_state = D_model.state_dict()
else:
D_model_state = None
ckpt_state = {
'epoch': epoch,
'G_model_state': G_model_state,
'D_model_state': D_model_state,
'G_optim_state': G_optim_state,
'D_optim_state': D_optim_state,
}
return ckpt_state
def train(cfg, logger):
device = torch.device('cuda')
logger.info('>>> Build Modeling ...')
cmc_g = CMC_G()
cmc_d = CMC_D()
cmc_g.to(device)
cmc_d.to(device)
compute_losses = CMCLoss(cfg)
dataset = Dataset(cfg, is_train=True)
dataloader = data.DataLoader(
dataset=dataset,
num_workers=8,
batch_size=cfg.batch_size,
shuffle=True,
pin_memory=True,
drop_last=True,
)
G_optimizer = torch.optim.Adam(cmc_g.parameters(), lr=cfg.lr, weight_decay=cfg.weight_decay, betas=(0.9, 0.99))
D_optimizer = torch.optim.Adam(cmc_d.parameters(), lr=cfg.lr, weight_decay=cfg.weight_decay, betas=(0.9, 0.99))
total_iters = 0
for epoch in range(cfg.epochs):
epoch_iter = 0
dataset.shuffle()
disp_dict = {}
for idx, batch_data in enumerate(dataloader):
total_iters += cfg.batch_size
epoch_iter += cfg.batch_size
image, audio, unpair_image, unpair_audio, label, noise = batch_data
image = image.to(device)
audio = audio.to(device)
unpair_image = unpair_image.to(device)
unpair_audio = unpair_audio.to(device)
label = label.to(device)
noise = noise.to(device)
data_dict = {
'real_image': image,
'real_audio': audio,
'fake_image': unpair_image,
'fake_audio': unpair_audio,
'label': label,
'noise': noise,
}
# Forward G
G_stat_dict = cmc_g(image, audio, label, noise)
# Backward D
# compute_losses.set_requires_grad(model.D_CMC, True)
D_optimizer.zero_grad()
D_stat_dict = cmc_d(image, audio, unpair_image, unpair_audio, G_stat_dict, detach=True)
D_loss_dict = compute_losses.optim_D(D_stat_dict)
D_loss_sums = sum(loss for loss in D_loss_dict.values())
D_loss_sums.backward()
D_optimizer.step()
# Backward G
# compute_losses.set_requires_grad(model.D_CMC, False)
G_optimizer.zero_grad()
D_stat_dict = cmc_d(image, audio, unpair_image, unpair_audio, G_stat_dict, detach=False)
G_loss_dict = compute_losses.optim_G(G_stat_dict, D_stat_dict, data_dict)
G_loss_sums = sum(loss for loss in G_loss_dict.values())
G_loss_sums.backward()
G_optimizer.step()
# # update D
# compute_losses.set_requires_grad(model.D_CMC, True)
# D_optimizer.zero_grad()
# D_loss_dict = compute_losses.optim_D(stat_dict)
# D_loss_sums = sum(loss for loss in D_loss_dict.values())
# D_loss_sums.backward(retain_graph=True)
# D_optimizer.step()
# # repeat G
# # for i in range(cfg.repeat_nums):
# compute_losses.set_requires_grad(model.D_CMC, False)
# G_optimizer.zero_grad()
# G_loss_dict = compute_losses.optim_G(stat_dict)
# G_loss_sums = sum(loss for loss in G_loss_dict.values())
# G_loss_sums.backward()
# G_optimizer.step()
for key in D_loss_dict.keys():
if key not in disp_dict.keys():
disp_dict[key] = 0
disp_dict[key] += D_loss_dict[key]
for key in G_loss_dict.keys():
if key not in disp_dict.keys():
disp_dict[key] = 0
disp_dict[key] += G_loss_dict[key]
log_str = 'EPOCH[%03d/%03d] ITER[%05d]' % (epoch, cfg.epochs, epoch_iter)
if total_iters % cfg.disp_freq == 0:
for key in sorted(disp_dict.keys()):
disp_dict[key] = disp_dict[key] / cfg.disp_freq
log_str += ' %s:%.4f ' % (key, disp_dict[key])
disp_dict[key] = 0 # reset disp_dict
logger.info(log_str)
if epoch % cfg.save_freq == 0:
os.makedirs(os.path.join(cfg.save_path, 'ckpt'), exist_ok=True)
ckpt_name = os.path.join(cfg.save_path, 'ckpt', 'ckpt_epoch_%d' % epoch)
save_ckpt(get_ckpt_state(cmc_g, cmc_d, G_optimizer, D_optimizer, epoch=epoch), ckpt_name)
logger.info(f'>>> Save checkpoint: {ckpt_name}')