-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
257 lines (197 loc) · 10.1 KB
/
train.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
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
import os
import cv2
import time
import yaml
import random
import torch
import torchvision
import argparse
import numpy as np
import torch.nn as nn
import matplotlib as mpl
mpl.use('Agg')
import matplotlib.pyplot as plt
from shutil import copyfile
from data_util import custom_to_pil, save_image
from vqgan_util import load_config, load_vqgan, vqgan_encoder
from dataset import TrainDataset
from model.restoration_network import RAHC
from model.discriminitor import NetD
from omegaconf import OmegaConf
from taming.models.vqgan import VQModel, GumbelVQ
from loss import L1Loss, BCELoss, SSIMLoss, CharbonnierLoss, EdgeLoss, PSNRLoss, ContrastLoss, CosLoss
def setup_seed(seed):
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
np.random.seed(seed)
random.seed(seed)
torch.backends.cudnn.deterministic = True
# set random_seed
setup_seed(2023)
parser = argparse.ArgumentParser()
parser.add_argument('--train-dataset', type=str, default='./datasets/HAC/train', help='dataset folder for train (change for different datasets)')
parser.add_argument('--val-dataset', type=str, default='./datasets/HAC/test', help='dataset forder for val')
parser.add_argument('--save-epoch-freq', type=int, default=5 , help='how often to save model')
parser.add_argument('--print-freq', type=int, default=20, help='how often to print training information')
parser.add_argument('--val', type=bool, default=False, help='val during training or not')
parser.add_argument('--val-freq', type=int, default=5, help='how often to val model')
parser.add_argument('--train-visual-freq', type=int, default=20, help='how often to visualize training process')
parser.add_argument('--val-visual-freq', type=int, default=30, help='how often to visualize validation process')
parser.add_argument('--resume', type=str, default=None, help='continue training from this checkpoint')
parser.add_argument('--start-epoch', type=int, default=1, help='start epoch')
parser.add_argument('--output-dir', type=str, default='./checkpoints', help='model saved folder')
parser.add_argument('--log-dir', type=str, default='./logs', help='save visual image')
parser.add_argument('--epochs', type=int, default=300, help='total number of epoch')
parser.add_argument('--image-size', type=int, default=256, help='image crop size')
parser.add_argument('--batch_size', type=int, default=8, help='batch size')
parser.add_argument('--num-workers', type=int, default=16, help='num of workers per GPU to use')
parser.add_argument('--lr', type=float, default=0.0002, help='learning rate')
def build_model(opt):
device_ids = [i for i in range(torch.cuda.device_count())]
if torch.cuda.device_count() > 1:
print("\n\nLet's use", torch.cuda.device_count(), "GPUs!\n\n")
config_vqgan = load_config("logs/vqgan_gumbel_f8/configs/model.yaml", display=False)
vqgan, vqgan_dict = load_vqgan(config_vqgan, ckpt_path="logs/vqgan_gumbel_f8/checkpoints/last.ckpt", is_gumbel=True)
vqgan = vqgan.cuda()
discriminitor = NetD().cuda()
model = RAHC().cuda()
pretrained_dict = {k[9:]: v for k, v in vqgan_dict.items() if 'quantize' in k}
model.mapping_network.quantize.load_state_dict(pretrained_dict, strict=False)
for param in model.mapping_network.quantize.parameters():
param.requires_grad = False
param.retain_graph = True
optimizer_rest = torch.optim.Adam(list(model.encoders.parameters())+
list(model.decoders.parameters())+
list(model.intro.parameters())+
list(model.ending.parameters())+
list(model.middle_pre.parameters())+
list(model.middle_aft.parameters())+
list(model.ups.parameters())+
list(model.downs.parameters())+
list(model.post_conv.parameters()),
lr=opt.lr,
betas=(0.9, 0.999),eps=1e-8)
scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(optimizer_rest, opt.epochs,
eta_min=1e-6)
optimizer_map = torch.optim.Adam(list(model.mapping_network.conv.parameters())+
list(model.mapping_network.att.parameters()),
lr=opt.lr,
betas=(0.9, 0.999), eps=1e-8)
optimizer_dis = torch.optim.Adam(discriminitor.parameters(),
lr=opt.lr,
betas=(0.9, 0.999), eps=1e-8)
if len(device_ids) > 1:
model = nn.DataParallel(model, device_ids=device_ids)
l1loss = L1Loss().cuda()
bceloss = BCELoss().cuda()
cosloss = CosLoss().cuda()
loss = {'l1':l1loss, 'bce':bceloss, 'cos':cosloss}
return model, vqgan, discriminitor, optimizer_rest, optimizer_map, optimizer_dis, loss, scheduler
def get_trainval_loader(opt):
train_dataset = TrainDataset(opt)
train_dataloader = torch.utils.data.DataLoader(
train_dataset,
batch_size=opt.batch_size,
shuffle=True,
num_workers=opt.num_workers)
return train_dataloader
def load_checkpoint(opt, model, optimizer):
print(f"=> loading checkpoint '{opt.resume}'")
checkpoint = torch.load(opt.resume, map_location='cpu')
opt.start_epoch = checkpoint['epoch'] + 1
model.load_state_dict(checkpoint['model'])
optimizer.load_state_dict(checkpoint['optimizer'])
print(f"=> loaded successfully '{opt.resume}' (epoch {checkpoint['epoch']})")
del checkpoint
torch.cuda.empty_cache()
def save_checkpoint(epoch, model, optimizer):
print('==> Saving Epoch: {}'.format(epoch))
state = {
'model': model.state_dict(),
'optimizer': optimizer.state_dict(),
'epoch': epoch,
}
file_name = os.path.join(opt.output_dir, f'ckpt_epoch_{epoch}.pth')
torch.save(state, file_name)
copyfile(file_name, os.path.join(opt.output_dir, 'latest.pth'))
def train(epoch, model, vqgan, discriminitor, train_loader, optimizer_rest, optimizer_map, optimizer_dis, train_loss, opt):
epoch_start_time = time.time()
model.train()
total_epoch_train_loss = []
for batch_iter, data in enumerate(train_loader):
image, target, label = data['in_img'].cuda(), data['gt_img'].cuda(), data['label'].cuda()
output, f_rv = model(image)
# =====================discriminitor=============================
optimizer_dis.zero_grad()
dis_out_t = discriminitor(target)
type = torch.zeros_like(label).cuda()
dis_loss_t = train_loss['bce'](dis_out_t, type)
dis_loss_t.backward()
dis_out_d = discriminitor(output.detach())
dis_loss_d = train_loss['bce'](dis_out_d, label)
dis_loss_d.backward()
optimizer_dis.step()
# =====================mapping network=============================
vqgan_results = vqgan_encoder(target, vqgan).detach()
optimizer_map.zero_grad()
map_net_loss = train_loss['cos'](f_rv, vqgan_results)
map_net_loss.backward()
optimizer_map.step()
# =====================restoration network=============================
optimizer_rest.zero_grad()
dis_out_o = discriminitor(output)
losses = train_loss['l1'](output, target) + 0.1 * train_loss['bce'](dis_out_o, type)
losses.backward()
optimizer_rest.step()
total_epoch_train_loss.append(losses.cpu().data)
if (batch_iter+1) % opt.print_freq == 0:
print('Epoch: {}, Epoch_iter: {}, Loss: {}'.format(epoch, batch_iter+1, losses))
if (batch_iter+1) % opt.train_visual_freq == 0:
print('Saving training image epoch: {}'.format(epoch))
save_image(epoch, 'train',[image, output, target] , opt)
print('End of epoch %d / %d \t Time Taken: %d sec' %
(epoch, opt.epochs, time.time() - epoch_start_time))
return np.mean(total_epoch_train_loss) / opt.batch_size
def val(epoch, model, val_loader, val_loss, opt):
model.eval()
total_epoch_val_loss = []
for i, data in enumerate(val_loader):
image, target = data['in_img'].cuda(), data['gt_img'].cuda()
output, _ = model(image)
losses = val_loss['l1'](output, target)
total_epoch_val_loss.append(losses.cpu().data)
if (i+1) % opt.val_visual_freq == 0:
print('Saving validation image epoch: {}'.format(epoch))
save_image(epoch, 'val', [image, final, target], opt)
return np.mean(total_epoch_val_loss) / opt.batch_size
def main(opt):
model, vqgan, discriminitor, optimizer_rest, optimizer_map, optimizer_dis, loss, scheduler = build_model(opt)
scheduler.step()
if opt.val:
train_loader, val_loader = get_trainval_loader(opt)
else:
train_loader = get_trainval_loader(opt)
total_train_loss = []
total_val_loss = []
if opt.resume:
assert os.path.isfile(opt.resume)
load_checkpoint(opt, model, optimizer_rest)
for epoch in range(opt.start_epoch, opt.epochs + 1):
epoch_train_loss = train(epoch, model, vqgan, discriminitor, train_loader, optimizer_rest, optimizer_map, optimizer_dis, loss, opt)
total_train_loss.append(epoch_train_loss)
if epoch % opt.save_epoch_freq == 0:
print('saving the model at the end of epoch %d' % epoch)
save_checkpoint(epoch, model, optimizer_rest)
if opt.val and epoch % opt.val_freq == 0:
epoch_val_loss = val(epoch, model, val_loader, loss, opt)
print('Epoch: {}, Validation_Loss: {}'.format(epoch, epoch_val_loss))
total_val_loss.append(epoch_val_loss)
scheduler.step()
plt.subplot(211)
plt.plot(total_train_loss)
plt.subplot(212)
plt.plot(total_val_loss)
plt.savefig(opt.log_dir + '/loss.png')
if __name__ == '__main__':
opt = parser.parse_args()
main(opt)