-
Notifications
You must be signed in to change notification settings - Fork 5
/
Copy pathtrain_Deblur.py
190 lines (156 loc) · 6.82 KB
/
train_Deblur.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
import os
import sys
# add dir
dir_name = os.path.dirname(os.path.abspath(__file__))
sys.path.append(os.path.join(dir_name, './auxiliary/'))
print(dir_name)
import argparse
import options
######### parser ###########
opt = options.Options().init(argparse.ArgumentParser(description='image debluring')).parse_args()
print(opt)
import utils
######### Set GPUs ###########
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = opt.gpu
import torch
torch.backends.cudnn.benchmark = True
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
import random
import time
import numpy as np
from einops import rearrange, repeat
import datetime
from losses import CharbonnierLoss
from warmup_scheduler import CosineAnnealingWithRestartsLR, CosineAnnealingRestartCyclicLR
from timm.utils import NativeScaler
from utils.loader import get_deblur_training_data
def print_network(net):
num_params = 0
for param in net.parameters():
num_params += param.numel()
print('Total number of parameters: %d' % num_params)
######### Logs dir ###########
log_dir = os.path.join(dir_name, 'log', opt.arch + opt.env)
if not os.path.exists(log_dir):
os.makedirs(log_dir)
logname = os.path.join(log_dir, datetime.datetime.now().isoformat() + '.txt')
print("Now time is : ", datetime.datetime.now().isoformat())
result_dir = os.path.join(log_dir, 'results')
model_dir = opt.save_dir
utils.mkdir(result_dir)
utils.mkdir(model_dir)
# ######### Set Seeds ###########
random.seed(1234)
np.random.seed(1234)
torch.manual_seed(1234)
torch.cuda.manual_seed_all(1234)
######### Model ###########
model_restoration = utils.get_arch(opt)
print_network(model_restoration)
with open(logname, 'a') as f:
f.write(str(opt) + '\n')
f.write(str(model_restoration) + '\n')
######### Optimizer ###########
start_epoch = 1
if opt.optimizer.lower() == 'adam':
optimizer = optim.Adam(model_restoration.parameters(), lr=opt.lr_initial, betas=(0.9, 0.999), eps=1e-8,
weight_decay=opt.weight_decay)
elif opt.optimizer.lower() == 'adamw':
optimizer = optim.AdamW(model_restoration.parameters(), lr=opt.lr_initial, betas=(0.9, 0.999), eps=1e-8,
weight_decay=opt.weight_decay)
else:
raise Exception("Error optimizer...")
scheduler = CosineAnnealingRestartCyclicLR(optimizer, periods=[opt.thre, opt.nepoch-opt.thre], restart_weights=[1, 1] ,eta_mins=[opt.lr_initial, opt.LR_MIN])
######### DataParallel ###########
model_restoration = torch.nn.DataParallel(model_restoration)
model_restoration.cuda()
######### Loss ###########
criterion = CharbonnierLoss().cuda()
######### DataLoader ###########
print('===> Loading datasets')
train_dataset = get_deblur_training_data(opt.train_dir, opt.train_ps)
train_loader = DataLoader(dataset=train_dataset, batch_size=opt.batch_size, shuffle=True,
num_workers=opt.train_workers, pin_memory=True, drop_last=False)
val_dataset = get_deblur_training_data(opt.val_dir, opt.train_ps)
val_loader = DataLoader(dataset=val_dataset, batch_size=opt.batch_size, shuffle=False,
num_workers=opt.eval_workers, pin_memory=True, drop_last=False)
len_trainset = train_dataset.__len__()
len_valset = val_dataset.__len__()
print("Sizeof training set: ", len_trainset, ", sizeof validation set: ", len_valset)
######### train ###########
print('===> Start Epoch {} End Epoch {}'.format(start_epoch, opt.nepoch))
best_psnr = 0
best_epoch = 0
best_iter = 0
loss_scaler = NativeScaler()
torch.cuda.empty_cache()
global_step = 0
eval_now = len(train_loader)
print("Eval Freq: ", eval_now)
for epoch in range(start_epoch, opt.nepoch + 1):
epoch_start_time = time.time()
epoch_loss = 0
model_restoration.train()
for i, data in enumerate(train_loader, 0):
global_step += 1
model_restoration.zero_grad()
optimizer.zero_grad()
target = data[0].cuda()
input_ = data[1].cuda()
restored = model_restoration(input_)
restored = torch.clamp(restored, 0, 1)
loss = criterion(restored, target)
loss.backward()
if opt.use_grad_clip:
torch.nn.utils.clip_grad_norm_(model_restoration.parameters(), 0.01)
optimizer.step()
epoch_loss += loss.item()
if global_step % 50 == 0:
print("Epoch: %d, LearningRate: %.6f, global step: %d, loss: %.4f, time:%.4f" %(epoch, scheduler.get_lr()[0],
global_step, epoch_loss, time.time() - epoch_start_time))
epoch_loss = 0.0
epoch_start_time = time.time()
#### Evaluation ####
if (i + 1) % eval_now == 0 and i > 0:
with torch.no_grad():
model_restoration.eval()
psnr_val_rgb = []
for ii, data_val in enumerate((val_loader), 0):
target = data_val[0].cuda()
input_ = data_val[1].cuda()
with torch.cuda.amp.autocast():
restored = model_restoration(input_)
restored = torch.clamp(restored, 0, 1)
psnr_val_rgb.append(utils.batch_PSNR(restored, target, False).item())
psnr_val_rgb = sum(psnr_val_rgb) / len_valset
if psnr_val_rgb > best_psnr:
best_psnr = psnr_val_rgb
best_epoch = epoch
best_iter = i
torch.save({'epoch': epoch,
'state_dict': model_restoration.state_dict(),
'optimizer': optimizer.state_dict()
}, os.path.join(model_dir, "model_best.pth"))
print(
"[Ep %d it %d\t PSNR SIDD: %.4f\t] ---- [best_Ep_SIDD %d best_it_SIDD %d Best_PSNR_SIDD %.4f] " % (
epoch, i, psnr_val_rgb, best_epoch, best_iter, best_psnr))
with open(logname, 'a') as f:
f.write(
"[Ep %d it %d\t PSNR SIDD: %.4f\t] ---- [best_Ep_SIDD %d best_it_SIDD %d Best_PSNR_SIDD %.4f] " \
% (epoch, i, psnr_val_rgb, best_epoch, best_iter, best_psnr) + '\n')
model_restoration.train()
torch.cuda.empty_cache()
scheduler.step()
torch.save({'epoch': epoch,
'state_dict': model_restoration.state_dict(),
'optimizer': optimizer.state_dict()
}, os.path.join(model_dir, "model_latest.pth"))
if epoch % opt.checkpoint == 0:
torch.save({'epoch': epoch,
'state_dict': model_restoration.state_dict(),
'optimizer': optimizer.state_dict()
}, os.path.join(model_dir, "model_epoch_{}.pth".format(epoch)))
print("Now time is : ", datetime.datetime.now().isoformat())