Restoring-Extremely-Dark-Images-In-Real-Time/train.py at main · scott-mao/Restoring-Extremely-Dark-Images-In-Real-Time · GitHub

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
opt={'base_lr':1e-4}
opt['reduce_lr_by'] = 0.1
opt['atWhichReduce'] = [300000]
opt['batch_size'] = #Enter batch size
opt['atWhichSave'] = [2,100002,150002,200002,250002,300002,350002,400002,450002,500002, 600000,700000,800000,900000,1000000]
opt['iterations'] = 1000005
dry_run_iterations = 100
dry_run = False
metric_average_file = 'metric_average.txt'
test_amplification_file = 'test_amplification.txt'
train_amplification_file = 'train_amplification.txt'
# These are folders
save_weights = 'weights'
save_images = 'images' # make this change below also
save_csv_files = 'csv_files'

import numpy as np
import os
import shutil
import torch
from torch.utils.data import DataLoader
import torch.optim as optim
import glob

from common_classes import load_data, run_test
from network import Net
import argparse

# Argument for EDSR
parser = argparse.ArgumentParser(description='EDSR')
parser.add_argument('--n_resblocks', type=int, default=32,
                    help='number of residual blocks')
parser.add_argument('--n_feats', type=int, default=64,
                    help='number of feature maps')
parser.add_argument('--res_scale', type=float, default=1,
                    help='residual scaling')
parser.add_argument('--scale', type=str, default=2,
                    help='super resolution scale')
parser.add_argument('--patch_size', type=int, default=256,
                    help='output patch size')
parser.add_argument('--n_colors', type=int, default=4,
                    help='number of input color channels to use')
parser.add_argument('--o_colors', type=int, default=3,
                    help='number of output color channels to use')
args = parser.parse_args()

os.environ["CUDA_DEVICE_ORDER"]="PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"]="1"

shutil.rmtree(metric_average_file, ignore_errors = True)
shutil.rmtree(test_amplification_file, ignore_errors = True)
shutil.rmtree(train_amplification_file, ignore_errors = True)

shutil.rmtree(save_weights, ignore_errors = True)
shutil.rmtree(save_images, ignore_errors = True)
shutil.rmtree(save_csv_files, ignore_errors = True)

os.makedirs(save_weights)
os.makedirs(save_images)
os.makedirs(save_csv_files)

train_files = glob.glob('/SID_cvpr_18_dataset/Sony/short/0*_00_0.1s.ARW')
train_files +=glob.glob('/SID_cvpr_18_dataset/Sony/short/2*_00_0.1s.ARW')
if dry_run:
    train_files = train_files[:2]
    opt['iterations'] = dry_run_iterations

gt_files = []
for x in train_files:
    gt_files += glob.glob('/SID_cvpr_18_dataset/Sony/long/*'+x[-17:-12]+'*.ARW')

dataloader_train = DataLoader(load_data(train_files,gt_files,train_amplification_file,20,gt_amp=True,training=True), batch_size=1, shuffle=True, num_workers=0, pin_memory=True)

test_files = glob.glob('/SID_cvpr_18_dataset/Sony/short/1*_00_0.1s.ARW')
if dry_run:
    test_files = test_files[:2]

gt_files = []
for x in test_files:
    gt_files = gt_files+ glob.glob('/SID_cvpr_18_dataset/Sony/long/*'+x[-17:-12]+'*.ARW')
dataloader_test = DataLoader(load_data(test_files,gt_files,test_amplification_file,2,gt_amp=True,training=False), batch_size=1, shuffle=False, num_workers=0, pin_memory=True)

for i,img in enumerate(dataloader_train):
    print('Input image size : {}, GT image size : {}'.format(img[0].size(), img[1].size()))
    break

############ Training Begins

device = torch.device("cuda")
model = Net(args)
print(model)
print('\nTrainable parameters : {}\n'.format(sum(p.numel() for p in model.parameters() if p.requires_grad)))
model = model.to(device)
print('Device on cuda: {}'.format(next(model.parameters()).is_cuda))

iter_num = 0
l1_loss = torch.nn.L1Loss()

optimizer = torch.optim.Adam(model.parameters(), lr=opt['base_lr'])
optimizer.zero_grad()
loss_list = ['L1_loss']
loss_iter_list = ['Iteration']
iter_LR = ['Iter_LR']


while iter_num<opt['iterations']:
    for _, img in enumerate(dataloader_train):
        low = img[0].to(device)
        gt = img[1].to(device)
        model.train()
        pred = model(low)
        iter_num +=1
        #loss2 = feature_loss(pred, gt, which='relu2')
        loss1 = l1_loss(pred,gt)
        #print('l1,vgg loss: ',loss1.item(), loss2.item())
        loss = loss1
        loss.backward()
        optimizer.step()
        optimizer.zero_grad()

        if iter_num>opt['iterations']:
            break

        if iter_num%10==0:
            print(iter_num)
            if iter_num%100==0:
                loss_list.append('{},{},{}'.format(loss1.item(),-1,-1))
                loss_iter_list.append(iter_num)
                iter_LR.append(optimizer.param_groups[0]['lr'])

        if iter_num in opt['atWhichReduce']:
            for group in optimizer.param_groups:
                old_lr = group['lr']
                group['lr'] = opt['reduce_lr_by']*old_lr
                if group['lr']<1e-5:
                    group['lr']=1e-5
                print('Changing LR from {} to {}'.format(old_lr,group['lr']))

        if iter_num in opt['atWhichSave']:
            print('testing......')
            if iter_num == opt['atWhichSave'][0]:
                mode = 'w'
            else:
                mode = 'a'
            run_test(model, dataloader_test, iter_num, save_images, save_csv_files, metric_average_file, mode, training=True)
            torch.save({'model': model.state_dict()},os.path.join(save_weights,'weights_{}'.format(iter_num)))


np.savetxt(os.path.join(save_csv_files,'loss_curve.csv'),[p for p in zip(loss_iter_list,loss_list,iter_LR)],delimiter=',',fmt='%s')