forked from VITA-Group/USAID
-
Notifications
You must be signed in to change notification settings - Fork 0
/
USAID_train.py
192 lines (159 loc) · 8.47 KB
/
USAID_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
from utils import *
import os
import argparse
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
import torchvision.utils as utils
from torch.autograd import Variable
from torch.utils.data import DataLoader
from tensorboardX import SummaryWriter
from models import DnCNN
from dataset import prepare_data, Dataset
from USAID_dataloader import *
from FPN.models.fpn import fpn
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
os.environ["CUDA_VISIBLE_DEVICES"] = "3"
parser = argparse.ArgumentParser(description="DnCNN")
parser.add_argument("--batchSize", type=int, default=64, help="Training batch size")
parser.add_argument("--cropSize", type=int, default=48, help="Image crop size")
parser.add_argument("--num_of_layers", type=int, default=20, help="Number of total layers")
parser.add_argument("--epochs", type=int, default=100, help="Number of training epochs")
parser.add_argument("--lr", type=float, default=1e-3, help="Initial learning rate")
parser.add_argument("--coef_MSE", type=float, default=5e-1, help="Coefficient for MSE in total loss, (1-coef_MSE) for Seg Loss")
parser.add_argument("--outf", type=str, default="logs", help='path of log files')
parser.add_argument("--mode", type=str, default="B", help='with known noise level (S) or blind training (B)')
parser.add_argument("--noiseL", type=float, default=25, help='noise level; ignored when mode=B')
parser.add_argument("--val_noiseL", type=float, default=25, help='noise level used on validation set')
parser.add_argument("--num_of_SegClass", type=int, default=21, help='Number of Segmentation Classes, default VOC = 21')
opt = parser.parse_args()
save_dir = opt.outf
print ('save models to directory : ', save_dir)
if not os.path.exists(save_dir):
os.mkdirs(save_dir)
def main():
# Load dataset
print('Loading dataset ...\n')
# VOC dataset loading
dataset_train = MultiDataSet(cropSize=opt.cropSize, testFlag=False, Scale=False)
dataset_val = MultiDataSet(cropSize=opt.cropSize, testFlag=True, Scale=False)
loader_train = DataLoader(dataset=dataset_train, num_workers=4, batch_size=opt.batchSize, shuffle=True)
loader_val = DataLoader(dataset=dataset_val, num_workers=1, batch_size=1, shuffle=False)
print("# of training samples: %d\n" % int(len(dataset_train)))
print("# of validation samples: %d\n" % int(len(dataset_val)))
# Denoiser
net = DnCNN(channels=3, num_of_layers=opt.num_of_layers)
net.apply(weights_init_kaiming)
criterion = nn.MSELoss(size_average=False).cuda()
model = nn.DataParallel(net).cuda()
seg = fpn(opt.num_of_SegClass)
seg_criterion = FocalLoss(gamma=2).cuda()
seg = nn.DataParallel(seg).cuda()
optimizer = optim.Adam(model.parameters(), lr=opt.lr)
scheduler = optim.lr_scheduler.MultiStepLR(optimizer, milestones=[10,40, 80, 120, 140], gamma=0.1)
# training
writer = SummaryWriter(save_dir)
step = 0
noiseL_B=[0,55] # ingnored when opt.mode=='S'
for epoch in range(opt.epochs):
scheduler.step()
for param_group in optimizer.param_groups:
current_lr = param_group["lr"]
print('learning rate %f' % current_lr)
# train
for i, data in enumerate(loader_train, 0):
img_train = data
model.train()
seg.train()
model.zero_grad()
seg.zero_grad()
optimizer.zero_grad()
# training step
if opt.mode == 'S':
noise = torch.FloatTensor(img_train.size()).normal_(mean=0, std=opt.noiseL/255.)
if opt.mode == 'B':
noise = torch.zeros(img_train.size())
stdN = np.random.uniform(noiseL_B[0], noiseL_B[1], size=noise.size()[0])
for n in range(noise.size()[0]):
sizeN = noise[0,:,:,:].size()
noise[n,:,:,:] = torch.FloatTensor(sizeN).normal_(mean=0, std=stdN[n]/255.)
imgn_train = img_train + noise
img_train, imgn_train = Variable(img_train.cuda()), Variable(imgn_train.cuda())
noise = Variable(noise.cuda())
out_train = model(imgn_train)
loss = criterion(out_train, noise) / (imgn_train.size()[0]*2)
out_train = torch.clamp(imgn_train-model(imgn_train), 0., 1.)
psnr_train = batch_PSNR(out_train, img_train, 1.)
# demean segmentation inputs
seg_input = out_train.data.cpu().numpy()
for n in range(out_train.size()[0]):
seg_input[n, :, :, :] = rgb_demean(seg_input[n, :, :, :])
seg_input = Variable(torch.from_numpy(seg_input).cuda())
seg_output = seg(seg_input)
target = (get_NoGT_target(seg_output)).data.cpu()
target_ = resize_target(target, seg_output.size(2))
target_ = torch.from_numpy(target_).long()
target_ = target_.cuda()
seg_loss = seg_criterion(seg_output, target_)
for param in seg.parameters():
param.requires_grad = False
totalLoss = opt.coef_MSE * loss + (1 - opt.coef_MSE) * seg_loss
totalLoss.backward()
optimizer.step()
if (i+1) % 1000 == 0:
print("[epoch %d][%d/%d] [SegClass: %d] loss: %.4f PSNR_train: %.4f" %
(epoch+1, i+1, len(loader_train), opt.num_of_SegClass, loss.item(), psnr_train))
# if you are using older version of PyTorch, you may need to change loss.item() to loss.data[0]
if step % 10 == 0:
# Log the scalar values
writer.add_scalar('loss', loss.item(), step)
writer.add_scalar('PSNR on training data', psnr_train, step)
step += 1
## the end of every 20 epoch， do validation
if (epoch+1) % 20 == 0:
model.eval()
psnr_val = 0
niqe_val = 0
ssim_val = 0
with torch.no_grad():
for i, data in enumerate(loader_val, 0):
img_val = data
noise = torch.FloatTensor(img_val.size()).normal_(mean=0, std=opt.noiseL / 255.)
imgn_val = img_val + noise
img_val, imgn_val = Variable(img_val.cuda()), Variable(imgn_val.cuda())
out_val = torch.clamp(imgn_val - model(imgn_val), 0., 1.)
psnr_val += batch_PSNR(out_val, img_val, 1.)
ssim_val += batch_SSIM(out_val, img_val, 1.)
if epoch == opt.epochs - 1:
niqe_val += batch_NIQE(out_val)
psnr_val /= len(loader_val)
ssim_val /= len(loader_val)
writer.add_scalar('PSNR on validation data', psnr_val, epoch)
torch.save(model.state_dict(), os.path.join(save_dir,
str(opt.num_of_SegClass)
+ '_USAID_epoch'
+ str(epoch+1) + '_'
+ str(psnr_val) + '.pth'))
print("\n[epoch %d] [SegClass: %d] PSNR_val: %.2f SSIM_val: %.4f"
% (epoch + 1, opt.num_of_SegClass, psnr_val, ssim_val))
print("**********************************************************************")
if epoch == opt.epochs - 1:
niqe_val /= len(loader_val)
torch.save(model.state_dict(),
os.path.join(save_dir, str(opt.num_of_SegClass) + '_USAID_final.pth'))
print("\n[epoch %d] [SegClass: %d] PSNR_val: %.2f SSIM_val: %.4f NIQE_val: %.4f"
% (epoch + 1, opt.num_of_SegClass, psnr_val, ssim_val, niqe_val))
print("\n========== END ===========")
# log the images
out_train = torch.clamp(imgn_train-model(imgn_train), 0., 1.)
Img = utils.make_grid(img_train.data, nrow=8, normalize=True, scale_each=True)
Imgn = utils.make_grid(imgn_train.data, nrow=8, normalize=True, scale_each=True)
Irecon = utils.make_grid(out_train.data, nrow=8, normalize=True, scale_each=True)
writer.add_image('clean image', Img, epoch)
writer.add_image('noisy image', Imgn, epoch)
writer.add_image('reconstructed image', Irecon, epoch)
# save model
torch.save(model.state_dict(), os.path.join(save_dir, str(opt.num_of_SegClass) + '_USAID_lastest.pth'))
if __name__ == "__main__":
main()