-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain_noisy_L.py
275 lines (255 loc) · 12.6 KB
/
train_noisy_L.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
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
#!/usr/bin/env python
# -*- coding: utf-8 -*-
# @Time : 2021-12-7 16:54
# @Author : 26731
# @File : train_tri.py
# @Software: PyCharm
import os
import argparse
import string
import random
import cv2
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 models import HN, DnCNN, IRCNN, DPAUNet, DPUNet, SUNet, FFDNet, HN, DnCNN_RL, \
HN2
from dataset import prepare_data, Dataset
from utils import *
import matplotlib.pyplot as plt
from torch.utils.tensorboard import SummaryWriter
# os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
# torch.cuda.set_device(0)
parser = argparse.ArgumentParser(description="SWCNN")
config = get_config('configs/config.yaml')
parser.add_argument("--preprocess", type=bool, default=False, help='run prepare_data or not')
parser.add_argument("--batchSize", type=int, default=8, help="Training batch size")
parser.add_argument("--num_of_layers", type=int, default=17, help="Number of total layers(DnCNN)")
parser.add_argument("--epochs", type=int, default=100, help="Number of training epochs")
parser.add_argument('--config', type=str, default='configs/config.yaml',
help="training configuration")
parser.add_argument("--milestone", type=int, default=30, help="When to decay learning rate; should be less than epochs")
parser.add_argument("--lr", type=float, default=1e-3, help="Initial learning rate")
parser.add_argument("--alpha", type=float, default=0.3, help="The opacity of the watermark")
parser.add_argument("--outf", type=str, default=config['train_model_out_path_DPAUNet'], help='path of log files')
parser.add_argument("--mode", type=str, default="S", help='with known noise level (S) or blind training (B)')
parser.add_argument("--mode_wm", type=str, default="S", help='with known alpha level (S) or blind training (B)')
parser.add_argument("--net", type=str, default="DPAUNet", help='Network used in training')
parser.add_argument("--loss", type=str, default="L1", help='The loss function used for training')
parser.add_argument("--self_supervised", type=str, default="True", help='T stands for TRUE and F stands for FALSE')
parser.add_argument("--PN", type=bool, default="True", help='Whether to use perception network')
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("--GPU_id", type=str, default="2", help='GPU_id')
opt = parser.parse_args()
os.environ["CUDA_VISIBLE_DEVICES"] = opt.GPU_id
if opt.self_supervised == "True":
model_name_3 = "n2n"
else:
model_name_3 = "n2c"
if opt.mode == "S":
model_name_4 = "S" + str(opt.noiseL)
else:
model_name_4 = "B"
if opt.loss == "L2":
model_name_5 = "L2"
else:
model_name_5 = "L1"
if opt.mode_wm == "S":
model_name_6 = "aS"
else:
model_name_6 = "aB"
tensorboard_name = opt.net + model_name_3 + model_name_4 + model_name_5 + model_name_6 + "alpha" + str(opt.alpha)
model_name = tensorboard_name + ".pth"
print()
def main():
# Load dataset
print('Loading dataset ...\n')
dataset_train = Dataset(train=True, mode='color', data_path=config['train_data_path'])
dataset_val = Dataset(train=False, mode='color', data_path=config['train_data_path'])
loader_train = DataLoader(dataset=dataset_train, num_workers=0, batch_size=opt.batchSize, shuffle=True) # 4
print("# of training samples: %d\n" % int(len(dataset_train)))
# load network
if opt.net == "HN":
net = HN()
elif opt.net == "SUNet":
net = SUNet()
elif opt.net == "DPUNet":
net = DPUNet()
elif opt.net == "DPAUNet":
net = DPAUNet()
elif opt.net == "HN2":
net = HN2()
elif opt.net == "DnCNN":
net = DnCNN(channels=3, num_of_layers=opt.num_of_layers)
elif opt.net == "DnCNN_RL":
net = DnCNN_RL(channels=3, num_of_layers=opt.num_of_layers)
elif opt.net == "FFDNet":
net = FFDNet(False)
elif opt.net == "IRCNN":
net = IRCNN(in_nc=3, out_nc=3)
else:
assert False
# TensorBoard was used to visually record the training results
writer = SummaryWriter("runs/" + tensorboard_name)
# 这个网络提取特征图，计算感知损失
model_vgg = load_froze_vgg16()
device_ids = [0]
model = nn.DataParallel(net, device_ids=device_ids).cuda()
# load loss function
if opt.loss == "L2":
criterion = nn.MSELoss(size_average=False)
else:
criterion = nn.L1Loss(size_average=False)
# criterion = nn.L1Loss(size_average=False)
# criterion_MSE = nn.MSELoss(size_average=False)
# Move to GPU
# Load the trained network and continue training
# model.load_state_dict(torch.load(os.path.join(opt.outf, "sec2_n2n_per_IRCNN_noise0.pth")))
criterion.cuda()
# Optimizer
optimizer = optim.Adam(model.parameters(), lr=opt.lr)
step = 0
noiseL_B = [0, 55] #
for epoch in range(opt.epochs):
if epoch < opt.milestone:
current_lr = opt.lr
else:
current_lr = opt.lr / 10.
# set learning rate
for param_group in optimizer.param_groups:
param_group["lr"] = current_lr
print('learning rate %f' % current_lr)
# train
for i, data in enumerate(loader_train, 0):
# training step
model.train()
model.zero_grad()
optimizer.zero_grad()
img_train = data
random_img = random.randint(1, 12)
if opt.mode_wm == "S":
imgn_train = add_watermark_noise(img_train, 40, True, random_img, alpha=opt.alpha)
else:
imgn_train = add_watermark_noise_B(img_train, 40, True, random_img, alpha=opt.alpha)
if opt.self_supervised == "True":
if opt.mode_wm == "S":
imgn_train_2 = add_watermark_noise(img_train, 40, True, random_img, alpha=opt.alpha)
else:
imgn_train_2 = add_watermark_noise_B(img_train, 40, True, random_img, alpha=opt.alpha)
else:
imgn_train_2 = img_train
imgn_train_mid = torch.Tensor(imgn_train)
if opt.mode == 'S':
if opt.noiseL != 0:
noise_gauss = torch.FloatTensor(img_train.size()).normal_(mean=0, std=opt.noiseL / 255.)
else:
noise_gauss = torch.zeros(img_train.size())
stdN = np.random.uniform(noiseL_B[0], noiseL_B[1], size=noise_gauss.size()[0])
for n in range(noise_gauss.size()[0]):
sizeN = noise_gauss[0, :, :, :].size()
noise_gauss[n, :, :, :] = torch.FloatTensor(sizeN).normal_(mean=0, std=stdN[n] / 255.)
if opt.noiseL == 0:
imgn_train = torch.Tensor(imgn_train)
else:
imgn_train = torch.Tensor(imgn_train) + noise_gauss
imgn_train_2 = torch.Tensor(imgn_train_2)
img_train, imgn_train = Variable(img_train.cuda()), Variable(imgn_train.cuda())
imgn_train_mid = Variable(imgn_train_mid.cuda())
imgn_train_2 = Variable(imgn_train_2.cuda())
if opt.net == "FFDNet":
noise_sigma = opt.noiseL / 255.
noise_sigma = torch.FloatTensor(np.array([noise_sigma for idx in range(img_train.shape[0])]))
noise_sigma = Variable(noise_sigma)
noise_sigma = noise_sigma.cuda()
out_train = model(imgn_train, noise_sigma)
elif opt.net == "SUNet":
out_dn, out_train = model(imgn_train)
elif opt.net == "DPUNet" or opt.net == "DPAUNet":
out_train, out_dn, out_wm = model(imgn_train)
feature_out_wm = model_vgg(out_wm)
else:
out_train = model(imgn_train)
feature_out = model_vgg(out_train)
feature_img = model_vgg(imgn_train_2)
if opt.net == "HN":
loss = (1.0 * criterion(out_train, imgn_train_2) / imgn_train.size()[
0] * 2) + (0.024 * criterion(feature_out, feature_img) / (feature_img.size()[0] / 2))
elif opt.net == "SUNet":
loss = (1.0 * criterion(out_train, img_train) / imgn_train.size()[
0] * 2) + (0.024 * criterion(feature_out, feature_img) / (feature_img.size()[0] / 2)) + (
1.0 * criterion(out_dn, imgn_train_mid) / imgn_train.size()[
0] * 2)
elif opt.net == "DPUNet" or opt.net == "DPAUNet":
loss = (1.0 * criterion(out_train, img_train) / imgn_train.size()[0] * 2) + (0.024 * criterion(feature_out, feature_img) / (feature_img.size()[0] / 2)) + (1.0 * criterion(out_dn, imgn_train_mid) / imgn_train.size()[0] * 2) + (
1.0 * criterion(out_wm, img_train) / imgn_train.size()[0] * 2) + (0.024 * criterion(feature_out_wm, feature_img) / (feature_img.size()[0] / 2))
else:
loss = (1.0 * criterion(out_train, img_train) / imgn_train.size()[
0] * 2) + (0.024 * criterion(feature_out, feature_img) / (feature_img.size()[0] / 2))
loss.backward()
optimizer.step()
# results
model.eval()
if opt.net == "FFDNet":
out_train = torch.clamp(model(imgn_train, noise_sigma), 0., 1.)
elif opt.net == "SUNet":
out_train = torch.clamp(model(imgn_train)[1], 0., 1.)
elif opt.net == "DPUNet" or opt.net == "DPAUNet":
out_train = torch.clamp(model(imgn_train)[0], 0., 1.)
else:
out_train = torch.clamp(model(imgn_train), 0., 1.)
psnr_train = batch_PSNR(out_train, img_train, 1.)
print("[epoch %d][%d/%d] loss: %.4f PSNR_train: %.4f" %
(epoch + 1, i + 1, len(loader_train), loss.item(), psnr_train))
step += 1
if step % 10 == 0:
writer.add_scalar("PSNR", psnr_train, step)
writer.add_scalar("loss", loss.item(), step)
## the end of each epoch
model.eval()
# Save the trained network parameters
torch.save(model.state_dict(), os.path.join(opt.outf, model_name))
# validate
psnr_val = 0
with torch.no_grad():
for k in range(len(dataset_val)):
img_val = torch.unsqueeze(dataset_val[k], 0)
# Cut the picture into multiples of 32
_, _, w, h = img_val.shape
w = int(int(w / 32) * 32)
h = int(int(h / 32) * 32)
img_val = img_val[:, :, 0:w, 0:h]
noise_gauss = torch.FloatTensor(img_val.size()).normal_(mean=0, std=opt.val_noiseL / 255.)
imgn_val = add_watermark_noise(img_val, 0, alpha=opt.alpha)
img_val = torch.Tensor(img_val)
if opt.noiseL == 0:
imgn_val = torch.Tensor(imgn_val)
else:
imgn_val = torch.Tensor(imgn_val) + noise_gauss
img_val, imgn_val = Variable(img_val.cuda(), volatile=True), Variable(imgn_val.cuda(), volatile=True)
if opt.net == "FFDNet":
noise_sigma = opt.noiseL / 255.
noise_sigma = torch.FloatTensor(np.array([noise_sigma for idx in range(img_val.shape[0])]))
noise_sigma = Variable(noise_sigma)
noise_sigma = noise_sigma.cuda()
out_val = torch.clamp(model(imgn_val, noise_sigma), 0., 1.)
elif opt.net == "SUNet":
out_val = torch.clamp(model(imgn_val)[1], 0., 1.)
elif opt.net == "DPUNet" or opt.net == "DPAUNet":
out_val = torch.clamp(model(imgn_val)[0], 0., 1.)
else:
out_val = torch.clamp(model(imgn_val), 0., 1.)
psnr_val += batch_PSNR(out_val, img_val, 1.)
psnr_val /= len(dataset_val)
writer.add_scalar("PSNR_val", psnr_val, epoch + 1)
print("\n[epoch %d] PSNR_val: %.4f" % (epoch + 1, psnr_val))
writer.close()
if __name__ == "__main__":
# data preprocess
if opt.preprocess:
prepare_data(data_path=config['train_data_path'], patch_size=256, stride=128, aug_times=1, mode='color')
main()