forked from IDKiro/CBDNet-pytorch
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_all.py
272 lines (214 loc) · 11.4 KB
/
train_all.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
from __future__ import division
from __future__ import print_function
import os, time, scipy.io, shutil
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
import glob
import re
import cv2
from utils import *
from model import *
def load_CRF():
CRF = scipy.io.loadmat('matdata/201_CRF_data.mat')
iCRF = scipy.io.loadmat('matdata/dorfCurvesInv.mat')
B_gl = CRF['B']
I_gl = CRF['I']
B_inv_gl = iCRF['invB']
I_inv_gl = iCRF['invI']
if os.path.exists('matdata/201_CRF_iCRF_function.mat')==0:
CRF_para = np.array(CRF_function_transfer(I_gl, B_gl))
iCRF_para = 1. / CRF_para
scipy.io.savemat('matdata/201_CRF_iCRF_function.mat', {'CRF':CRF_para, 'iCRF':iCRF_para})
else:
Bundle = scipy.io.loadmat('matdata/201_CRF_iCRF_function.mat')
CRF_para = Bundle['CRF']
iCRF_para = Bundle['iCRF']
return CRF_para, iCRF_para, I_gl, B_gl, I_inv_gl, B_inv_gl
def load_checkpoint(checkpoint_dir):
if os.path.exists(checkpoint_dir + 'checkpoint.pth.tar'):
# load existing model
model_info = torch.load(checkpoint_dir + 'checkpoint.pth.tar')
print('==> loading existing model:', checkpoint_dir + 'checkpoint.pth.tar')
model = CBDNet()
model.cuda()
model.load_state_dict(model_info['state_dict'])
optimizer = torch.optim.Adam(model.parameters())
optimizer.load_state_dict(model_info['optimizer'])
cur_epoch = model_info['epoch']
else:
# create model
model = CBDNet()
model.cuda()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
cur_epoch = 0
return model, optimizer, cur_epoch
def save_checkpoint(state, is_best, filename='checkpoint.pth.tar'):
torch.save(state, checkpoint_dir + 'checkpoint.pth.tar')
if is_best:
shutil.copyfile(checkpoint_dir + 'checkpoint.pth.tar',checkpoint_dir + 'model_best.pth.tar')
def adjust_learning_rate(optimizer, epoch, lr_update_freq):
if not epoch % lr_update_freq and epoch:
for param_group in optimizer.param_groups:
param_group['lr'] = param_group['lr'] * 0.1
return optimizer
if __name__ == '__main__':
input_syn_dir = './dataset/synthetic/'
input_real_dir = './dataset/real/'
checkpoint_dir = './checkpoint/all/'
result_dir = './result/all/'
PS = 512
REAPET = 10
save_freq = 50
lr_update_freq = 100
CRF_para, iCRF_para, I_gl, B_gl, I_inv_gl, B_inv_gl = load_CRF()
train_syn_fns = glob.glob(input_syn_dir + '*.bmp')
train_real_fns = glob.glob(input_real_dir + 'Batch_*')
origin_syn_imgs = [None] * len(train_syn_fns)
noise_syn_imgs = [None] * len(train_syn_fns)
noise_syn_levels = [None] * len(train_syn_fns)
origin_real_imgs = [None] * len(train_real_fns)
noise_real_imgs = [None] * len(train_real_fns)
for i in range(len(train_syn_fns)):
origin_syn_imgs[i] = []
noise_syn_imgs[i] = []
noise_syn_levels[i] = []
for i in range(len(train_real_fns)):
origin_real_imgs[i] = []
noise_real_imgs[i] = []
model, optimizer, cur_epoch = load_checkpoint(checkpoint_dir)
criterion = fixed_loss()
criterion = criterion.cuda()
for epoch in range(cur_epoch, 201):
cnt=0
losses = AverageMeter()
optimizer = adjust_learning_rate(optimizer, epoch, lr_update_freq)
model.train()
print('Training on synthetic noisy images...')
for ind in np.random.permutation(len(train_syn_fns)):
train_syn_fn = train_syn_fns[ind]
if not len(origin_syn_imgs[ind]):
origin_syn_img = cv2.imread(train_syn_fn)
origin_syn_img = origin_syn_img[:,:,::-1] / 255.0
origin_syn_imgs[ind] = np.array(origin_syn_img).astype('float32')
# re-add noise
if epoch % save_freq == 0:
noise_syn_imgs[ind] = []
noise_syn_levels[ind] = []
if len(noise_syn_imgs[ind]) < 1:
noise_syn_img, noise_syn_level = AddRealNoise(origin_syn_imgs[ind][:, :, :], CRF_para, iCRF_para, I_gl, B_gl, I_inv_gl, B_inv_gl)
noise_syn_imgs[ind].append(noise_syn_img)
noise_syn_levels[ind].append(noise_syn_level)
st = time.time()
for nind in np.random.permutation(len(noise_syn_imgs[ind])):
temp_origin_img = origin_syn_imgs[ind]
temp_noise_img = noise_syn_imgs[ind][nind]
temp_noise_level = noise_syn_levels[ind][nind]
if np.random.randint(2, size=1)[0] == 1:
temp_origin_img = np.flip(temp_origin_img, axis=1)
temp_noise_img = np.flip(temp_noise_img, axis=1)
temp_noise_level = np.flip(temp_noise_level, axis=1)
if np.random.randint(2, size=1)[0] == 1:
temp_origin_img = np.flip(temp_origin_img, axis=0)
temp_noise_img = np.flip(temp_noise_img, axis=0)
temp_noise_level = np.flip(temp_noise_level, axis=0)
if np.random.randint(2, size=1)[0] == 1:
temp_origin_img = np.transpose(temp_origin_img, (1, 0, 2))
temp_noise_img = np.transpose(temp_noise_img, (1, 0, 2))
temp_noise_level = np.transpose(temp_noise_level, (1, 0, 2))
temp_noise_img_chw = hwc_to_chw(temp_noise_img)
temp_origin_img_chw = hwc_to_chw(temp_origin_img)
temp_noise_level_chw = hwc_to_chw(temp_noise_level)
cnt += 1
st = time.time()
input_var = torch.from_numpy(temp_noise_img_chw.copy()).type(torch.FloatTensor).unsqueeze(0)
target_var = torch.from_numpy(temp_origin_img_chw.copy()).type(torch.FloatTensor).unsqueeze(0)
noise_level_var = torch.from_numpy(temp_noise_level_chw.copy()).type(torch.FloatTensor).unsqueeze(0)
input_var, target_var, noise_level_var = input_var.cuda(), target_var.cuda(), noise_level_var.cuda()
noise_level_est, output = model(input_var)
loss = criterion(output, target_var, noise_level_est, noise_level_var, 1)
losses.update(loss.item())
optimizer.zero_grad()
loss.backward()
optimizer.step()
print('[{0}][{1}]\t'
'lr: {lr:.5f}\t'
'Loss: {loss.val:.4f} ({loss.avg:.4f})\t'
'Time: {time:.3f}'.format(
epoch, cnt,
lr=optimizer.param_groups[-1]['lr'],
loss=losses,
time=time.time()-st))
if epoch % save_freq == 0:
if not os.path.isdir(result_dir + '%04d'%epoch):
os.makedirs(result_dir + '%04d'%epoch)
output_np = output.squeeze().cpu().detach().numpy()
output_np = chw_to_hwc(np.clip(output_np, 0, 1))
temp = np.concatenate((temp_origin_img, temp_noise_img, output_np), axis=1)
scipy.misc.toimage(temp*255, high=255, low=0, cmin=0, cmax=255).save(result_dir + '%04d/train_%d_%d.jpg'%(epoch, ind, nind))
print('Training on real noisy images...')
for r in range(REAPET):
for ind in np.random.permutation(len(train_real_fns)):
train_real_fn = train_real_fns[ind]
if not len(origin_real_imgs[ind]):
train_real_origin_fns = glob.glob(train_real_fn + '/*Reference.bmp')
train_real_noise_fns = glob.glob(train_real_fn + '/*Noisy.bmp')
origin_real_img = cv2.imread(train_real_origin_fns[0])
origin_real_img = origin_real_img[:,:,::-1] / 255.0
origin_real_imgs[ind] = np.array(origin_real_img).astype('float32')
for train_real_noise_fn in train_real_noise_fns:
noise_real_img = cv2.imread(train_real_noise_fn)
noise_real_img = noise_real_img[:,:,::-1] / 255.0
noise_real_img = np.array(noise_real_img).astype('float32')
noise_real_imgs[ind].append(noise_real_img)
st = time.time()
for nind in np.random.permutation(len(noise_real_imgs[ind])):
H = origin_real_imgs[ind].shape[0]
W = origin_real_imgs[ind].shape[1]
ps_temp = min(H, W, PS) - 1
xx = np.random.randint(0, W-ps_temp)
yy = np.random.randint(0, H-ps_temp)
temp_origin_img = origin_real_imgs[ind][yy:yy+ps_temp, xx:xx+ps_temp, :]
temp_noise_img = noise_real_imgs[ind][nind][yy:yy+ps_temp, xx:xx+ps_temp, :]
if np.random.randint(2, size=1)[0] == 1:
temp_origin_img = np.flip(temp_origin_img, axis=1)
temp_noise_img = np.flip(temp_noise_img, axis=1)
if np.random.randint(2, size=1)[0] == 1:
temp_origin_img = np.flip(temp_origin_img, axis=0)
temp_noise_img = np.flip(temp_noise_img, axis=0)
if np.random.randint(2, size=1)[0] == 1:
temp_origin_img = np.transpose(temp_origin_img, (1, 0, 2))
temp_noise_img = np.transpose(temp_noise_img, (1, 0, 2))
temp_noise_img_chw = hwc_to_chw(temp_noise_img)
temp_origin_img_chw = hwc_to_chw(temp_origin_img)
cnt += 1
st = time.time()
input_var = torch.from_numpy(temp_noise_img_chw.copy()).type(torch.FloatTensor).unsqueeze(0)
target_var = torch.from_numpy(temp_origin_img_chw.copy()).type(torch.FloatTensor).unsqueeze(0)
input_var, target_var = input_var.cuda(), target_var.cuda()
noise_level_est, output = model(input_var)
loss = criterion(output, target_var, noise_level_est, 0, 0)
losses.update(loss.item())
optimizer.zero_grad()
loss.backward()
optimizer.step()
print('[{0}][{1}]\t'
'lr: {lr:.5f}\t'
'Loss: {loss.val:.4f} ({loss.avg:.4f})\t'
'Time: {time:.3f}'.format(
epoch, cnt,
lr=optimizer.param_groups[-1]['lr'],
loss=losses,
time=time.time()-st))
if epoch % save_freq == 0:
if not os.path.isdir(result_dir + '%04d'%epoch):
os.makedirs(result_dir + '%04d'%epoch)
output_np = output.squeeze().cpu().detach().numpy()
output_np = chw_to_hwc(np.clip(output_np, 0, 1))
temp = np.concatenate((temp_origin_img, temp_noise_img, output_np), axis=1)
scipy.misc.toimage(temp*255, high=255, low=0, cmin=0, cmax=255).save(result_dir + '%04d/train_%d_%d.jpg'%(epoch, ind + len(train_syn_fns) + r * len(train_real_fns), nind))
save_checkpoint({
'epoch': epoch + 1,
'state_dict': model.state_dict(),
'optimizer' : optimizer.state_dict()}, is_best=0)