-
Notifications
You must be signed in to change notification settings - Fork 4
/
training_fusion_lrr.py
322 lines (283 loc) · 11.2 KB
/
training_fusion_lrr.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
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
# -*- coding:utf-8 -*-
# @Author: Li Hui, Jiangnan University
# @Email: hui_li_jnu@163.com
# @File : training_fusion_lrr.py
# @Time : 2020/6/29 21:01
import os
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
import time
import scipy.io as scio
import torch
from torch.optim import Adam
from torch.autograd import Variable
from visdom import Visdom
# from net import LRR_NET, Vgg16
# from net_nuclear import LRR_NET, Vgg16
from net_lista import LRR_NET, Vgg16
from args import Args as args
import utils
import random
EPSILON = 1e-5
def load_data(path, train_num):
imgs_path, _ = utils.list_images(path)
imgs_path = imgs_path[:train_num]
random.shuffle(imgs_path)
return imgs_path
def main():
fusion_type = 'cat' # cat, add
# True - RGB, False - gray
if args.channel == 1:
img_flag = False
else:
img_flag = True
path = args.path_ir
train_num = 20000
data = load_data(path, train_num)
w_vi = 0.5
w_ir_list = [1.0, 2.0, 3.0, 4.0, 5.0, 6.0]
lam2_vi_list = [0.1, 0.5, 1.0, 1.5, 2.0, 2.5]
lam3_gram_list = [1500, 2000, 2500]
num_block_list = [2, 4, 6, 8]
# for idx_wir in range(2, 3):
# w_ir = w_ir_list[idx_wir]
# for idx_lam2 in range(3, 4):
# lam2_vi = lam2_vi_list[idx_lam2]
# for idx_lam3 in range(1, 2):
# lam3_gram = lam3_gram_list[idx_lam3]
# for idx_num in range(1, 2):
# num_block = num_block_list[idx_num]
# train(data, img_flag, fusion_type, lam2_vi, w_vi, w_ir, lam3_gram, num_block)
w_ir = w_ir_list[2]
lam2_vi = lam2_vi_list[3]
lam3_gram = lam3_gram_list[1]
num_block = num_block_list[1]
train(data, img_flag, fusion_type, lam2_vi, w_vi, w_ir, lam3_gram, num_block)
def train(data, img_flag, fusion_type, lam2_vi, w_vi, w_ir, lam3_gram, num_block):
batch_size = args.batch_size
# fusion network
model_or = LRR_NET(args.s, args.n, args.channel, args.stride, num_block, fusion_type)
# model = torch.nn.DataParallel(model_or, list(range(torch.cuda.device_count()))).cuda()
# model = torch.nn.DataParallel(model, device_ids=[0])
model = model_or
if args.resume_model is not None:
print('Resuming, initializing fusion net using weight from {}.'.format(args.resume_model))
model.load_state_dict(torch.load(args.resume_model))
optimizer = Adam(model.parameters(), args.lr, weight_decay=0.9)
mse_loss = torch.nn.MSELoss()
# visdom
viz = Visdom()
# Loss network - VGG16
vgg = Vgg16()
utils.init_vgg16(vgg, os.path.join(args.vgg_model_dir, "vgg16.pth"))
if args.cuda:
model.cuda()
vgg.cuda()
# tbar = trange(args.epochs)
print('Start training.....')
# creating save path
temp_path_model = os.path.join(args.save_fusion_model)
temp_path_loss = os.path.join(args.save_loss_dir)
if os.path.exists(temp_path_model) is False:
os.mkdir(temp_path_model)
if os.path.exists(temp_path_loss) is False:
os.mkdir(temp_path_loss)
Loss_list1 = []
Loss_list2 = []
Loss_list3 = []
Loss_list4 = []
Loss_list5 = []
Loss_list_all = []
count = 0
viz_index = 0
loss_p1 = 0.
loss_p2 = 0.
loss_p3 = 0.
loss_p4 = 0.
loss_p5 = 0.
loss_all = 0.
model.train()
for e in range(args.epochs):
img_paths, batch_num = utils.load_dataset(data, batch_size)
for idx in range(batch_num):
image_paths_ir = img_paths[idx * batch_size:(idx * batch_size + batch_size)]
img_ir = utils.get_train_images(image_paths_ir, height=args.Height, width=args.Width, flag=img_flag)
image_paths_vi = [x.replace('lwir', 'visible') for x in image_paths_ir]
img_vi = utils.get_train_images(image_paths_vi, height=args.Height, width=args.Width, flag=img_flag)
count += 1
optimizer.zero_grad()
batch_ir = Variable(img_ir, requires_grad=False)
batch_vi = Variable(img_vi, requires_grad=False)
if args.cuda:
batch_ir = batch_ir.cuda()
batch_vi = batch_vi.cuda()
# normalize for each batch
batch_ir_in = utils.normalize_tensor(batch_ir)
batch_vi_in = utils.normalize_tensor(batch_vi)
# get fusion image
output = model(batch_ir_in, batch_vi_in)
# out = {'fea_x_l': fea_x_l, 'fea_x_s': fea_x_s,
# 'fea_y_l': fea_y_l, 'fea_y_s': fea_y_s,
# 'x_l': x_l, 'x_h': x_h,
# 'y_l': y_l, 'y_h': y_h,
# 'fl': fl, 'fh': fh, 'fuse': f}
# fused image
out_f = output['fuse']
out_f = utils.normalize_tensor(out_f)
out_f = out_f * 255
# ---------- LOSS FUNCTION ----------
loss_pixel_vi = 10 * mse_loss(out_f, batch_vi)
# --- Feature loss ----
vgg_outs = vgg(out_f)
vgg_irs = vgg(batch_ir)
vgg_vis = vgg(batch_vi)
t_idx = 0
loss_fea_vi = 0.
loss_fea_ir = 0.
loss_gram_ir = 0.
weights_fea = [lam2_vi, 0.01, 0.5, 0.1]
weights_gram = [0, 0, 0.1, lam3_gram]
# w_ir = 4.0
# w_vi = 0.5
for fea_out, fea_ir, fea_vi, w_fea, w_gram in zip(vgg_outs, vgg_irs, vgg_vis, weights_fea, weights_gram):
if t_idx == 0:
loss_fea_vi = w_fea * mse_loss(fea_out, fea_vi)
if t_idx == 1 or t_idx == 2:
# relu2_2, relu3_3, relu4_3
loss_fea_ir += w_fea * mse_loss(fea_out, w_ir * fea_ir + w_vi * fea_vi)
if t_idx == 3:
gram_out = utils.gram_matrix(fea_out)
gram_ir = utils.gram_matrix(fea_ir)
loss_gram_ir += w_gram * mse_loss(gram_out, gram_ir)
t_idx += 1
# total loss
total_loss = loss_pixel_vi + loss_fea_vi + loss_fea_ir + loss_gram_ir
# total loss
total_loss.backward()
optimizer.step()
loss_p1 += loss_pixel_vi
loss_p2 += loss_fea_vi
loss_p3 += loss_fea_ir
loss_p4 += loss_gram_ir
loss_all += total_loss
step = 10
if count % step == 0:
loss_p1 /= step
loss_p2 /= step
loss_p3 /= step
loss_p4 /= step
loss_p5 /= step
loss_all /= step
if e == 0 and count == step:
viz.line([loss_all.item()], [0.], win='total_loss', opts=dict(title='Total Loss'))
viz.line([loss_p1.item()], [0.], win='pixel_loss', opts=dict(title='Pixel Loss'))
viz.line([loss_p2.item()], [0.], win='shallow_loss', opts=dict(title='Shallow Loss'))
viz.line([loss_p3.item()], [0.], win='middle_loss', opts=dict(title='Middle Loss'))
viz.line([loss_p4.item()], [0.], win='deep_loss', opts=dict(title='Deep Loss'))
mesg = "{}\t lam2 {}\t w_ir {}\t lam3 {}\t Count {} \t Epoch {}/{} \t Batch {}/{} \t block num: {} \n " \
"pixel vi loss: {:.6f} \t fea vi loss: {:.6f} \t " \
"fea ir loss: {:.6f} \t gram ir loss: {:.6f} \n " \
"total loss: {:.6f} \n". \
format(time.ctime(), lam2_vi, w_ir, lam3_gram, count, e + 1, args.epochs, idx + 1, batch_num,
num_block, loss_p1, loss_p2, loss_p3, loss_p4, loss_all)
print(mesg)
viz.line([loss_all.item()], [viz_index], win='total_loss', update='append')
viz.line([loss_p1.item()], [viz_index], win='pixel_loss', update='append')
viz.line([loss_p1.item()], [viz_index], win='shallow_loss', update='append')
viz.line([loss_p3.item()], [viz_index], win='middle_loss', update='append')
viz.line([loss_p4.item()], [viz_index], win='deep_loss', update='append')
viz_index = viz_index + 1
Loss_list1.append(loss_p1.item())
Loss_list2.append(loss_p2.item())
Loss_list3.append(loss_p3.item())
Loss_list4.append(loss_p4.item())
Loss_list_all.append(total_loss.item())
loss_p1 = 0.
loss_p2 = 0.
loss_p3 = 0.
loss_p4 = 0.
loss_p5 = 0.
loss_all = 0.
if count % 1000 == 0:
temp_loss = str(lam2_vi) + "wir_" + str(w_ir) + "_lam3_gram_" + str(lam3_gram) + "_epoch_" + str(e + 1) + "_batch_" + str(idx + 1) + \
"_block_" + str(num_block) + str(time.ctime()).replace(' ', '_').replace(':', '_') + ".mat"
# save 1 loss
loss_filename_path = "loss_1_lam2_" + temp_loss
save_loss_path = os.path.join(temp_path_loss, loss_filename_path)
scio.savemat(save_loss_path, {'loss_1': Loss_list1})
# save 2 loss
loss_filename_path = "loss_2_lam2_" + temp_loss
save_loss_path = os.path.join(temp_path_loss, loss_filename_path)
scio.savemat(save_loss_path, {'loss_2': Loss_list2})
# save 3 loss
loss_filename_path = "loss_3_lam2_" + temp_loss
save_loss_path = os.path.join(temp_path_loss, loss_filename_path)
scio.savemat(save_loss_path, {'loss_3': Loss_list3})
# save 4 loss
loss_filename_path = "loss_4_lam2_" + temp_loss
save_loss_path = os.path.join(temp_path_loss, loss_filename_path)
scio.savemat(save_loss_path, {'loss_4': Loss_list4})
# save 5 loss
loss_filename_path = "loss_5_lam2_" + temp_loss
save_loss_path = os.path.join(temp_path_loss, loss_filename_path)
scio.savemat(save_loss_path, {'loss_5': Loss_list5})
# save total loss
loss_filename_path = "loss_all_lam2_" + temp_loss
save_loss_path = os.path.join(temp_path_loss, loss_filename_path)
scio.savemat(save_loss_path, {'loss_all': Loss_list_all})
if count % 2000 == 0:
# save model ever 2000 iter.
model.eval()
model.cpu()
save_model_filename = "lrr_net_lam2_" + str(lam2_vi) + "wir_" + str(w_ir) + "_lam3_gram_" + str(lam3_gram) + \
"_epoch_" + str(e + 1) + "_count_" + str(count) + "_block_" + str(num_block) + ".model"
save_model_path = os.path.join(temp_path_model, save_model_filename)
torch.save(model.state_dict(), save_model_path)
print('Saving model at ' + save_model_path + '......')
##############
model.train()
model.cuda()
# save model
model.eval()
model.cpu()
save_model_filename = "lrr_net_lam2_" + str(lam2_vi) + "wir_" + str(w_ir) + "_lam3_gram_" + str(lam3_gram) + \
"_epoch_" + str(e + 1) + "_block_" + str(num_block) + ".model"
save_model_path = os.path.join(temp_path_model, save_model_filename)
torch.save(model.state_dict(), save_model_path)
##############
model.train()
model.cuda()
print("\nCheckpoint, trained model saved at: " + save_model_path)
final_temp = str(lam2_vi) + "_wir_" + str(w_ir) + "_lam3_gram_" + str(lam3_gram) + "_epoch_" + str(args.epochs)+ "_block_" + str(num_block)
# save 1 loss
loss_filename_path = "final_loss_1_lam2_" + final_temp + ".mat"
save_loss_path = os.path.join(temp_path_loss, loss_filename_path)
scio.savemat(save_loss_path, {'loss_1': Loss_list1})
# save 2 loss
loss_filename_path = "final_loss_2_lam2_" + final_temp + ".mat"
save_loss_path = os.path.join(temp_path_loss, loss_filename_path)
scio.savemat(save_loss_path, {'loss_2': Loss_list2})
# save 3 loss
loss_filename_path = "final_loss_3_lam2_" + final_temp + ".mat"
save_loss_path = os.path.join(temp_path_loss, loss_filename_path)
scio.savemat(save_loss_path, {'loss_3': Loss_list3})
# save 4 loss
loss_filename_path = "final_loss_4_lam2_" + final_temp + ".mat"
save_loss_path = os.path.join(temp_path_loss, loss_filename_path)
scio.savemat(save_loss_path, {'loss_4': Loss_list4})
# save 5 loss
loss_filename_path = "final_loss_5_lam2_" + final_temp + ".mat"
save_loss_path = os.path.join(temp_path_loss, loss_filename_path)
scio.savemat(save_loss_path, {'loss_5': Loss_list5})
# save total loss
loss_filename_path = "final_loss_all_lam2_" + final_temp + ".mat"
save_loss_path = os.path.join(temp_path_loss, loss_filename_path)
scio.savemat(save_loss_path, {'loss_all': Loss_list_all})
# save model
model.eval()
model.cpu()
save_model_filename = "final_lrr_net_lam2_" + final_temp + ".model"
save_model_path = os.path.join(temp_path_model, save_model_filename)
torch.save(model.state_dict(), save_model_path)
print("\nDone, trained model saved at", save_model_path)
if __name__ == "__main__":
main()