-
Notifications
You must be signed in to change notification settings - Fork 10
/
Copy pathmain.py
320 lines (286 loc) · 15.1 KB
/
main.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
# -*- coding: utf-8 -*-
"""
Created on Mon May 24 19:00:20 2021
@author: WIN
"""
import os
import cv2
import tensorflow as tf
import core.common as common
import core.utils as utils
import core.backbone as backbone
from core.config import cfg
import numpy as np
from core.dataset import Dataset
from tqdm import tqdm
from skimage import transform,data
from crossfusiongan import *
import tensorflow.contrib.slim as slim
import math
from crossGenerator import Generator
from Discriminator import Discriminator1, Discriminator2,Discriminator3
from LOSS import SSIM_LOSS, L1_LOSS, Fro_LOSS, _tf_fspecial_gauss,discriminator_loss ,generator_loss
os.environ["CUDA_VISIBLE_DEVICES"] = "0"
image_width = 512
image_height = 512
epochs = 400
images_vi = tf.compat.v1.placeholder(tf.float32, [None, None, None, 3], name='images_vi')
labels_vi = tf.compat.v1.placeholder(tf.float32, [None, None, None, 3], name='labels_vi')
images_de = tf.compat.v1.placeholder(tf.float32, [None, None, None, 3], name='images_de')
labels_de = tf.compat.v1.placeholder(tf.float32, [None, None, None, 3], name='labels_de')
images_in = tf.compat.v1.placeholder(tf.float32, [None, None, None, 3], name='images_in')
labels_in = tf.compat.v1.placeholder(tf.float32, [None, None, None, 3], name='labels_in')
label_sbbox = tf.placeholder(dtype=tf.float32, name='label_sbbox')
label_mbbox = tf.placeholder(dtype=tf.float32, name='label_mbbox')
label_lbbox = tf.placeholder(dtype=tf.float32, name='label_lbbox')
true_sbboxes = tf.placeholder(dtype=tf.float32, name='sbboxes')
true_mbboxes = tf.placeholder(dtype=tf.float32, name='mbboxes')
true_lbboxes = tf.placeholder(dtype=tf.float32, name='lbboxes')
trainable = tf.placeholder(dtype=tf.bool, name='training')
lr = tf.placeholder(tf.float32, None, name='learning_rate') # 训练中的学习率
lr0 = tf.placeholder(tf.float32, None, name='learning_rate0') # 训练中的学习率
lr1 = tf.placeholder(tf.float32, None, name='learning_rate1') # 训练中的学习率
#input_image=tf.concat([images_ir, images_vi], axis=-1)
G = Generator('Generator')
fusion_image = G.transform(images_vi,images_de,images_in)
D1 = Discriminator1('Discriminator1')
D2 = Discriminator2('Discriminator2')
D3 = Discriminator3('Discriminator3')
pos = D1.discrim(labels_vi, reuse = False)
pos2 = D2.discrim(labels_vi, reuse = False)
pos3= D3.discrim(labels_vi, reuse = False)
neg = D1.discrim(fusion_image, reuse=True)
neg2 = D2.discrim(fusion_image, reuse=True)
neg3 = D3.discrim(fusion_image, reuse=True)
# Loss for Generator
G_loss_GAN_D1 = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=neg, labels=tf.ones_like(neg))) # 生成器对抗损失（vi）
# G_loss_GAN_D1 = generator_loss(True, 'lsgan', real=pos, fake=neg)
G_loss_GAN_D2 = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=neg2, labels=tf.ones_like(neg2))) # 生成器对抗损失（de）
G_loss_GAN_D3 = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=neg3, labels=tf.ones_like(neg3))) # 生成器对抗损失（in）
#G_loss_GAN = G_loss_GAN_D1 + G_loss_GAN_D2+G_loss_GAN_D3
G_loss_GAN1 = G_loss_GAN_D1
G_loss_GAN2 = G_loss_GAN_D1+G_loss_GAN_D2
G_loss_GAN3 = G_loss_GAN_D1+G_loss_GAN_D2+G_loss_GAN_D3
LOSS_IN = Fro_LOSS(fusion_image - labels_in)
LOSS_VI = L1_LOSS(gradient(fusion_image) - gradient(labels_vi))
G_loss_norm = LOSS_IN / 48000000 + 0.000000004 * LOSS_VI # 传统损失函数
# G_loss_norm = 1.2 * LOSS_VI
G_loss1 = G_loss_GAN1*5 + 0.6 * G_loss_norm
G_loss2 = G_loss_GAN2*5 + 0.6 * G_loss_norm
G_loss3 = G_loss_GAN3*5 + 0.6 * G_loss_norm
# G_loss = 0.6 * G_loss_norm
# Loss for Discriminator1
D1_loss_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=pos, labels=tf.ones_like(pos)))
D1_loss_fake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=neg, labels=tf.zeros_like(neg)))
D1_loss = D1_loss_fake + D1_loss_real
# D1_loss = discriminator_loss(True, 'lsgan', real=pos, fake=neg)
D1_loss = D1_loss*5
# Loss for Discriminator2
D2_loss_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=pos2, labels=tf.ones_like(pos2)))
D2_loss_fake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=neg2, labels=tf.zeros_like(neg2)))
D2_loss = D2_loss_fake + D2_loss_real
D2_loss = D2_loss*5
# # Loss for Discriminator3
D3_loss_real = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=pos3, labels=tf.ones_like(pos3)))
D3_loss_fake = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits=neg3, labels=tf.zeros_like(neg3)))
D3_loss = D3_loss_fake + D3_loss_real
D3_loss = D3_loss*5
#检测模型
model = YOLOV3(fusion_image, trainable)
giou_loss, conf_loss, prob_loss = model.compute_loss(
label_sbbox, label_mbbox, label_lbbox,
true_sbboxes, true_mbboxes, true_lbboxes)
y_loss=giou_loss + conf_loss + prob_loss
g_loss1=G_loss1+y_loss
g_loss2=G_loss2+y_loss
g_loss3=G_loss3+y_loss
g_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='Generator')#取出Generator的参数
d1_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='Discriminator1')
d2_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='Discriminator2')
d3_vars = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, scope='Discriminator3')
y_vars= []
for var in tf.global_variables():
var_name = var.op.name
#print(var_name)
var_name_mess = str(var_name).split('/')
if var_name_mess[0] in ['darknet','conv52','conv53','conv54','conv55','conv56','conv57','conv58','conv59','conv60','conv61','conv62','conv63','conv64','conv65','conv66','conv67','conv68','conv_lobj_branch','conv_mobj_branch','conv_sobj_branch','conv_sbbox', 'conv_mbbox', 'conv_lbbox']:
y_vars.append(var)
first_stage_trainable_var_list= []
for var in tf.trainable_variables():
var_name = var.op.name
#print(var_name)
var_name_mess = str(var_name).split('/')
if var_name_mess[0] in ['conv_sbbox', 'conv_mbbox', 'conv_lbbox']:
first_stage_trainable_var_list.append(var)
# train_g1 = tf.train.AdamOptimizer(learning_rate=lr1).minimize(g_loss1, var_list=g_vars)
#train_g2 = tf.train.AdamOptimizer(learning_rate=lr1).minimize(g_loss2, var_list=g_vars)
train_g3 = tf.train.AdamOptimizer(learning_rate=lr1).minimize(g_loss3, var_list=g_vars)
train_d1 = tf.train.AdamOptimizer(learning_rate=lr1).minimize(D1_loss, var_list=d1_vars)
train_d2 = tf.train.AdamOptimizer(learning_rate=lr1).minimize(D2_loss, var_list=d2_vars)
train_d3 = tf.train.AdamOptimizer(learning_rate=lr1).minimize(D3_loss, var_list=d3_vars)
train_y = tf.train.AdamOptimizer(learning_rate=lr1).minimize(y_loss, var_list=y_vars)
train_y0 = tf.train.AdamOptimizer(learning_rate=lr1).minimize(y_loss, var_list=first_stage_trainable_var_list)
moving_ave = tf.train.ExponentialMovingAverage(0.9).apply(tf.trainable_variables())
vars1= []
for var in tf.global_variables():
var_name = var.op.name
#print(var_name)
var_name_mess = str(var_name).split('/')
if var_name_mess[0] in ['darknet','conv52','conv53','conv54','conv55','conv56','conv57','conv58','conv59','conv60','conv61','conv62','conv63','conv64','conv65','conv66','conv67','conv68','conv_lobj_branch','conv_mobj_branch','conv_sobj_branch','conv_sbbox', 'conv_mbbox', 'conv_lbbox']:
vars1.append(var)
vars2= []
for var in tf.global_variables():
var_name = var.op.name
#print(var_name)
var_name_mess = str(var_name).split('/')
if var_name_mess[0] in ['Generator','Generator_1','Generator_2','Generator_3','Generator_4','Generator_5','Generator_6','Generator_7']:
vars2.append(var)
vars3= []
for var in tf.global_variables():
var_name = var.op.name
#print(var_name)
var_name_mess = str(var_name).split('/')
if var_name_mess[0] in ['Discriminator1']:
vars3.append(var)
vars4= []
for var in tf.global_variables():
var_name = var.op.name
#print(var_name)
var_name_mess = str(var_name).split('/')
if var_name_mess[0] in ['Discriminator2']:
vars4.append(var)
vars5= []
for var in tf.global_variables():
var_name = var.op.name
#print(var_name)
var_name_mess = str(var_name).split('/')
if var_name_mess[0] in ['Discriminator3']:
vars5.append(var)
load_vars = vars1 + vars2 + vars3
vars = vars1 + vars2 + vars3 + vars4 + vars5
print(load_vars)
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
with tf.control_dependencies([train_y0]):
with tf.control_dependencies([moving_ave]):
train_y0 = tf.no_op()
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
with tf.control_dependencies([train_y]):
with tf.control_dependencies([moving_ave]):
train_y = tf.no_op()
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
with tf.control_dependencies([train_d1]):
with tf.control_dependencies([moving_ave]):
train_d1 = tf.no_op()
# with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
# with tf.control_dependencies([train_g1]):
# with tf.control_dependencies([moving_ave]):
# train_g1 = tf.no_op()
# with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
# with tf.control_dependencies([train_g2]):
# with tf.control_dependencies([moving_ave]):
# train_g2 = tf.no_op()
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
with tf.control_dependencies([train_g3]):
with tf.control_dependencies([moving_ave]):
train_g3 = tf.no_op()
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
with tf.control_dependencies([train_d2]):
with tf.control_dependencies([moving_ave]):
train_d2 = tf.no_op()
with tf.control_dependencies(tf.get_collection(tf.GraphKeys.UPDATE_OPS)):
with tf.control_dependencies([train_d3]):
with tf.control_dependencies([moving_ave]):
train_d3 = tf.no_op()
trainset = Dataset('train')
# testset = Dataset('test')
with tf.Session() as sess:
sess.run(tf.compat.v1.global_variables_initializer())#初始化模型参数
# loader1 = tf.train.Saver(var_list=y_vars) # 模型保存器
# loader1.restore(sess,'kitti/pre')
# loader2 = tf.train.Saver(var_list=load_vars) # 模型保存器
# loader2.restore(sess, 'cross/cross/model-124000')
saver = tf.train.Saver(var_list=vars,max_to_keep=50) # 模型保存器
saver.restore(sess, 'cross/model-164000')
step=164000
for epoch in range(epochs):
pbar = tqdm(trainset)
for train_data in pbar:
#忽略
if (step >= 0) and (step <= 100000):
lrate = 0.000001 + 0.5 * (0.0001 - 0.000001) * (
1 + math.cos((step) / (100000 ) * np.pi)) # 得到该训练epoch的学习率
if step > 100000:
lrate = 0.000001
#学习率在下面
if step<=10000:
lrate1 = 0.0001*(step+1000)/(10000+1000)
if (step>10000)and(step <= 204000):
lrate1=0.000001 + 0.5 * (0.0001 - 0.000001) *(1 + math.cos((step - 10000) / (204000 - 10000) * np.pi))# 得到该训练epoch的学习率
if step>204000:
lrate1=0.000001
#喂值
feed_dict = {lr: lrate,
lr1: lrate1,
images_vi: train_data[0], images_de: train_data[1], images_in: train_data[2],
labels_vi: train_data[0], labels_de: train_data[1], labels_in: train_data[2],
label_sbbox: train_data[3],
label_mbbox: train_data[4],
label_lbbox: train_data[5],
true_sbboxes: train_data[6],
true_mbboxes: train_data[7],
true_lbboxes: train_data[8],
trainable: True}
if step<124000:
_, _, _, gl, dl1, yl = sess.run([train_g1, train_y, train_d1, g_loss1, D1_loss, y_loss],
feed_dict=feed_dict)
if (step>=124000)and(step<164000):
_,_,_,_, gl, dl1,dl2, yl = sess.run([train_g2, train_y,train_d1, train_d2,g_loss2, D1_loss,D2_loss, y_loss],
feed_dict=feed_dict)
if step >= 164000:
_, _, _, _,_, gl, dl1, dl2,dl3, yl = sess.run(
[train_g3, train_y, train_d1, train_d2, train_d3, g_loss3, D1_loss, D2_loss,D3_loss, y_loss],
feed_dict=feed_dict)
step+=1
print('epoch: %d, step: %d,learnrate: %f, Generator Loss: %f, DisLoss: %f,YOLO Loss: %f' % (epoch, step, lrate1,gl,dl3,yl))
if (step%100)==0:
#print(train_data[0])
g = sess.run(fusion_image, feed_dict = {images_vi: train_data[0], images_de: train_data[1],images_in: train_data[2]})
for j in range(1):
cv2.imwrite('crossresult/' + str(step) + "_" + str(j)+'.png', g[j]*255.)
if (step%4000)==0:
# tbar = tqdm(testset)
# loss = 0
# yloss = 0
# # saver.save(sess, 'kitti/model', step)
# # saver1.save(sess, 'kitti/modelone', step)
# #
# # saver1.restore(sess, 'kitti/'+'modelone'+"-"+str(step))
# #
# # saver3 = tf.train.Saver(tf.trainable_variables())
# # saver3.save(sess, 'kitti/model', step)
# for test_data in tbar:
# gl, dl1, yl = sess.run([g_loss, D1_loss, y_loss],
# feed_dict={
# images_vi: test_data[0], images_de: test_data[1],
# images_in: test_data[2],
# labels_vi: test_data[0], labels_de: test_data[1],
# labels_in: test_data[2],
# label_sbbox: test_data[3],
# label_mbbox: test_data[4],
# label_lbbox: test_data[5],
# true_sbboxes: test_data[6],
# true_mbboxes: test_data[7],
# true_lbboxes: test_data[8],
# trainable: True})
# loss = loss + gl + dl1
# yloss = yloss + yl
# print('Loss: %f %f %d' % (gl + dl1 , yl, step))
# loss = loss / 472.
# yloss = yloss / 472.
# saver1.save(sess, 'cross/modelone', step)
#
# saver1.restore(sess, 'cross/'+'modelone'+"-"+str(step))
#
# saver3 = tf.train.Saver(tf.trainable_variables())
# saver3.save(sess, 'cross/model', step)
saver.save(sess, 'cross/model',step)