-
Notifications
You must be signed in to change notification settings - Fork 0
/
train1.py
160 lines (147 loc) · 8.38 KB
/
train1.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
from model import *
import keras.backend as K
from keras.optimizers import Adam
from data_generator import image_generator
from config import *
import os
import tensorflow as tf
from keras.backend.tensorflow_backend import set_session
from PIL import Image
os.environ["CUDA_VISIBLE_DEVICES"] = "3"
config = tf.ConfigProto()
# config.gpu_options.per_process_gpu_memory_fraction = 1
config.gpu_options.allow_growth = True
set_session(tf.Session(config=config))
train_step_per_epoch = len(os.listdir(image_source_dir + 'trainA')) / batch_size + 1
test_step_per_epoch = len(os.listdir(image_source_dir + 'testA')) / batch_size + 1
train_image_generator = image_generator(image_source_dir + 'trainA',
image_source_dir + 'trainB', batch_size=batch_size,
shuffle=True)
test_image_generator = image_generator(image_source_dir + 'testA',
image_source_dir + 'testB', batch_size=batch_size,
shuffle=False)
#warning,the original metric(acc) and loss function(mse,mae) is defined with axis=-1,because the output is 2D(batch_size*N),in our cases the output is 4D(batch_size*H*W*C), we should adapt them to our case
def acc1(y_true, y_pred):
return K.mean(K.equal(y_true, K.round(y_pred)), axis=[1,2,3])
#for patch discriminator,output is batch_size*H*W*1
def mse_custom(y_true, y_pred):
return K.mean(K.square(y_pred - y_true), axis=[1,2,3])
#for generator,output is batch_size*H*W*3
def mae_custom(y_true, y_pred):
return K.mean(K.abs(y_pred - y_true), axis=[1,2,3])
def mape_custom(y_true, y_pred):
diff = K.abs((y_true - y_pred) / K.clip(K.abs(y_true),K.epsilon(),None))
return 100. * K.mean(diff, axis=[1,2,3])
opt1 = Adam(lr=lr)
opt2 = Adam(lr=lr)
discriminator = get_discriminator_1(n_layers=int(np.log(downsample)/np.log(2)),instance_norm=use_instance_norm,name='discriminator_a')
discriminator.compile(optimizer=opt1, loss=mse_custom, metrics=[acc1])
print(discriminator.summary())
generator_A2B = get_generator(name='a2b',instance_norm=use_instance_norm)
generator_B2A = get_generator(name='b2a',instance_norm=use_instance_norm)
# generator_A2B.compile(optimizer='Adam', loss='mae', metrics=['mean_absolute_percentage_error'])
print(generator_A2B.summary())
generator_train = get_generator_training_model_1(generator_A2B, generator_B2A, discriminator)
print(generator_train.summary())
generator_train.compile(optimizer=opt2, loss=[mae_custom, mae_custom, mae_custom, mae_custom, mse_custom, mse_custom ],
metrics=[mape_custom],
loss_weights=[5, 5, 10, 10, 1, 1])
#generator_train.compile(optimizer=opt3, loss=[mae_custom, mae_custom, mse_custom, mse_custom, ],
# metrics=[mape_custom],
# loss_weights=[10, 10, 1, 1])
if os.path.exists(combined_filepath):
generator_train.load_weights(combined_filepath, by_name=True)
generator_A2B.load_weights(generator_a2b_filepath, by_name=True)
generator_B2A.load_weights(generator_b2a_filepath, by_name=True)
print('weights loaded!')
def imsave(img,filename):
img = (img + 1) * 127.5
img = np.clip(img, 0, 255)
img = img.astype(np.uint8)
img = Image.fromarray(img)
img.save(filename)
real_A = np.zeros((batch_size, image_size/downsample, image_size/downsample, 3))
real_A[:,:,:,0]=1
real_B = np.zeros((batch_size, image_size/downsample, image_size/downsample, 3))
real_B[:,:,:,1]=1
fake = np.zeros((batch_size, image_size/downsample, image_size/downsample, 3))
fake[:,:,:,2]=1
best_loss = 1000
for i in range(epoch):
train_step = 0
for imgA, imgB in train_image_generator:
train_step += 1
if train_step > train_step_per_epoch:
test_step = 0
total_loss = 0
total_mape = 0
for imgA, imgB in test_image_generator:
test_step += 1
if test_step > test_step_per_epoch:
break
# print generator_train.metrics_names
#gloss, _, _, _, _, mape1, mape2, mape3, mape4 = generator_train.test_on_batch(
# [imgA, imgB], [imgA, imgB, real, real])
gloss, _, _, _, _, _, _, mape1, mape2, mape3, mape4, mape5, mape6 = generator_train.test_on_batch([imgA, imgB], [imgA, imgB,imgA, imgB, real_A, real_B])
total_loss += gloss
#total_mape += sum([mape1, mape2, mape3, mape4])
total_mape += sum([mape1, mape2, mape3, mape4, mape5, mape6])
print('epoch:{} test loss g:{} \n test mape:{}'.format(i + 1, total_loss / (test_step - 1),
total_mape / (test_step - 1)))
if total_loss / (test_step - 1) < best_loss:
print('test loss improved from {} to {}'.format(best_loss, total_loss / (test_step - 1)))
best_loss = total_loss / (test_step - 1)
generator_train.save_weights(combined_filepath, overwrite=True)
generator_A2B.save_weights(generator_a2b_filepath, overwrite=True)
generator_B2A.save_weights(generator_b2a_filepath, overwrite=True)
break
#print(discriminator_A.summary())
#print(discriminator_B.summary())
fakeB = generator_A2B.predict(imgA)
fakeA = generator_B2A.predict(imgB)
if debug:
fn=output_dir + str(i + 1) + '_' + str(train_step) + '_fake_a2b.png'
imsave(fakeB[0],fn)
#print("{} saved".format(fn))
fn=output_dir + str(i + 1) + '_' + str(train_step) + '_real_a.png'
imsave(imgA[0],fn)
#print("{} saved".format(fn))
fn=output_dir + str(i + 1) + '_' + str(train_step) + '_fake_b2a.png'
imsave(fakeA[0],fn)
#print("{} saved".format(fn))
fn=output_dir + str(i + 1) + '_' + str(train_step) + '_real_b.png'
imsave(imgB[0],fn)
#print("{} saved".format(fn))
# print('realB:', imgB[0], imgB.shape)
# print descriminator.trainable
#d_fakeA = discriminator_A.predict(fakeA)
#d_realA = discriminator_A.predict(imgA)
#d_fakeB = discriminator_B.predict(fakeB)
#d_realB = discriminator_B.predict(imgB)
#print('d_real_A:', np.squeeze(d_realA[0]), d_realA.shape)
#print('d_fake_A:', np.squeeze(d_fakeA[0]))
#print('d_real_B:', np.squeeze(d_realB[0]))
#print('d_fake_B:', np.squeeze(d_fakeB[0]))
discriminator.trainable = True
generator_A2B.trainable = False
generator_B2A.trainable = False
loss_fakeA, fake_accA = discriminator.train_on_batch(fakeA, fake)
loss_realA, real_accA = discriminator.train_on_batch(imgA, real_A)
loss_fakeB, fake_accB = discriminator.train_on_batch(fakeB, fake)
loss_realB, real_accB = discriminator.train_on_batch(imgB, real_B)
print('epoch:{} train step:{}, loss d_fake:{:.2}, loss d_real:{:.2}, fake_acc:{:.2}, real_acc:{:.2}'.format(i + 1, train_step,loss_fakeA + loss_fakeB,loss_realA + loss_realB,(fake_accA + fake_accB) * 0.5,(real_accA + real_accB) * 0.5))
# print descriminator.metrics_names
if i+1<pretrain_epoch or (train_step>pretrain_step_start and train_step<pretrain_step_end):
#if (i+1)%2==1 or i<pretrain_epoch:# or (train_step>pretrain_step_start and train_step<pretrain_step_end):
continue
discriminator.trainable = False
generator_A2B.trainable = True
generator_B2A.trainable = True
#print(generator_train.summary())
# print generator_train.metrics_names
#loss = generator_train.train_on_batch([imgA, imgB], [imgA, imgB, imgA, imgB, real, real])
#loss = generator_train.train_on_batch([imgA, imgB], [imgA, imgB, real, real])
loss = generator_train.train_on_batch([imgA, imgB], [imgA, imgB, imgA, imgB, real_A, real_B])
# print descriminator.trainable
# print('epoch:{} train step:{} loss cycle:{:.2} loss fool:{:.2}'.format(i + 1, train_step,loss[1] + loss[2],loss[3] + loss[4]))
print('epoch:{} train step:{} loss identity:{:.2} loss cycle:{:.2} loss gan:{:.2}'.format(i + 1, train_step,loss[1] + loss[2],loss[3] + loss[4],loss[5] + loss[6]))