-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathcgan.py
156 lines (127 loc) · 5.72 KB
/
cgan.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
import ljqpy, os, sys, time, random
import numpy as np
import keras.backend as K
from collections import defaultdict
from keras.datasets import mnist
from keras.utils.generic_utils import Progbar
from keras.optimizers import Adam
from keras.models import *
from keras.layers import *
from keras.preprocessing import image
from PIL import Image
time.clock()
np.random.seed(1333)
K.set_image_dim_ordering('tf')
from model import BuildGenerator, BuildDiscriminator
# params
nb_epochs = 200
batch_size = 1
p_lambda = 10
imgsize = 256
adam_lr = 0.000005
adam_beta_1 = 0.5
imgdirA = '/mnt/smb25/ImageNet/horse2zebra/trainA'
imgdirB = '/mnt/smb25/ImageNet/horse2zebra/trainB'
testimgdirA = '/mnt/smb25/ImageNet/horse2zebra/testA'
testimgdirB = '/mnt/smb25/ImageNet/horse2zebra/testB'
modeldir = 'data/'
testimgdir = 'images/'
modelG = BuildGenerator(Input(shape=(imgsize,imgsize,3)))
modelF = BuildGenerator(Input(shape=(imgsize,imgsize,3)))
modelDG = BuildDiscriminator(Input(shape=(imgsize,imgsize,3)))
modelDF = BuildDiscriminator(Input(shape=(imgsize,imgsize,3)))
modelG.summary()
modelDG.summary()
try:
modelG.load_weights( os.path.join(modeldir, 'modelG.h5' ))
modelF.load_weights( os.path.join(modeldir, 'modelF.h5' ))
modelDG.load_weights(os.path.join(modeldir, 'modelDG.h5'))
modelDF.load_weights(os.path.join(modeldir, 'modelDF.h5'))
except Exception as e:
print(e)
modelDG.compile(optimizer=Adam(adam_lr, adam_beta_1), loss='mse')
modelDF.compile(optimizer=Adam(adam_lr, adam_beta_1), loss='mse')
modelG.compile(optimizer=Adam(adam_lr, adam_beta_1), loss='mse')
modelF.compile(optimizer=Adam(adam_lr, adam_beta_1), loss='mse')
imageReal = Input(shape=(imgsize,imgsize,3))
imageFake = Input(shape=(imgsize,imgsize,3))
DGReal, DGFake = modelDG(imageReal), modelDG(imageFake)
combDG = Model(inputs=[imageReal, imageFake], outputs=[DGReal, DGFake])
combDG.compile(optimizer=Adam(adam_lr, adam_beta_1), loss='mse')
imageReal = Input(shape=(imgsize,imgsize,3))
imageFake = Input(shape=(imgsize,imgsize,3))
DFReal, DFFake = modelDF(imageReal), modelDF(imageFake)
combDF = Model(inputs=[imageReal, imageFake], outputs=[DFReal, DFFake])
combDF.compile(optimizer=Adam(adam_lr, adam_beta_1), loss='mse')
imageA = Input(shape=(imgsize,imgsize,3))
imageB = Input(shape=(imgsize,imgsize,3))
modelDG.trainable = False
modelDF.trainable = False
fakeB, fakeA = modelG(imageA), modelF(imageB)
disG, disF = modelDG(fakeB), modelDF(fakeA)
cycGF, cycFG = modelF(fakeB), modelG(fakeA)
combM = Model(inputs=[imageA, imageB], outputs=[disG, disF, cycGF, cycFG])
combM.compile(optimizer=Adam(adam_lr, adam_beta_1), loss=['mse','mse','mae','mae'], loss_weights=[1, 1, p_lambda, p_lambda])
def ImgGenerator(imgdir):
imglst = [os.path.join(imgdir, x) for x in os.listdir(imgdir)]
while True:
random.shuffle(imglst)
for fn in imglst:
img = image.load_img(fn, target_size=(imgsize, imgsize))
img = (image.img_to_array(img) - 127.5) / 127.5
yield np.expand_dims(img, axis=0)
# if the memory is enough for all imgs ...
def ImgGeneratorS(imgdir):
imglst = [os.path.join(imgdir, x) for x in os.listdir(imgdir)]
X = np.zeros((len(imglst), imgsize, imgsize, 3))
for i, fn in enumerate(imglst):
if i % 20 == 0: print('%d/%d' % (i, len(imglst)))
img = image.load_img(fn, target_size=(imgsize, imgsize))
X[i] = image.img_to_array(img)
X = (X - 127.5) / 127.5
ids = list(range(len(imglst)))
while True:
random.shuffle(ids)
for ii in ids: yield X[ii:ii+1]
genA = ImgGeneratorS(imgdirA)
genB = ImgGeneratorS(imgdirB)
testA = ImgGenerator(testimgdirA)
testB = ImgGenerator(testimgdirB)
nb_batches = len(os.listdir(imgdirA)) // batch_size
ones = np.ones( (batch_size, imgsize//16, imgsize//16, 1) )
zeros = np.zeros( (batch_size, imgsize//16, imgsize//16, 1) )
recordG, recordF = [], []
for epoch in range(nb_epochs):
print('Epoch %d of %d' % (epoch+1, nb_epochs))
progress_bar = Progbar(target=nb_batches)
lossDG, lossDF = 0.5, 0.5
for index in range(nb_batches):
A_image_batch, B_image_batch = next(genA), next(genB)
generateG = modelG.predict_on_batch(A_image_batch)
recordG.append(generateG)
if len(recordG) > 100: recordG = recordG[-50:]
lossDG = combDG.train_on_batch([B_image_batch, random.choice(recordG)], [ones, zeros])[0]
generateF = modelF.predict_on_batch(B_image_batch)
recordF.append(generateF)
if len(recordF) > 100: recordF = recordF[-50:]
lossDF = combDF.train_on_batch([A_image_batch, random.choice(recordF)], [ones, zeros])[0]
for _ in range(1):
A_image_batch, B_image_batch = next(genA), next(genB)
_, lossG, lossF, losscycGF, losscycFG = combM.train_on_batch([A_image_batch, B_image_batch], [ones, ones, A_image_batch, B_image_batch])
progress_bar.update(index+1, values=[('DG',lossDG),('G',lossG),('DF',lossDF),('F',lossF),('cycGF',losscycGF),('cycFG',losscycFG)])
print('Testing for epoch {}:'.format(epoch + 1))
modelG.save_weights( os.path.join(modeldir, 'modelG.h5' ), True)
modelF.save_weights( os.path.join(modeldir, 'modelF.h5' ), True)
modelDG.save_weights(os.path.join(modeldir, 'modelDG.h5'), True)
modelDF.save_weights(os.path.join(modeldir, 'modelDF.h5'), True)
tA = np.concatenate( [next(testA) for x in range(4)], axis=0 )
tB = np.concatenate( [next(testB) for x in range(4)], axis=0 )
gG = modelG.predict(tA, batch_size=1)
gF = modelF.predict(tB, batch_size=1)
tA = tA.reshape(-1,imgsize,3)
tB = tB.reshape(-1,imgsize,3)
gG = gG.reshape(-1,imgsize,3)
gF = gF.reshape(-1,imgsize,3)
img = np.concatenate([tA, gG, tB, gF], axis=1)
img = (img * 127.5 + 127.5).astype(np.uint8)
Image.fromarray(img).save(os.path.join(testimgdir, 'plot_epoch_{0:03d}_generated.png'.format(epoch)))