-
Notifications
You must be signed in to change notification settings - Fork 8
/
train.py
363 lines (289 loc) · 14 KB
/
train.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
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
import os, random
os.environ["KERAS_BACKEND"] = "tensorflow"
from keras.layers import Conv2D, MaxPooling2D, GlobalAveragePooling2D
import h5py
import numpy as np
from keras.utils import plot_model
from keras.layers import Input, merge, Lambda
from keras.layers.core import Reshape, Dense, Dropout, Activation, Flatten
from keras.layers.advanced_activations import LeakyReLU
from keras.layers.convolutional import Convolution2D, MaxPooling2D, ZeroPadding2D, UpSampling2D, AveragePooling2D, \
Conv2DTranspose
from keras.layers.normalization import *
from keras.optimizers import *
from keras import initializers
import matplotlib.pyplot as plt
import cPickle, random, sys, keras
from keras.models import Model
from tqdm import tqdm
import time
import os, sys
from functools import partial
normal = partial(initializers.normal, scale=.02)
## load and preprocess the dataset ##
batch_size = 256
num_ep = 7
num_pp = 8
epochs = 300
img_rows, img_cols = 64, 64
c_dim = num_pp
date = 2018
print ('Loading data...')
f = h5py.File('/data')
print ('Finished loading....')
epsilon_std = 1
def sampling_np(args):
z_mean, z_log_var = args
epsilon = np.random.normal(loc=0., scale=epsilon_std, size=(z_mean.shape[0], z_dim), )
return z_mean + np.exp(z_log_var / 2) * epsilon
def generate_dataset(ee):
## save to numpyz###############
c = np.random.randint(num_pp, size=x_train.shape[0])
c_train = keras.utils.to_categorical(c, num_pp)
c = np.random.randint(num_pp, size=x_test.shape[0])
c_test = keras.utils.to_categorical(c, num_pp)
[z_train, mean_var_train] = encoder.predict(x_train)
encoded_xtrain = decoder.predict([z_train, c_train])
[z_test, mean_var_test] = encoder.predict(x_test)
encoded_xtest = decoder.predict([z_test, c_test])
np.savez('/Z_' + str(date) + 'epoch'+str(ee)+'_64_64_VAE_GAN_labelfull_v2.npz',
encoded_xtrain, y_train1, y_train2, c_train, encoded_xtest, y_test1, y_test2, c_test)
np.savez('/X_' + str(date) + 'epoch'+str(ee)+ '_fi_512_VAE_GAN_labelfull_v2.npz',
z_train, y_train1, y_train2, c_train, z_test, y_test1, y_test2, c_test)
opt = RMSprop(lr=0.0003, decay=1e-6)
dopt = RMSprop(lr=0.0003, decay=1e-6)
def KL_loss(y_true, y_pred):
z_mean = y_pred[:, 0:z_dim]
z_log_var = y_pred[:, z_dim:2 * z_dim]
kl_loss = - 0.5 * K.sum(1 + z_log_var - K.square(z_mean) - K.exp(z_log_var), axis=-1)
return K.mean(kl_loss)
def sampling(args):
z_mean, z_log_sigma = args
epsilon = K.random_normal(shape=(K.shape(z_mean)[0], z_dim), mean=0.,
stddev=epsilon_std)
return z_mean + K.exp(K.square(z_log_sigma) / 2) * epsilon
def model_encoder(z_dim, input_shape, units=512, dropout=0.3):
k = 8
x = Input(input_shape)
h = Conv2D(units / 8, (k, k), strides=(2, 2), border_mode='same')(x)
h = BatchNormalization(momentum=0.8)(h)
h = Dropout(dropout)(h)
# h = MaxPooling2D(pool_size=(2, 2))(h)
h = LeakyReLU(0.2)(h)
h = Conv2D(units / 4, (k, k), strides=(2, 2), border_mode='same')(h)
h = BatchNormalization(momentum=0.8)(h)
h = Dropout(dropout)(h)
# h = MaxPooling2D(pool_size=(2, 2))(h)
h = LeakyReLU(0.2)(h)
h = Conv2D(units / 2, (k, k), strides=(2, 2), border_mode='same')(h)
h = BatchNormalization(momentum=0.8)(h)
h = Dropout(dropout)(h)
# h = MaxPooling2D(pool_size=(2, 2))(h)
h = LeakyReLU(0.2)(h)
h = Conv2D(units, (k, k), strides=(2, 2), border_mode='same')(h)
h = BatchNormalization(momentum=0.8)(h)
h = Dropout(dropout)(h)
h = LeakyReLU(0.2)(h)
h = Flatten()(h)
mean = Dense(z_dim, name="encoder_mean")(h)
logvar = Dense(z_dim, name="encoder_sigma", activation='sigmoid')(h)
z = Lambda(sampling, output_shape=(z_dim,))([mean, logvar])
h2 = keras.layers.concatenate([mean, logvar])
return Model(x, [z, h2], name='Encoder')
def model_decoder(z_dim, c_dim):
k = 8
x = Input(shape=(z_dim,))
auxiliary_c = Input(shape=(c_dim,), name='aux_input_c')
h = keras.layers.concatenate([x, auxiliary_c])
h = Dense(4 * 4 * 128, activation='relu')(h)
h = Reshape((4, 4, 128))(h)
# h = LeakyReLU(0.2)(h)
h = Conv2DTranspose(units, (k, k), strides=(2, 2), padding='same', activation='relu')(h) # 32*32*64
# h = Dropout(dropout)(h)
h = BatchNormalization(momentum=0.8)(h)
# h = LeakyReLU(0.2)(h)
# h = UpSampling2D(size=(2, 2))(h)
h = Conv2DTranspose(units / 2, (k, k), strides=(2, 2), padding='same', activation='relu')(h) # 64*64*64
# h = Dropout(dropout)(h)
h = BatchNormalization(momentum=0.8)(h)
# h = LeakyReLU(0.2)(h)
# h = UpSampling2D(size=(2, 2))(h)
h = Conv2DTranspose(units / 2, (k, k), strides=(2, 2), padding='same', activation='relu')(h) # 8*6*64
# h = Dropout(dropout)(h)
h = BatchNormalization(momentum=0.8)(h)
h = Conv2DTranspose(3, (k, k), strides=(2, 2), padding='same', activation='tanh')(h) # 8*6*64
return Model([x, auxiliary_c], h, name="Decoder")
################################################ Build the discrminator ###########################################################################
input_shape = (img_rows, img_cols, 3)
loss_weights_1= Input(shape=(1,), name='disc_1')
loss_weights_2= Input(shape=(1,),name='disc_2')
loss_weights_3= Input(shape=(1,),name='disc_3')
targets1 = Input(shape = (1,),name='disc_4')
targets2 = Input(shape = (num_pp,),name='disc_5')
targets3 = Input(shape = (num_ep,),name='disc_6')
d_input = Input(input_shape,name='disc_7')
rep_field = 8
x = Conv2D(32, (rep_field, rep_field), strides=(2, 2), padding='same', name='id_conv1')(d_input)
# x = BatchNormalization(momentum=0.9)(x)
x = LeakyReLU(0.2)(x)
# x = AveragePooling2D((2, 2), padding='same')(x)
# x = Dropout(0.3)(x)
x = Conv2D(64, (rep_field, rep_field), strides=(2, 2), padding='same', name='id_conv2')(x)
# x = BatchNormalization(momentum=0.9)(x)
x = LeakyReLU(0.2)(x)
# x = AveragePooling2D((2, 2), padding='same')(x)
# x = Dropout(0.3)(x)
x = Conv2D(128, (rep_field, rep_field), strides=(2, 2), padding='same', name='id_conv3')(x)
# x = BatchNormalization(momentum=0.9)(x)
x = LeakyReLU(0.2)(x)
# x = AveragePooling2D((2, 2), padding='same')(x)
# x = Dropout(0.3)(x)
x = Conv2D(256, (rep_field, rep_field), strides=(2, 2), padding='same', name='id_conv4')(x)
# x = BatchNormalization(momentum=0.9)(x)
x = LeakyReLU(0.2)(x)
x = Flatten()(x)
x = Dense(256, name='ds')(x)
x = LeakyReLU(0.2)(x)
x = Dropout(0.5)(x)
output_binary = Dense(1, activation='sigmoid', name='bin_real')(x)
output_identity = Dense(num_pp, activation='softmax', name='id_real')(x)
output_expression = Dense(num_ep, activation='softmax', name='exp_real')(x)
discriminator = Model([d_input, loss_weights_1, loss_weights_2,loss_weights_3, targets1, targets2, targets3], [output_binary, output_identity, output_expression])
from keras import losses
loss =loss_weights_1*losses.binary_crossentropy(targets1,output_binary) + \
loss_weights_2*losses.categorical_crossentropy(targets2,output_identity)+ \
loss_weights_3*losses.categorical_crossentropy(targets3,output_expression)
discriminator.add_loss(loss)
discriminator.compile( optimizer=dopt, loss = None)
discriminator.summary()
print (discriminator.metrics_names)
#plot_model(discriminator, to_file = '/media/vivo/New Volume/FERG_DB_256/stats/disc_0605_model.png')
def make_trainable(net, val):
net.trainable = val
for l in net.layers:
l.trainable = val
make_trainable(discriminator, False)
discriminator.trainable = False
# #### Build GAN model ####
z_dim = 128
units = 256
GANloss_weights_vae = Input(shape = (1,))
GANtargets_vae = Input(shape = (z_dim*2,))
#ee = 100
auxiliary_c = Input(shape=(c_dim,), name='aux_input_c')
encoder = model_encoder(z_dim=z_dim, input_shape=(img_rows, img_cols, 3), units=units, dropout=0.3)
# encoder.load_weights('/media/vivo/New Volume/FERG_DB_256/model/VAEGAN_5th_encoder_MUG_8pp_real_' + str(date)+'epochs'+str(ee)+ '.h5')
encoder.compile(loss='binary_crossentropy', optimizer=opt)
encoder.summary()
decoder = model_decoder(z_dim=z_dim, c_dim=c_dim)
# decoder.load_weights('/media/vivo/New Volume/FERG_DB_256/model/VAEGAN_5th_decoder_MUG_8pp_real_' + str(date)+'epochs'+str(ee)+ '.h5')
decoder.compile(loss='binary_crossentropy', optimizer=opt)
decoder.summary()
### Generate Image set ###
# generate_dataset(ee=ee)
###
### GAN formulation ###
[z, mean_var] = encoder(d_input)
xpred = decoder([z, auxiliary_c])
output_binary, output_identity, output_expression = discriminator([xpred, loss_weights_1, loss_weights_2,loss_weights_3, targets1, targets2, targets3])
GAN = Model([d_input, auxiliary_c, GANloss_weights_vae, loss_weights_1,loss_weights_2,loss_weights_3, GANtargets_vae, targets1, targets2, targets3],\
[mean_var, output_binary, output_identity, output_expression])
GANloss = GANloss_weights_vae*KL_loss(GANtargets_vae, mean_var) + \
loss_weights_1*losses.binary_crossentropy(targets1,output_binary) + \
loss_weights_2*losses.categorical_crossentropy(targets2, output_identity)+ \
loss_weights_3*losses.categorical_crossentropy(targets3, output_expression)
GAN.add_loss(GANloss)
GAN.compile(optimizer = opt, loss = None)
GAN.summary()
print (GAN.metrics_names)
# plot_model(GAN, to_file = 'GAN_model.png')
def plotGeneratedImages(epoch, idx=0, examples=10, dim=(10, 10), figsize=(10, 10)):
n = num_pp # how many digits we will display
pp_avg = 4500
plt.figure(figsize=(16, 4))
sample = x_ori[idx:idx + n, :, :, :]
c = np.asarray([0, 1, 2, 3, 4, 5, 6, 7])
c = keras.utils.to_categorical(c, num_pp)
[z, mean_var] = encoder.predict(sample)
generated_images = decoder.predict([z, c])
for i in range(n):
# display original
ax = plt.subplot(2, n, i + 1)
ori = sample[i].reshape(img_rows, img_cols, 3)
ori = np.uint8(ori * 127.5 + 127.5)
plt.imshow(ori)
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
# display reconstruction
ax = plt.subplot(2, n, i + 1 + n)
rec = generated_images[i].reshape(img_rows, img_cols, 3)
rec = np.uint8(rec * 127.5 + 127.5)
plt.imshow(rec)
ax.get_xaxis().set_visible(False)
ax.get_yaxis().set_visible(False)
# Path to be created
plt.savefig(path + '/GAN_MUG_results_' + str(date) + '_generated_image_epoch_%d.tif' % epoch)
plt.close()
def train_for_n(nb_epoch=50000, plt_frq=25, BATCH_SIZE=256):
batchCount = x_train.shape[0] / BATCH_SIZE
for ee in xrange(1, nb_epoch + 1):
print '-' * 15, 'Epoch %d' % ee, '-' * 15
plotGeneratedImages(epoch=ee + 40, idx=75)
# val_bin_acc, val_id_acc, val_ep_acc = val_test()
for e in tqdm(range(batchCount)):
# for didx in xrange(0,k):
idx = random.sample(range(0, x_train.shape[0]),
BATCH_SIZE) # train discriminator twice more than the generator
image_batch = x_train[idx, :, :, :] # real data
c = np.random.randint(num_pp, size=BATCH_SIZE)
c = keras.utils.to_categorical(c, num_pp)
[z, mean_var] = encoder.predict(image_batch)
generated_images = decoder.predict([z, c])
y1_batch = y_train1[idx, :]
y2_batch = y_train2[idx, :]
# generated_images = generator.predict([image_batch, c_, z])
y0_dist_real = np.random.uniform(0.9, 1.0, size=[BATCH_SIZE, 1])
y0_dist_fake = np.random.uniform(0, 0.1, size=[BATCH_SIZE, 1])
make_trainable(discriminator, True)
discriminator.trainable = True
loss_weights_1 = np.ones(shape = (batch_size,))*1/4.0
loss_weights_2 = np.ones(shape = (batch_size,))*1/2.0
loss_weights_3 = np.ones(shape = (batch_size,))*1/4.0
d_loss_real = discriminator.train_on_batch([image_batch, loss_weights_1, loss_weights_2, loss_weights_3,y0_dist_real, y1_batch, y2_batch],y= None)
loss_weights_1 = np.ones(shape=(batch_size,))*1.0
loss_weights_2 = np.ones(shape=(batch_size,)) * 0
loss_weights_3 = np.ones(shape=(batch_size,)) * 0
d_loss_fake = discriminator.train_on_batch([generated_images,loss_weights_1,loss_weights_2,loss_weights_3, y0_dist_fake, c, y2_batch], y = None)
make_trainable(discriminator, False)
discriminator.trainable = False
for ii in xrange(0, 2):
idx = random.sample(range(0, x_train.shape[0]),
BATCH_SIZE) # train discriminator twice more than the generator
image_batch = x_train[idx, :, :, :] # real data
c = np.random.randint(num_pp, size=BATCH_SIZE)
c = keras.utils.to_categorical(c, num_pp)
mean_var_ref = np.ones((BATCH_SIZE, z_dim * 2))
y1_batch = y_train1[idx, :]
y2_batch = y_train2[idx, :]
y0_batch = np.ones((BATCH_SIZE, 1)) #0.002, 0.09, 0.8, 0.108
GANloss_weights_vae = np.ones(shape = (batch_size,))*0.002
loss_weights_1 = np.ones(shape = (batch_size,))*0.078
loss_weights_2 = np.ones(shape = (batch_size,))*0.8
loss_weights_3 = np.ones(shape = (batch_size,))*0.12
g_loss = GAN.train_on_batch([image_batch, c, GANloss_weights_vae, loss_weights_1, loss_weights_2, loss_weights_3, mean_var_ref, y0_batch, c, y2_batch], y = None)
if ee % 25 == 0:
GAN.save('/VAEGAN_real_' + str(date) + 'epochs' + str(
ee) + '.h5')
encoder.save('/encoder_MUG_8pp_real_' + str(
date) + 'epochs' + str(ee) + '.h5')
decoder.save('/decoder_MUG_8pp_real_' + str(
date) + 'epochs' + str(ee) + '.h5')
discriminator.save('/discriminator_MUG_8pp_real_' + str(
date) + 'epochs' + str(ee) + '.h5')
start_time = time.time()
path = "/path_" + str(date)
if os.path.isdir(path) == False:
os.mkdir(path);
train_for_n(nb_epoch=epochs, plt_frq=500, BATCH_SIZE=batch_size)
process_time = time.time() - start_time
print("Elapsed: %s " % (process_time))