-
Notifications
You must be signed in to change notification settings - Fork 7
/
model.py
430 lines (346 loc) · 13.3 KB
/
model.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
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
import tensorflow as tf
import numpy as np
import math
import time
c = - 0.5 * math.log(2 * math.pi)
def binary_crossentropy(t, o, eps=1e-8):
return -(t * tf.log(o + eps) + (1.0 - t) * tf.log(1.0 - o + eps))
def log_normal2(x, mean, log_var, eps=1e-5):
return c - log_var / 2 - tf.pow(x - mean, 2) / (2 * tf.exp(log_var) + eps)
def kl_divergence(mean1, log_var1, mean2, log_var2, eps=1e-8):
mean_term = 0.5 * (tf.exp(log_var1) + tf.pow(mean1 - mean2, 2)) \
/ (tf.exp(log_var2) + eps)
return mean_term + 0.5 * log_var2 - 0.5 * log_var1 - 0.5
class Model(object):
def __init__(self, config):
self.config = config
self.learning_rate = config['learning_rate']
self.adv_learning_rate = config['adv_learning_rate']
self.batch_size = config['batch_size']
self.A, self.B, self.n_chan = config['input_shape']
self.img_size = self.A * self.B * self.n_chan
self.z_size = config['z_size']
self.is_gaussian = config['is_gaussian']
self.is_attacking = config['is_attacking']
self.mean = config['data_mean']
self.std = config['data_std']
self.max_pixel = config['data_max_pixel']
# Input placeholders
self.x = tf.placeholder(
tf.float32,
shape=(self.batch_size, self.A, self.B, self.n_chan),
name='x'
)
# Adversarial noise
self.noise = tf.Variable(
tf.zeros([self.A, self.B, self.n_chan]),
trainable=False,
name='noise'
)
# Placeholders for adversarial attack
self.noise_placeholder = tf.placeholder(
tf.float32,
shape=self.noise.get_shape(),
name='noise_placeholder'
)
self.C = tf.placeholder(dtype=tf.float32, shape=[], name='C')
self.adv_target_mean = tf.placeholder(
tf.float32,
[self.z_size],
name=('adv_target_mean')
)
self.adv_target_log_var = tf.placeholder(
tf.float32,
[self.z_size],
name=('adv_target_log_var')
)
self.adv_target_output = tf.placeholder(
tf.float32,
shape=(self.A, self.B, self.n_chan),
name='adv_target_output'
)
# Operation to set noise and reset noise
self.op_update_noise = self.noise.assign(self.noise_placeholder)
self.op_reset_noise = self.noise.assign(
tf.zeros([self.A, self.B, self.n_chan])
)
# Build graph
self.x_in = self.x
if self.is_attacking:
self.x_in += self.noise
self.x_in = tf.maximum(self.x_in, -self.mean / self.std)
self.x_in = tf.minimum(
self.x_in, (self.max_pixel - self.mean) / self.std)
self.x_in = tf.reshape(self.x_in, [self.batch_size, -1])
self._create_network()
self._create_loss_optimizer()
self._create_attack_optimizer()
self._create_output_attack_optimizer()
# Initializing the tensor flow variables
init = tf.global_variables_initializer()
# Launch the session
self.sess = tf.InteractiveSession()
self.sess.run(init)
# Model saver
self._saver = tf.train.Saver()
def _create_network(self):
pass
def _compute_latent_loss(self):
Lz = tf.reduce_sum(
kl_divergence(
self.z_mean,
self.z_log_var,
0.,
1.),
axis=1
)
return tf.reduce_mean(Lz)
def _compute_reconstr_loss(self):
flat_x_in = tf.reshape(self.x_in, [self.batch_size, -1])
if self.is_gaussian:
Lx = tf.reduce_sum(
-log_normal2(
flat_x_in,
self.x_reconstr_mean,
self.x_reconstr_log_var,
),
axis=1
)
else:
Lx = tf.reduce_sum(
binary_crossentropy(flat_x_in, self.x_reconstr_mean),
1
)
return tf.reduce_mean(Lx)
def _create_loss_optimizer(self):
self.reconstr_Lx = self._compute_reconstr_loss()
self.reconstr_Lz = self._compute_latent_loss()
self.reconstr_loss = self.reconstr_Lx + self.reconstr_Lz
def ClipIfNotNone(grad):
if grad is None:
return grad
return tf.clip_by_value(grad, -1, 1)
params = tf.trainable_variables()
grads = tf.gradients(self.reconstr_loss, params)
grads, _ = tf.clip_by_global_norm(grads, 3.)
self.reconstr_grads = [ClipIfNotNone(grad) for grad in grads]
self.recontr_optimizer = tf.train.AdamOptimizer(self.learning_rate)
self.reconstr_update = self.recontr_optimizer.apply_gradients(
zip(self.reconstr_grads, params)
)
def _create_attack_optimizer(self):
self.adversarial_loss = self.C * tf.reduce_sum(self.noise * self.noise)
self.adversarial_loss += (
tf.reduce_sum(
kl_divergence(
self.z_mean[0],
self.z_log_var[0],
self.adv_target_mean,
self.adv_target_log_var
)
)
)
adv_grads = tf.gradients(self.adversarial_loss, [self.noise])
adv_grads, _ = tf.clip_by_global_norm(adv_grads, 3)
self.adv_grad = tf.clip_by_value(adv_grads[0], -1, 1)
self.adv_optimizer = tf.train.AdamOptimizer(self.adv_learning_rate)
self.adv_update = self.adv_optimizer.apply_gradients(
zip([self.adv_grad], [self.noise])
)
def _create_output_attack_optimizer(self):
diff = self.adv_target_output - tf.reshape(
self.x_reconstr_mean[0],
shape=(self.A, self.B, self.n_chan)
)
self.adversarial_loss_output = (
self.C *
tf.reduce_sum(self.noise * self.noise)
)
self.adversarial_loss_output += tf.reduce_sum(diff * diff)
adv_grads = tf.gradients(self.adversarial_loss_output, [self.noise])
adv_grads, _ = tf.clip_by_global_norm(adv_grads, 3)
self.adv_grad_output = tf.clip_by_value(adv_grads[0], -1, 1)
self.adv_output_optimizer = tf.train.AdamOptimizer(
self.adv_learning_rate
)
self.adv_output_update = self.adv_output_optimizer.apply_gradients(
zip([self.adv_grad_output], [self.noise])
)
def batch_evaluate(self, X):
assert X.shape[0] == self.batch_size
loss = self.sess.run(
self.reconstr_loss,
feed_dict={self.x: X}
)
return loss
def batch_transform(self, X):
assert X.shape[0] == self.batch_size
z_mean, z_log_var = self.sess.run(
(self.z_mean, self.z_log_var),
feed_dict={self.x: X}
)
return z_mean, z_log_var
def batch_evaluate_attack(self, X, adv_target_mean=None,
adv_target_log_var=None, C=0,
adv_target_output=None):
assert X.shape[0] == self.batch_size
assert (
(adv_target_mean is not None and adv_target_log_var is not None) or
adv_target_output is not None
)
if adv_target_output is None:
loss, grad = self.sess.run(
(self.adversarial_loss, self.adv_grad),
feed_dict={
self.x: X,
self.adv_target_mean: adv_target_mean,
self.adv_target_log_var: adv_target_log_var,
self.C: C
}
)
else:
loss, grad = self.sess.run(
(self.adversarial_loss_output, self.adv_grad_output),
feed_dict={
self.x: X,
self.adv_target_output: adv_target_output,
self.C: C
}
)
return loss, grad
def batch_optimize_attack(self, X,
adv_target_mean=None,
adv_target_log_var=None,
C=0,
adv_target_output=None,
num_iter=50):
assert X.shape[0] == self.batch_size
assert (
(adv_target_mean is not None and adv_target_log_var is not None) or
adv_target_output is not None
)
losses = []
best_loss = float('Inf')
init_noise = np.random.uniform(
-1e-8,
1e-8,
size=self.config['input_shape']
).astype(np.float32)
self.set_noise(init_noise)
for i in range(num_iter):
if adv_target_output is None:
_, loss = self.sess.run(
(self.adv_update, self.adversarial_loss),
feed_dict={
self.x: X,
self.adv_target_mean: adv_target_mean,
self.adv_target_log_var: adv_target_log_var,
self.C: C
}
)
else:
_, loss = self.sess.run(
(self.adv_output_update, self.adversarial_loss_output),
feed_dict={
self.x: X,
self.adv_target_output: adv_target_output,
self.C: C
}
)
losses.append(loss)
if best_loss > loss:
best_loss = loss
best_noise = self.sess.run(self.noise)
return best_noise, losses
def batch_fit(self, X):
assert X.shape[0] == self.batch_size
_, loss = self.sess.run(
(self.reconstr_update, self.reconstr_loss),
feed_dict={self.x: X}
)
return loss
def batch_reconstruct(self, X):
assert X.shape[0] <= self.batch_size
diff = None
if X.shape[0] < self.batch_size:
diff = self.batch_size - X.shape[0]
X = np.vstack((X, np.zeros((diff,) + X.shape[1:])))
X_reconstr = self.sess.run(
self.x_reconstr_mean,
feed_dict={self.x: X}
)
if diff is not None:
X_reconstr = X_reconstr[:diff]
return X_reconstr
def reset_noise(self):
self.sess.run(self.op_reset_noise)
def set_noise(self, noise):
assert noise.shape == tuple((self.A, self.B, self.n_chan))
self.sess.run(
self.op_update_noise,
feed_dict={self.noise_placeholder: noise}
)
def evaluate(self, X):
num_samples = X.shape[0]
num_batches = math.floor(num_samples / self.batch_size)
total_loss = 0
for idx in range(0, num_samples, self.batch_size):
batch_x = X[idx:(idx + self.batch_size)]
if batch_x.shape[0] < self.batch_size:
continue
total_loss += self.batch_evaluate(batch_x)
return total_loss / num_batches
def fit(self, X, X_val, num_epochs, ckpt_path, random_sampling=False,
img_path=None, X_test=None, f_reconstruct=None, epoch_rec=10,
verbose=True):
num_samples = X.shape[0]
num_batches = math.ceil(num_samples / self.batch_size)
losses = []
val_losses = []
best_val = float('inf')
self.reset_noise()
for epoch in range(num_epochs):
tic = time.time()
total_loss = 0
for i in range(num_batches):
if random_sampling:
idx = np.random.choice(X.shape[0], self.batch_size)
else:
idx = list(
range(
i * self.batch_size,
min((i + 1) * self.batch_size, X.shape[0])
)
)
if len(idx) < self.batch_size:
idx += np.random.choice(
X.shape[0],
self.batch_size - len(idx)
).tolist()
continue
total_loss += self.batch_fit(X[idx])
val_losses.append(self.evaluate(X_val))
if val_losses[-1] < best_val:
best_val = val_losses[-1]
self.save(ckpt_path + ("/model.epoch=%d.val_loss=%.4f.ckpt"
% (epoch + 1, best_val)))
if (epoch % epoch_rec == 0 and X_test is not None and
img_path is not None and f_reconstruct is not None):
f_reconstruct(
self,
X_test,
img_path + ('/rec.epoch=%d.png' % (epoch + 1)),
normalize=self.is_gaussian
)
losses.append(total_loss / num_batches)
if verbose:
print('Epoch %d/%d: loss=%f, elapsed=%.4fs' %
(epoch + 1, num_epochs, losses[-1], time.time() - tic))
print('\tval_loss=%f' % val_losses[-1])
return losses, val_losses
def load(self, ckpt_path):
self._saver.restore(self.sess, ckpt_path)
def save(self, ckpt_path):
self._saver.save(self.sess, ckpt_path)
def close(self):
self.sess.close()