-
Notifications
You must be signed in to change notification settings - Fork 4
/
CRBM.py
345 lines (289 loc) · 14.6 KB
/
CRBM.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
import tensorflow as tf
import numpy as np
import scipy.stats as stats
import logging
class CRBM():
'''
CRBM based on a paper 'Unsupervised feature learning for audio classification using convolutional
deep belief networks' by Honglak Lee, Yan Largman, Peter Pham, and Andrew Y, Ng.
'''
def __init__(self, filter_shape, visible_shape, k, params_id, stddev=1.0, binary=False):
'''
:param filter_shape: tuple of integer (f_h, f_w). shape of filter. f_h, f_w are height and width respectively.
:param visible_shape: tuple of integer (v_h, v_w). shape of visibles. v_h, v_w are height and width respectively.
:param k: integer. also called group, number of filters.
:param params_id: str. unique identifier for parameters.
:param stddev: float. standard deviation during parameters initialization
:param binary: bool. if visible units are binary or not. False means they are real values.
'''
assert filter_shape[0] <= visible_shape[0]
assert filter_shape[1] <= visible_shape[1]
if binary:
raise Exception('not implemented yet')
graph = tf.Graph()
with graph.as_default() as g:
with tf.variable_scope(params_id) as crbm_scope:
self.w = tf.get_variable('weights', shape=(k,) + filter_shape,
initializer=tf.random_normal_initializer(mean=0.0, stddev=stddev))
self.w_r = self.w[:,:,::-1]
self.hb = tf.get_variable('hidden_biases', shape=(k,),
initializer=tf.random_normal_initializer(mean=0.0, stddev=stddev))
self.vb = tf.get_variable('visible_biases', shape=(1,),
initializer=tf.random_normal_initializer(mean=0.0, stddev=stddev))
self.sess = tf.Session(graph=graph)
# initialize parameters
self.sess.run(tf.global_variables_initializer())
self.hidden_shape = (visible_shape[0]-filter_shape[0]+1, visible_shape[1]-filter_shape[1]+1)
self.filter_shape = filter_shape
self.visible_shape = visible_shape
self.k = k
self.params_id = params_id
self.binary = binary
self.graph = graph
def generate_hidden_units_probabilities(self, visible):
'''
generate probabilities of hidden units being 1, given visible units
:param v: numpy array of shape (b,) + self.visible_shape. b is batch size.
:return: numpy array of shape (b,) + self.k + self.hidden_shape. b is batch size. each element is between [0,1]
'''
assert visible.shape[1:] == self.visible_shape
w = self.w
hb = self.hb
sess = self.sess
graph = self.graph
with graph.as_default() as g:
visible = tf.convert_to_tensor(visible)
# fit data.
visible = tf.expand_dims(visible, axis=-1)
w = tf.expand_dims(w, axis=0)
w = tf.transpose(w, perm=[2,3,0,1]) # (h, w, 1, k)
tf_convolution = tf.nn.conv2d(visible, w, [1,1,1,1], 'VALID') + hb
tf_convolution = tf.transpose(tf_convolution, perm=[0,3,1,2]) # (b, k, h, w)
convoluted = sess.run(tf_convolution)
probabilities = 1.0/(1+np.power(np.e, -convoluted))
return probabilities
def generate_hidden_units(self, visible):
'''
generate hidden units (1s and 0s), given visible units.
:param v: numpy array of shape (b,) + self.visible_shape. visible units. b is batch size.
:return: numpy array of shape (b,) + self.k + self.hidden_shape. b is batch size. each element is between [0,1]
'''
assert visible.shape[1:] == self.visible_shape
hidden_shape = self.hidden_shape
k = self.k
batch_size = visible.shape[0]
probabilities = self.generate_hidden_units_probabilities(visible)
# efficient sampling
random_0to1 = np.random.rand(batch_size,k,hidden_shape[0],hidden_shape[1])
eval_func = np.vectorize(lambda d: 1 if d > 0 else 0)
hidden_units = eval_func(probabilities - random_0to1)
return hidden_units
def generate_visible_units_expectations(self, hidden):
'''
generate expectations(mean) of visible units.
:param hidden: numpy array of shape (b,) + self.k + self.hidden_shape.
:return: numpy array of shape (b,) + self.visible_shape.
'''
assert hidden.shape[1:] == (self.k,) + self.hidden_shape
if self.binary:
raise Exception('not implemented yet')
filter_shape = self.filter_shape
w_r = self.w_r
vb = self.vb
sess = self.sess
graph = self.graph
filter_height = filter_shape[0]
filter_width = filter_shape[1]
with graph.as_default() as g:
hidden = tf.convert_to_tensor(hidden, dtype=tf.float32)
w_r = tf.expand_dims(w_r, axis=0)
w_r = tf.transpose(w_r, perm=[2,3,1,0]) # (h, w, k, 1)
padding = tf.convert_to_tensor([[0,0],[filter_height-1,filter_height-1],
[filter_width-1,filter_width-1],[0,0]])
hidden = tf.transpose(hidden, perm=[0,2,3,1])
hidden = tf.pad(hidden, padding) # (b, h, w, k)
tf_expectation = tf.nn.conv2d(hidden, w_r, [1,1,1,1], 'VALID') + vb # (b, h, w, 1)
expectations = sess.run(tf.squeeze(tf_expectation, axis=[-1]))
return expectations
def generate_visible_units(self, hidden, sigma=1):
'''
generate visible units.
:param hidden: numpy array of shape (b,) + self.k + self.hidden_shape.
:param sigma: float. variance of generating visible units.
:return: numpy array of shape (b,) + self.visible_shape.
'''
assert hidden.shape[1:] == (self.k,) + self.hidden_shape
if self.binary:
raise Exception('not implemented yet')
expectations = self.generate_visible_units_expectations(hidden)
sample_func = np.vectorize(lambda ep : stats.norm.rvs(ep,sigma))
visible = sample_func(expectations)
return visible
class Trainer:
'''
Trainer to CRBM using constrastive divergence(CD).
'''
def __init__(self, crbm, summary_enabled=False, summary_dir=None, summary_flush_secs=60, init_step=0):
'''
an instance of CRBM to be trained.
:param crbm: CRBM.
:param summary_enabled: bool. enable summary
:param summary_dir: str. when @summary_enable, this is where summary will be saved.
:param init_step: int. step to begin. for resuming from previous training, set it to the previous step + 1.
'''
if summary_enabled:
assert summary_dir != None
filter_shape = crbm.filter_shape
visible_shape = crbm.visible_shape
hidden_shape = crbm.hidden_shape
k = crbm.k
w = crbm.w
hb = crbm.hb
vb = crbm.vb
graph = crbm.graph
tf_dtype = tf.float32
summaries, summary_file = None, None
with graph.as_default() as g:
w = tf.expand_dims(w, axis=0)
w = tf.transpose(w, perm=[2,3,0,1]) # (h, w, 1, k)
hidden_in = tf.placeholder(tf_dtype, shape=[None, k, hidden_shape[0],hidden_shape[1]])
visible_in = tf.placeholder(tf_dtype, shape=[None, visible_shape[0], visible_shape[1]])
hidden_in_fitted = tf.transpose(hidden_in, perm=[0,2,3,1]) # (b, h, w, k)
visible_in_fitted = tf.expand_dims(visible_in, axis=-1) # (b, h, w, 1)
convolution = tf.nn.conv2d(visible_in_fitted, w, [1,1,1,1], 'VALID')
energy = -tf.reduce_sum(hidden_in_fitted*convolution, axis=[1,2,3]) \
- tf.reduce_sum(hb*tf.reduce_sum(hidden_in_fitted, axis=[1,2]), axis=[1]) \
- vb*tf.reduce_sum(visible_in_fitted, axis=[1,2,3])
hidden_rec_in = tf.placeholder(tf_dtype, shape=[None, k, hidden_shape[0],hidden_shape[1]])
visible_rec_in = tf.placeholder(tf_dtype, shape=[None, visible_shape[0], visible_shape[1]])
hidden_rec_in_fitted = tf.transpose(hidden_rec_in, perm=[0,2,3,1])
visible_rec_in_fitted = tf.expand_dims(visible_rec_in, axis=-1)
convolution_rec = tf.nn.conv2d(visible_rec_in_fitted, w, [1,1,1,1], 'VALID')
energy_rec = -tf.reduce_sum(hidden_rec_in_fitted*convolution_rec, axis=[1,2,3]) \
- tf.reduce_sum(hb*tf.reduce_sum(hidden_rec_in_fitted, axis=[1,2]), axis=[1]) \
- vb*tf.reduce_sum(visible_rec_in_fitted, axis=[1,2,3])
probabilities = tf.reduce_mean(tf.nn.sigmoid(convolution + hb)) # it is also the regularization term.
loss = tf.reduce_mean(energy - energy_rec)
learning_rate_in = tf.placeholder(tf_dtype)
optimizer = tf.train.GradientDescentOptimizer(learning_rate_in)
energy_mean = tf.reduce_mean(energy)
minimize_energy = optimizer.minimize(energy_mean)
energy_rec_mean = tf.reduce_mean(energy_rec)
maximize_energy_rec = optimizer.minimize(-energy_rec_mean)
if summary_enabled:
summary_file = tf.summary.FileWriter(summary_dir, graph, flush_secs=summary_flush_secs)
tf.summary.scalar('loss', loss)
tf.summary.scalar('probability', probabilities)
tf.summary.scalar('real_energy', energy_mean)
tf.summary.scalar('reconstructed_energy', energy_rec_mean)
w_gradient = tf.gradients(loss, w)
tf.summary.histogram('weights_gradient', w_gradient)
hb_gradient = tf.gradients(loss, hb)
tf.summary.histogram('hidden_biases_gradient', hb_gradient)
vb_gradient = tf.gradients(loss, vb)
tf.summary.histogram('visible_biases_gradient', vb_gradient)
summaries = tf.summary.merge_all()
self.crbm = crbm
self.probabilities = probabilities
self.loss = loss
self.energy_mean = energy_mean
self.energy_rec_mean = energy_rec_mean
self.minimize_energy = minimize_energy
self.maximize_energy_rec = maximize_energy_rec
self.hidden_in = hidden_in
self.visible_in = visible_in
self.hidden_rec_in = hidden_rec_in
self.visible_rec_in = visible_rec_in
self.learning_rate_in = learning_rate_in
self.summary_enabled = summary_enabled
self.summaries = summaries
self.summary_file = summary_file
self.step = init_step
self.logger = logging.getLogger()
def train(self, visible_units, gibbs=1, sigma=1.0, lr=0.0001):
'''
one iteration of training one self.crbm
:param visible_units: numpy array of shape, (b, ) + crbm.visible_shape
:param gibbs: int. gibb step. has to be greater than 1
:param sigma: float. noise when reconstructing.
:param verbose: bool.
:param lr: float. learning rate
:return: None.
'''
assert gibbs >= 1
probabilities = self.probabilities
crbm = self.crbm
loss = self.loss
energy_mean = self.energy_mean
energy_rec_mean = self.energy_rec_mean
minimize_energy = self.minimize_energy
maximize_energy_rec = self.maximize_energy_rec
hidden_in = self.hidden_in
visible_in = self.visible_in
hidden_rec_in = self.hidden_rec_in
visible_rec_in = self.visible_rec_in
learning_rate_in = self.learning_rate_in
summary_enabled = self.summary_enabled
summaries = self.summaries
summary_file = self.summary_file
logger = self.logger
sess = crbm.sess
hidden = crbm.generate_hidden_units(visible_units)
# reconstruct
hidden_rec = hidden
visible_rec = crbm.generate_visible_units(hidden_rec, sigma=sigma)
for _ in range(gibbs-1):
hidden_rec = crbm.generate_hidden_units(visible_rec)
visible_rec = crbm.generate_visible_units(hidden_rec, sigma=sigma)
# optimization
feed_dict = {
hidden_in: hidden,
visible_in: visible_units,
hidden_rec_in: hidden_rec,
visible_rec_in: visible_rec,
learning_rate_in: lr
}
sess.run(minimize_energy, feed_dict=feed_dict)
sess.run(maximize_energy_rec, feed_dict=feed_dict)
# logging.
log_msg = '''
error: {0}
real energy: {1}
reconstructed energy: {2}
aggregated hidden units activation probability: {3}
at step {4}
'''.format(sess.run(loss, feed_dict=feed_dict), sess.run(energy_mean, feed_dict=feed_dict),
sess.run(energy_rec_mean, feed_dict=feed_dict), sess.run(probabilities, feed_dict=feed_dict),
self.step)
logger.debug(log_msg)
if summary_enabled:
summary_file.add_summary(sess.run(summaries, feed_dict=feed_dict), global_step=self.step)
self.step += 1
return feed_dict
class Saver:
@staticmethod
def save(crbm, path, step):
'''
(path='my-model', step=0) ==> filename: 'my-model-0'
:param crbm: an instance of crbm. to be saved
:param path: str. path to save.
:param step: int.
:return:
'''
w = crbm.w
w_r = crbm.w_r
vb = crbm.vb
hb = crbm.hb
sess = crbm.sess
with crbm.graph.as_default() as g:
saver = tf.train.Saver(var_list=[w, vb, hb])
saver.save(sess, path, global_step=step, write_meta_graph=False)
@staticmethod
def restore(crbm, path):
with crbm.graph.as_default() as g:
sess = crbm.sess
tf.train.Saver().restore(sess, path)
# reset w_r from w. this step might not be needed.
with crbm.graph.as_default() as g:
with tf.variable_scope(crbm.params_id) as crbm_scope:
crbm.w_r = crbm.w[:,:,::-1]