-
Notifications
You must be signed in to change notification settings - Fork 14
/
Copy pathmodel.py
386 lines (323 loc) · 15.6 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
import tensorflow as tf
import logging
class IterE():
"""
This class build a computation graph that represents the
axiom embedding model
"""
def __init__(self, option, data):
self.data = data
self.dim = option.dim
self.option = option
self.model = option.model
self.axiom_weight = option.axiom_weight
self.device = option.device
self.clip_val = 10.0
self._build_graph()
def _build_graph(self):
self._build_loss_graph()
self._build_axiom_probality_graph()
self._build_test_graph()
def _build_loss_graph(self):
self._build_loss_input()
# loss for embeddings
self.pos_score = pos_score = self.triple_score(self.pos_hrt)
self.neg_score = neg_score = self.triple_score(self.neg_hrt)
self.pos_score_sig = tf.sigmoid(pos_score)
self.neg_score_sig = tf.sigmoid(neg_score)
self.neg_score_sig2 = tf.sigmoid(-neg_score)
if self.option.loss_type == "margin-based":
self.loss_embedding = tf.reduce_sum(tf.maximum(0., neg_score+self.option.margin - pos_score))
elif self.option.loss_type == 'log-likelihood':
self.loss_embedding = tf.reduce_sum(tf.nn.softplus(tf.concat([neg_score, -pos_score],0)))
elif self.option.loss_type == 'ANALOGY':
triples = tf.reshape(tf.concat([-neg_score, pos_score],0), [-1])
labels = tf.reshape(tf.concat([self.neg_labels, self.pos_labels], 0), [-1])
self.loss_embedding = tf.reduce_mean(-tf.log(tf.clip_by_value(tf.sigmoid(triples), 1e-32, 1.0)) *labels)
else:
raise NotImplementedError("Does not support %s loss"%self.option.loss_type)
self.loss_regularizer = 0
for var in [self.pos_h_embedding,self.pos_r_embedding, self.pos_t_embedding,
self.neg_h_embedding, self.neg_r_embedding, self.neg_t_embedding]:
if self.option.regularizer_type == 'L1':
self.loss_regularizer += tf.reduce_mean(tf.abs(var))
elif self.option.regularizer_type =='L2':
self.loss_regularizer += tf.reduce_mean(tf.square(var))
else:
raise NotImplementedError
self.loss_regularizer *= self.option.regularizer_weight
with tf.device(self.device):
self.loss = self.loss_embedding \
+ self.loss_regularizer \
if self.option.optimizer == 'Adam':
self.optimizer = tf.train.AdamOptimizer(self.learning_rate)
elif self.option.optimizer == 'Adagrad':
self.optimizer = tf.train.AdagradOptimizer(self.learning_rate)
elif self.option.optimizer == 'Gradientdescent':
self.optimizer = tf.train.GradientDescentOptimizer(self.learning_rate)
elif self.option.optimizer == 'AdagradDA':
self.optimizer = tf.train.AdagradDAOptimizer(self.learning_rate)
elif self.option.optimizer == 'RMSProp':
self.optimizer = tf.train.RMSPropOptimizer(self.learning_rate)
else:
raise NotImplementedError
self.grads_and_vars = self.optimizer.compute_gradients(self.loss, self.variables)
self.optimizer_step = self.optimizer.apply_gradients(self.grads_and_vars)
def _build_axiom_probality_graph(self):
self._build_axiom_probability_input()
with tf.device(self.device):
self.reflexive_probability = self.sim(head=self.reflexive_embed[:, 0, :],
tail=self.identity, arity=1)
self.symmetric_probability = self.sim(head=[self.symmetric_embed[:, 0, :],self.symmetric_embed[:, 0, :]],
tail=self.identity, arity=2)
self.transitive_probability = self.sim(head=[self.transitive_embed[:, 0, :],self.transitive_embed[:, 0, :]],
tail=self.transitive_embed[:, 0, :], arity=2)
self.inverse_probability1 = self.sim(head=[self.inverse_embed[:, 0,:],self.inverse_embed[:, 1,:]],
tail = self.identity, arity=2)
self.inverse_probability2 = self.sim(head=[self.inverse_embed[:,1,:],self.inverse_embed[:, 0,:]],
tail=self.identity, arity=2)
self.inverse_probability = (self.inverse_probability1+ self.inverse_probability2)/2
self.subproperty_probability = self.sim(head=self.subproperty_embed[:, 0,:],
tail=self.subproperty_embed[:, 1, :], arity=1)
self.equivalent_probability = self.sim(head=self.equivalent_embed[:, 0,:],
tail=self.equivalent_embed[:, 1,:], arity=1)
self.inferenceChain1_probability = self.sim(
head=[self.inferencechain_embed[:, 1, :], self.inferencechain_embed[:, 0, :]],
tail=self.inferencechain_embed[:, 2, :], arity=2)
self.inferenceChain2_probability = self.sim(
head=[self.inferencechain_embed[:, 2, :], self.inferencechain_embed[:, 1, :], self.inferencechain_embed[:, 0, :]],
tail=self.identity, arity=3)
self.inferenceChain3_probability = self.sim(
head=[self.inferencechain_embed[:, 1, :], self.inferencechain_embed[:, 2, :]],
tail=self.inferencechain_embed[:, 0, :], arity=2)
self.inferenceChain4_probability = self.sim(
head=[self.inferencechain_embed[:, 0, :], self.inferencechain_embed[:, 2, :]],
tail=self.inferencechain_embed[:, 1, :], arity=2)
def _build_test_graph(self):
self._build_test_input()
with tf.device(self.device):
self.test_score = self.triple_score(self.test_embed)
def _build_loss_input(self):
with tf.device('/cpu'):
# the input positive triples size should be [embedding_batchsize, 3]
# the input negative triples size should be [k*embedding_batchsize, 3]
self.pos_triples = tf.placeholder(tf.int32, [None, 3])
self.neg_triples = tf.placeholder(tf.int32, [None, 3])
self.pos_labels = tf.placeholder(tf.float32, [None, 1])
self.neg_labels = tf.placeholder(tf.float32, [None, 1])
self.learning_rate = tf.placeholder(tf.float32, [])
#bound = 6 / math.sqrt(self.dim)
bound = self.option.init_bound
self.entity_embeddings = tf.get_variable('ent_embed', [self.data.num_entity, self.dim],
initializer=tf.random_uniform_initializer(minval=-bound,
maxval=bound,
seed=123))
self.relation_embeddings = tf.get_variable('rel_embed', [self.data.num_relation, self.dim],
initializer=tf.random_uniform_initializer(minval=-bound,
maxval=bound,
seed=124))
self.variables = [self.entity_embeddings, self.relation_embeddings]
# embedding lookup for embeddings
# [None, 3, 20]
self.pos_hrt = self.embedding_lookup_triples(self.pos_triples, 1)
self.neg_hrt = self.embedding_lookup_triples(self.neg_triples, 1)
# normalize the entity embedding
entity_embeddings_norm = self.entity_embeddings
self.pos_h_embedding = tf.nn.embedding_lookup(entity_embeddings_norm, self.pos_triples[:, 0])
self.pos_t_embedding = tf.nn.embedding_lookup(entity_embeddings_norm, self.pos_triples[:, 2])
self.pos_r_embedding = tf.nn.embedding_lookup(entity_embeddings_norm, self.pos_triples[:, 1])
self.neg_h_embedding = tf.nn.embedding_lookup(entity_embeddings_norm, self.neg_triples[:, 0])
self.neg_t_embedding = tf.nn.embedding_lookup(entity_embeddings_norm, self.neg_triples[:, 2])
self.neg_r_embedding = tf.nn.embedding_lookup(entity_embeddings_norm, self.neg_triples[:, 1])
def _build_axiom_probability_input(self):
with tf.device('/cpu'):
self.reflexive_pool = tf.placeholder(tf.int32, [None, 1])
self.symmetric_pool = tf.placeholder(tf.int32, [None, 1])
self.transitive_pool = tf.placeholder(tf.int32, [None, 1])
self.inverse_pool = tf.placeholder(tf.int32, [None, 2])
self.subproperty_pool = tf.placeholder(tf.int32, [None,2])
self.equivalent_pool = tf.placeholder(tf.int32, [None,2])
self.inferenceChain_pool = tf.placeholder(tf.int32, [None,3])
# look up the embeddings for each axiom in axiom pool
# axiom_embed: [?, arity, dim]
self.reflexive_embed = tf.nn.embedding_lookup(self.relation_embeddings, self.reflexive_pool)
self.symmetric_embed = tf.nn.embedding_lookup(self.relation_embeddings, self.symmetric_pool)
self.transitive_embed = tf.nn.embedding_lookup(self.relation_embeddings, self.transitive_pool)
self.inverse_embed = tf.nn.embedding_lookup(self.relation_embeddings, self.inverse_pool)
self.subproperty_embed = tf.nn.embedding_lookup(self.relation_embeddings, self.subproperty_pool)
self.equivalent_embed = tf.nn.embedding_lookup(self.relation_embeddings, self.equivalent_pool)
self.inferencechain_embed = tf.nn.embedding_lookup(self.relation_embeddings, self.inferenceChain_pool)
self.identity = tf.expand_dims(tf.concat((tf.ones(self.dim-self.dim/4),tf.zeros(self.dim/4)),0),0)
def _build_test_input(self):
with tf.device('/cpu'):
# input test triples
self.input_test_triples = tf.placeholder(tf.int32, [None, 3])
# look up for embeddings of input triples
# test_embed including embeddings for h,r,t
self.test_embed = self.embedding_lookup_triples(self.input_test_triples, 1)
def embedding_lookup_triples(self, triples,num_triples):
output = None
for i in range(num_triples):
start_slice = i*3
# hrt: [?, 3, dim]
hrt = self.embedding_lookup_triple(triples, start_slice)
if i == 0:
output = hrt
else:
output = tf.concat([output, hrt], 1)
return output
def embedding_lookup_triple(self, triples, start_slice):
# h,r,t: [?, 1, dim]
# normalize the entity embeddings
#entity_embedding_norm = tf.nn.l2_normalize(self.entity_embeddings, dim=1)
entity_embedding_norm = self.entity_embeddings
h = tf.expand_dims(tf.nn.embedding_lookup(entity_embedding_norm, triples[:, start_slice + 0]), 1)
t = tf.expand_dims(tf.nn.embedding_lookup(entity_embedding_norm, triples[:, start_slice + 2]), 1)
r = tf.expand_dims(tf.nn.embedding_lookup(self.relation_embeddings, triples[:, start_slice + 1]), 1)
# outputembeddings: [?, 3, dim]
outputembeddings = tf.concat([h, r, t], 1)
return outputembeddings
def triple_score(self, triples):
# triples: [None, 3, dim]
# h,r,t: [None, 1, dim] -> [None, dim]
# [100 + 50(x) + 50(y)]
# x, y
# -y x
h,r,t = tf.split(triples, [1,1,1], 1)
h = tf.squeeze(h,1)
r = tf.squeeze(r,1)
t = tf.squeeze(t,1)
if self.model == 'ANALOGY':
# h_scalar: [None, dim/2]
# h_x, h_y: [None, dim/4]
h_scalar, h_x ,h_y = self.split_embedding(h)
r_scalar, r_x, r_y = self.split_embedding(r)
t_scalar, t_x, t_y = self.split_embedding(t)
# score_scalar: [None]
score_scalar = tf.reduce_sum(h_scalar * r_scalar * t_scalar, axis=1)
# score_block: [None]
score_block = tf.reduce_sum(h_x * r_x * t_x
+ h_x * r_y * t_y
+ h_y * r_x * t_y
- h_y * r_y * t_x, axis=1)
# score: [None]
score = score_scalar + score_block
return score
def split_embedding(self, embedding):
# embedding: [None, dim]
assert self.dim % 4 == 0
num_scalar = self.dim // 2
num_block = self.dim // 4
embedding_scalar = embedding[:, 0:num_scalar]
embedding_x = embedding[:, num_scalar:-num_block]
embedding_y = embedding[:, -num_block:]
return embedding_scalar, embedding_x, embedding_y
def axiom_loss(self, score, confidence, type):
if type == 1:
pi = score[0]
else:
pi_b = score[0]
if type==2:
pi_a = score[1]
elif type==3:
pi_a = score[1]*score[2]
else:
raise NotImplementedError
pi = pi_a * pi_b - pi_a +1
loss = tf.reduce_mean(tf.maximum(0.0, confidence-pi))
return loss
def axiom_loss_triple(self, input_score, confidence):
score = input_score[0]
loss = tf.reduce_mean(tf.clip_by_value(-confidence * tf.log(score), 1e-32, 1.0))
return loss
# calculate the similrity between two matrices
# head: [?, dim]
# tail: [?, dim] or [1,dim]
def sim(self, head=None, tail=None, arity=None):
if arity == 1:
A_scalar, A_x, A_y = self.split_embedding(head)
elif arity == 2:
M1_scalar, M1_x, M1_y = self.split_embedding(head[0])
M2_scalar, M2_x, M2_y = self.split_embedding(head[1])
A_scalar= M1_scalar * M2_scalar
A_x = M1_x*M2_x - M1_y*M2_y
A_y = M1_x*M2_y + M1_y*M2_x
elif arity==3:
M1_scalar, M1_x, M1_y = self.split_embedding(head[0])
M2_scalar, M2_x, M2_y = self.split_embedding(head[1])
M3_scalar, M3_x, M3_y = self.split_embedding(head[2])
M1M2_scalar = M1_scalar * M2_scalar
M1M2_x = M1_x * M2_x - M1_y * M2_y
M1M2_y = M1_x * M2_y + M1_y * M2_x
A_scalar = M1M2_scalar * M3_scalar
A_x = M1M2_x * M3_x - M1M2_y * M3_y
A_y = M1M2_x * M3_y + M1M2_y * M3_x
else:
raise NotImplemented
B_scala, B_x, B_y = self.split_embedding(tail)
similarity = tf.concat([(A_scalar - B_scala)**2, (A_x - B_x)**2, (A_x - B_x)**2, (A_y - B_y)**2, (A_y - B_y)**2 ], axis=1)
similarity = tf.sqrt(tf.reduce_sum(similarity, axis=1))
#recale the probability
probability = (tf.reduce_max(similarity)-similarity)/(tf.reduce_max(similarity)-tf.reduce_min(similarity))
return probability
# generate a probality for each axiom in axiom pool
def run_axiom_probability(self, sess, data):
if len(data.axiompool_reflexive) != 0: reflexive_prob = sess.run(self.reflexive_probability, {self.reflexive_pool: data.axiompool_reflexive})
else: reflexive_prob = []
if len(data.axiompool_symmetric) != 0: symmetric_prob = sess.run(self.symmetric_probability, {self.symmetric_pool: data.axiompool_symmetric})
else: symmetric_prob = []
if len(data.axiompool_transitive) != 0: transitive_prob = sess.run(self.transitive_probability, {self.transitive_pool: data.axiompool_transitive})
else: transitive_prob = []
if len(data.axiompool_inverse) != 0: inverse_prob = sess.run(self.inverse_probability, {self.inverse_pool: data.axiompool_inverse})
else: inverse_prob = []
if len(data.axiompool_subproperty) != 0: subproperty_prob = sess.run(self.subproperty_probability, {self.subproperty_pool: data.axiompool_subproperty})
else: subproperty_prob = []
if len(data.axiompool_equivalent) != 0: equivalent_prob = sess.run(self.equivalent_probability, {self.equivalent_pool: data.axiompool_equivalent})
else: equivalent_prob = []
if len(data.axiompool_inferencechain1) != 0:
inferencechain1_prob = sess.run(self.inferenceChain1_probability,
{self.inferenceChain_pool: data.axiompool_inferencechain1})
else:
inferencechain1_prob = []
if len(data.axiompool_inferencechain2) != 0:
inferencechain2_prob = sess.run(self.inferenceChain2_probability,
{self.inferenceChain_pool: data.axiompool_inferencechain2})
else:
inferencechain2_prob = []
if len(data.axiompool_inferencechain3) != 0:
inferencechain3_prob = sess.run(self.inferenceChain3_probability,
{self.inferenceChain_pool: data.axiompool_inferencechain3})
else:
inferencechain3_prob = []
if len(data.axiompool_inferencechain4) != 0:
inferencechain4_prob = sess.run(self.inferenceChain4_probability,
{self.inferenceChain_pool: data.axiompool_inferencechain4})
else:
inferencechain4_prob = []
output = [reflexive_prob, symmetric_prob, transitive_prob, inverse_prob,
subproperty_prob,equivalent_prob,inferencechain1_prob, inferencechain2_prob,
inferencechain3_prob, inferencechain4_prob]
return output
def run_train_batch(self, sess, feed):
_, loss, loss_reg, score_pos, score_neg, score_pos_sig, score_neg_sig, score_neg_sig2, grads_and_vars_clip,\
pos_embedding, neg_embeddding \
= sess.run([self.optimizer_step, self.loss,self.loss_regularizer,
self.pos_score, self.neg_score,
self.pos_score_sig, self.neg_score_sig, self.neg_score_sig2,self.grads_and_vars ,
self.pos_hrt, self.neg_hrt,
], feed_dict=feed)
# for debug
ent_embedding, rel_embedding = sess.run([self.entity_embeddings, self.relation_embeddings])
logging.info('score_pos:%s'%score_pos)
logging.info('score_neg:%s'%score_neg)
logging.info('score_pos_sig:%s'%score_pos_sig)
logging.info('score_neg_sig:%s'%score_neg_sig)
logging.info('score_neg_sig2(with -)%s'%score_neg_sig2)
logging.info('entity embedding:%s'%ent_embedding[:2, :10])
logging.info('relation embedding:%s'%rel_embedding[:2, :10])
logging.info('gradient:%s'%str(grads_and_vars_clip[0]))
logging.info('values:%s'%str(grads_and_vars_clip[1]))
return loss, loss_reg
def run_test(self, sess, feed):
test_score = sess.run(self.test_score, feed_dict = feed)
return test_score