forked from ot/entity2vec
-
Notifications
You must be signed in to change notification settings - Fork 0
/
entity_models.py
175 lines (116 loc) · 4.9 KB
/
entity_models.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
import logging
from collections import defaultdict, Counter
from multiprocessing.pool import Pool
from functools import partial
import numpy as np
from scipy.optimize import lbfgsb
import theano
import theano.tensor as T
from ext.gensim.utils import SaveLoad
from array_mmap_proxy import ArrayMMapProxy
def make_logistic_likelihood():
global log_l
lbda = T.dscalar('lbda')
w = T.dvector('w')
data = T.fmatrix('data')
pos_counts = T.vector('pos_counts', dtype='uint32')
neg_counts = T.vector('neg_counts', dtype='uint32')
fw = T.cast(w, 'float32') # lbfgsb only works with double
prods = T.dot(fw[:-1], data.T) + fw[-1]
ll = T.sum(pos_counts * T.log(1 + T.exp(prods))) + \
T.sum(neg_counts * T.log(1 + T.exp(-prods)))
ll_reg = ll + lbda * T.sum(fw[:-1]**2) / (fw.shape[0] - 1)
ll_reg_grad = T.cast(T.grad(ll_reg, wrt=w), 'float64') # again
log_l = theano.function([w, data, pos_counts, neg_counts, lbda],
(ll_reg, ll_reg_grad),
mode='FAST_RUN')
class EntityModel(SaveLoad):
def __init__(self):
self.vectors = None
self.entities = None
def _train_base(self, compute_vector, entity_word_seqs):
pool = Pool()
entities = {}
vectors = []
def idx_seqs():
for idx, (entity, seq) in enumerate(entity_word_seqs):
entities[entity] = idx
yield seq
for vec in pool.imap(compute_vector, idx_seqs()):
vectors.append(vec)
if len(vectors) % 1000 == 0:
logging.info('Computed %d vectors', len(vectors))
self.entities = entities
self.vectors = np.asarray(vectors)
def _count_word_vectors(self, model, word_idxs):
counts = Counter(word_idxs)
word_vecs = model.syn0[counts.keys()]
word_counts = counts.values()
return word_vecs, word_counts
def sample(bins, size):
points = np.random.randint(bins[-1], size=size)
return np.searchsorted(bins, points)
def compute_vector_lr(syn0_proxy, bins_proxy,
neg_words_mult, lbda, word_idxs):
syn0 = syn0_proxy.get()
bins = bins_proxy.get()
counts = defaultdict(lambda: np.zeros(2).astype(np.uint32))
for widx in word_idxs:
counts[widx][0] += 1
neg_words_idxs = sample(bins, int(neg_words_mult * len(word_idxs)))
for neg_widx in neg_words_idxs:
counts[neg_widx][1] += 1
vectors = syn0[counts.keys()]
count_pairs = np.vstack(counts.values())
f = lambda w, params=(vectors, count_pairs[:, 0], count_pairs[:, 1], lbda): log_l(w, *params)
x0 = np.zeros(syn0.shape[1] + 1)
opt = lbfgsb.fmin_l_bfgs_b(f, x0)
if opt[2]['warnflag']:
logging.debug('Error in optimization: %s', opt[2])
lr_vec = opt[0].astype(np.float32)
if not np.all(np.isfinite(lr_vec)):
logging.info('Error computing lr vector')
lr_vec[:] = 0
return lr_vec
class EntityModelLR(EntityModel):
def __init__(self, neg_cdf, neg_words_mult, lbda):
super(EntityModel, self).__init__()
self.neg_cdf = neg_cdf
self.neg_words_mult = neg_words_mult
self.lbda = lbda
def train(self, model, entity_word_seqs):
make_logistic_likelihood()
self._train_base(partial(compute_vector_lr,
ArrayMMapProxy(model.syn0),
ArrayMMapProxy.fromarray(self.neg_cdf),
self.neg_words_mult,
self.lbda),
entity_word_seqs)
def score(self, model, entity_vecs, words_idxs):
word_vecs, word_counts = self._count_word_vectors(model, words_idxs)
weights = entity_vecs[:, :-1]
bias = entity_vecs[:, -1:]
dotprods = np.dot(weights, word_vecs.T)
log_dot_prods = np.log(1 + np.exp(dotprods + bias))
return -np.dot(word_counts, log_dot_prods.T)
def compute_vector_centroid(syn0_proxy, word_idxs):
syn0 = syn0_proxy.get()
counts = Counter(word_idxs)
centroid_vec = np.dot(counts.values(), syn0[counts.keys()])
centroid_vec /= np.sqrt(np.sum(centroid_vec**2))
if not np.all(np.isfinite(centroid_vec)):
logging.info('Error computing centroid vector')
centroid_vec[:] = 0
return centroid_vec
class EntityModelCentroid(EntityModel):
def __init__(self):
super(EntityModel, self).__init__()
def train(self, model, entity_word_seqs):
self._train_base(partial(compute_vector_centroid,
ArrayMMapProxy(model.syn0)),
entity_word_seqs)
def score(self, model, entity_vecs, words_idxs):
word_vecs, word_counts = self._count_word_vectors(model, words_idxs)
centroid = np.dot(word_counts, word_vecs)
centroid /= np.sqrt(np.sum(centroid**2))
return np.dot(entity_vecs, centroid)