forked from williamleif/socialsent
-
Notifications
You must be signed in to change notification settings - Fork 0
/
embedding_transformer.py
195 lines (163 loc) · 7.71 KB
/
embedding_transformer.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
from socialsent import lexicons
from socialsent import util
import random
import numpy as np
import matplotlib.pyplot as plt
from itertools import combinations, product
from keras import backend as K
from keras.models import Model
from keras.layers.core import Dense, Lambda
from keras.optimizers import Adam, Optimizer
from keras.regularizers import Regularizer
from keras.constraints import Constraint
import theano.tensor as T
from socialsent.representations.embedding import Embedding
"""
Helper methods for learning transformations of word embeddings.
"""
class SimpleSGD(Optimizer):
def __init__(self, lr=5, momentum=0., decay=0.,
nesterov=False, **kwargs):
super(SimpleSGD, self).__init__(**kwargs)
self.__dict__.update(locals())
self.iterations = K.variable(0.)
self.lr = K.variable(lr)
self.momentum = K.variable(momentum)
self.decay = K.variable(decay)
def get_updates(self, params, constraints, loss):
grads = self.get_gradients(loss, params)
lr = self.lr * 0.99
self.updates = [(self.iterations, self.iterations + 1.)]
# momentum
self.weights = [K.variable(np.zeros(K.get_value(p).shape)) for p in params]
for p, g, m in zip(params, grads, self.weights):
v = self.momentum * m - lr * g # velocity
self.updates.append((m, v))
if self.nesterov:
new_p = p + self.momentum * v - lr * g
else:
new_p = p + v
# apply constraints
if p in constraints:
c = constraints[p]
new_p = c(new_p)
self.updates.append((p, new_p))
return self.updates
def get_config(self):
config = {'lr': float(K.get_value(self.lr)),
'momentum': float(K.get_value(self.momentum)),
'decay': float(K.get_value(self.decay)),
'nesterov': self.nesterov}
base_config = super(SimpleSGD, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
class Orthogonal(Constraint):
def __call__(self, p):
print "here"
u,s,v = T.nlinalg.svd(p)
return K.dot(u,K.transpose(v))
class OthogonalRegularizer(Regularizer):
def __init__(self, strength=0.):
self.strength = strength
def set_param(self, p):
self.p = p
def __call__(self, loss):
loss += K.sum(K.square(self.p.dot(self.p.T) - T.identity_like(self.p))) * self.strength
return loss
def get_config(self):
return {"name": self.__class__.__name__,
"strength": self.strength}
def orthogonalize(Q):
U, S, V = np.linalg.svd(Q)
return U.dot(V.T)
class DatasetMinibatchIterator:
def __init__(self, embeddings, positive_seeds, negative_seeds, batch_size=512, **kwargs):
self.words, embeddings1, embeddings2, labels = [], [], [], []
def add_examples(word_pairs, label):
for w1, w2 in word_pairs:
embeddings1.append(embeddings[w1])
embeddings2.append(embeddings[w2])
labels.append(label)
self.words.append((w1, w2))
add_examples(combinations(positive_seeds, 2), 1)
add_examples(combinations(negative_seeds, 2), 1)
add_examples(product(positive_seeds, negative_seeds), -1)
self.e1 = np.vstack(embeddings1)
self.e2 = np.vstack(embeddings2)
self.y = np.array(labels)
self.batch_size = batch_size
self.n_batches = (self.y.size + self.batch_size - 1) / self.batch_size
def shuffle(self):
perm = np.random.permutation(np.arange(self.y.size))
self.e1, self.e2, self.y, self.words = \
self.e1[perm], self.e2[perm], self.y[perm], [self.words[i] for i in perm]
def __iter__(self):
for i in range(self.n_batches):
batch = np.arange(i * self.batch_size, min(self.y.size, (i + 1) * self.batch_size))
yield {
'embeddings1': self.e1[batch],
'embeddings2': self.e2[batch],
'y': self.y[batch][:, np.newaxis]
}
def get_model(inputdim, outputdim, regularization_strength=0.01, lr=0.000, cosine=False, **kwargs):
transformation = Dense(inputdim, init='identity',
W_constraint=Orthogonal())
model = Model()
model.add_input(name='embeddings1', input_shape=(inputdim,))
model.add_input(name='embeddings2', input_shape=(inputdim,))
model.add_shared_node(transformation, name='transformation',
inputs=['embeddings1', 'embeddings2'],
outputs=['transformed1', 'transformed2'])
model.add_node(Lambda(lambda x: x[:, :outputdim]), input='transformed1', name='projected1')
model.add_node(Lambda(lambda x: -x[:, :outputdim]), input='transformed2', name='negprojected2')
if cosine:
model.add_node(Lambda(lambda x: x / K.reshape(K.sqrt(K.sum(x * x, axis=1)), (x.shape[0], 1))),
name='normalized1', input='projected1')
model.add_node(Lambda(lambda x: x / K.reshape(K.sqrt(K.sum(x * x, axis=1)), (x.shape[0], 1))),
name='negnormalized2', input='negprojected2')
model.add_node(Lambda(lambda x: K.reshape(K.sum(x, axis=1), (x.shape[0], 1))),
name='distances', inputs=['normalized1', 'negnormalized2'], merge_mode='mul')
else:
model.add_node(Lambda(lambda x: K.reshape(K.sqrt(K.sum(x * x, axis=1)), (x.shape[0], 1))),
name='distances', inputs=['projected1', 'negprojected2'], merge_mode='sum')
model.add_output(name='y', input='distances')
model.compile(loss={'y': lambda y, d: K.mean(y * d)}, optimizer=SimpleSGD())
return model
def apply_embedding_transformation(embeddings, positive_seeds, negative_seeds,
n_epochs=5, n_dim=10, force_orthogonal=False,
plot=False, plot_points=50, plot_seeds=False,
**kwargs):
print "Preparing to learn embedding tranformation"
dataset = DatasetMinibatchIterator(embeddings, positive_seeds, negative_seeds, **kwargs)
model = get_model(embeddings.m.shape[1], n_dim, **kwargs)
print "Learning embedding transformation"
# prog = util.Progbar(n_epochs)
for epoch in range(n_epochs):
dataset.shuffle()
loss = 0
for i, X in enumerate(dataset):
loss += model.train_on_batch(X)[0] * X['y'].size
Q, b = model.get_weights()
if force_orthogonal:
Q = orthogonalize(Q)
model.set_weights([Q, np.zeros_like(b)])
# prog.update(epoch + 1, exact_values=[('loss', loss / dataset.y.size)])
Q, b = model.get_weights()
new_mat = embeddings.m.dot(Q)[:,0:n_dim]
#print "Orthogonality rmse", np.mean(np.sqrt(
# np.square(np.dot(Q, Q.T) - np.identity(Q.shape[0]))))
if plot and n_dim == 2:
plot_words = positive_seeds + negative_seeds if plot_seeds else \
[w for w in embeddings if w not in positive_seeds and w not in negative_seeds]
plot_words = set(random.sample(plot_words, plot_points))
to_plot = {w: embeddings[w] for w in embeddings if w in plot_words}
lexicon = lexicons.load_lexicon()
plt.figure(figsize=(10, 10))
for w, e in to_plot.iteritems():
plt.text(e[0], e[1], w,
bbox=dict(facecolor='green' if lexicon[w] == 1 else 'red', alpha=0.1))
xmin, ymin = np.min(np.vstack(to_plot.values()), axis=0)
xmax, ymax = np.max(np.vstack(to_plot.values()), axis=0)
plt.xlim(xmin, xmax)
plt.ylim(ymin, ymax)
plt.show()
return Embedding(new_mat, embeddings.iw, normalize=n_dim!=1)