-
Notifications
You must be signed in to change notification settings - Fork 0
/
SkipGram.py
270 lines (212 loc) · 11.4 KB
/
SkipGram.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
import numpy as np
import tensorflow as tf
import tensorflow.compat.v1 as v1
import os
from datetime import datetime
class SkipGram(object):
def __init__(self, vocab_length, emb_length):
self.vocab_length = vocab_length
self.emb_length = emb_length
def build_graph(self, vocab_length, emb_length, context_size=None,
sampling='log-uniform', n_neg_samples=1, unigrams=None, distortion=0.75, tf_seed=None):
if sampling=='unigram':
assert unigrams is not None
assert len(unigrams)==vocab_length
if context_size is None:
context_size=1
g = tf.Graph()
with g.as_default():
# word indices
w = v1.placeholder(tf.int64, shape=(None), name='w')
# context indices
c = v1.placeholder(tf.int64, shape=(None, context_size), name='c')
learning_rate = v1.placeholder_with_default(1.0, shape=(), name='learning_rate')
l1_penalty = v1.placeholder_with_default(0.0, shape=(), name='l1_penalty')
l2_penalty = v1.placeholder_with_default(0.0, shape=(), name='l2_penalty')
emb_init = v1.initializers.he_normal(seed=tf_seed)
embeddings = tf.Variable(emb_init(shape=(vocab_length, emb_length)),
name='embedding')
l1_loss = l1_penalty * tf.reduce_mean(tf.abs(embeddings))
l2_loss = l2_penalty * tf.reduce_mean(tf.square(embeddings))
w_emb = tf.nn.embedding_lookup(
embeddings,
w,
name='w_emb')
c_emb = tf.nn.embedding_lookup(
embeddings,
c,
name='c_emb')
w_emb_reshaped = tf.reshape(w_emb, (-1, 1, emb_length))
c_logits = tf.reduce_sum(w_emb_reshaped * c_emb, axis=2, name='c_logits')
loss_normalizer = context_size * tf.reduce_logsumexp(tf.matmul(w_emb, tf.transpose(embeddings)), axis=1)
loss = tf.reduce_mean(
loss_normalizer - tf.reduce_sum(c_logits, axis=1),
name='loss')
regularized_loss = tf.identity(loss + l1_loss + l2_loss, name='regularized_loss')
if sampling=='uniform':
sampled_values = tf.random.uniform_candidate_sampler(
true_classes=c,
num_true=context_size,
num_sampled=n_neg_samples,
unique=True,
range_max=vocab_length)
elif sampling=='log-uniform':
sampled_values = tf.random.log_uniform_candidate_sampler(
true_classes=c,
num_true=context_size,
num_sampled=n_neg_samples,
unique=True,
range_max=vocab_length)
elif sampling=='unigram':
sampled_values = tf.random.fixed_unigram_candidate_sampler(
true_classes=c,
num_true=context_size,
num_sampled=n_neg_samples,
unique=True,
range_max=vocab_length,
unigrams=unigrams,
distortion=distortion
)
else:
raise AssertionError('Invalid sampling option')
sampled_loss = tf.reduce_mean(
tf.nn.sampled_softmax_loss(
weights=embeddings,
biases=tf.zeros(vocab_length),
labels=c,
inputs=w_emb,
num_sampled=n_neg_samples,
num_classes=vocab_length,
num_true = context_size,
sampled_values=sampled_values,
remove_accidental_hits=False),
name='sampled_loss')
training_loss = sampled_loss + l1_loss + l2_loss
optimizer = v1.train.GradientDescentOptimizer(learning_rate)
train_op = optimizer.minimize(training_loss, name='train_op')
return g
def show_w_emb(self):
g = self.build_graph(self.vocab_length, self.emb_length)
with v1.Session(graph=g) as sess:
sess.run(v1.global_variables_initializer())
w = list(range(self.vocab_length))
return sess.run('w_emb:0', feed_dict={'w:0':w})
def show_c_emb(self):
n = self.vocab_length
c = [[(i-1)%n, (i+1)%n] for i in range(self.vocab_length)]
context_size = len(c[0])
g = self.build_graph(self.vocab_length, self.emb_length, context_size=context_size)
with v1.Session(graph=g) as sess:
sess.run(v1.global_variables_initializer())
return sess.run('c_emb:0', feed_dict={'c:0':c})
def return_examples(self):
n = self.vocab_length
w = list(range(self.vocab_length))
c = [[(i-1)%n, (i+1)%n] for i in range(self.vocab_length)]
context_size = len(c[0])
g = self.build_graph(self.vocab_length, self.emb_length, context_size=context_size)
with v1.Session(graph=g) as sess:
sess.run(v1.global_variables_initializer())
return sess.run(['w_emb:0', 'c_emb:0', 'c_logits:0'], feed_dict={'w:0':w, 'c:0':c})
def loss(self, word_indices, context_indices, regularize=False, l1_penalty=0., l2_penalty=0., checkpoint_dir=None, use_batches=True,
batch_size=1024, n_loss_batches=1000, seed=None):
g = self.build_graph(self.vocab_length, self.emb_length, tf_seed=seed)
with g.as_default():
saver = v1.train.Saver()
with v1.Session(graph=g) as sess:
sess.run(v1.global_variables_initializer())
return self._loss(sess, saver, word_indices, context_indices,
regularize=regularize, l1_penalty=l1_penalty, l2_penalty=l2_penalty,
checkpoint_dir=checkpoint_dir, use_batches=use_batches, batch_size=batch_size)
def _loss(self, sess, saver, word_indices, context_indices, regularize=False, l1_penalty=0, l2_penalty=0., checkpoint_dir=None, use_batches=True, batch_size=1024, n_loss_batches=1000):
assert len(word_indices) == len(context_indices)
assert type(regularize)==type(True)
assert type(use_batches)==type(True)
random = np.random.RandomState()
if regularize:
loss_name='regularized_loss:0'
else:
loss_name='loss:0'
if checkpoint_dir is not None:
saver.restore(
sess,
tf.train.latest_checkpoint(checkpoint_dir))
if use_batches==False:
feed = {'w:0': word_indices, 'c:0': context_indices,
'l1_penalty:0':l1_penalty, 'l2_penalty:0':l2_penalty}
return sess.run(loss_name, feed_dict=feed)
elif use_batches==True:
n_samples = len(word_indices)
n_batches = len(range(0, n_samples, batch_size))
n_loss_batches = min(n_loss_batches, n_batches)
if n_loss_batches == n_batches:
batches_to_sample = np.arange(n_batches)
else:
batches_to_sample = random.choice(n_batches, n_loss_batches, replace=False)
losses = np.zeros(n_loss_batches)
weights = np.zeros(n_loss_batches)
for i, batch_idx in enumerate(batches_to_sample):
start = batch_idx * batch_size
end = start + batch_size
w, c = word_indices[start:end], context_indices[start:end]
feed = {'w:0':w, 'c:0':c,
'l1_penalty:0':l1_penalty, 'l2_penalty:0':l2_penalty}
losses[i] = sess.run(loss_name, feed_dict=feed)
weights[i] = len(w)
return np.average(losses, None, weights)
def train(self, word_indices, context_indices, l1_penalty=0., l2_penalty=0., sampling='log-uniform', neg_sample_rate=5,
unigrams=None, distortion=0.75, learning_rate=1, batch_size=64, n_epochs=1, checkpoint_dir=None,
load_prev=False, prev_epochs=0, print_reports=False, n_batch_reports=10, n_loss_batches=1000, seed=None):
assert len(word_indices) == len(context_indices)
word_indices = np.copy(word_indices)
context_indices = np.copy(context_indices)
random = np.random.RandomState(seed)
n_samples = len(word_indices)
n_batches = len(range(0, n_samples, batch_size))
report_at_batches = [int(round(x * n_batches / n_batch_reports)) for x in range(1, n_batch_reports+1)]
context_size = len(context_indices[0])
n_neg_samples = max(1, int(round(neg_sample_rate * context_size)))
g = self.build_graph(self.vocab_length, self.emb_length, context_size=context_size,
sampling=sampling, unigrams=unigrams, distortion=distortion, n_neg_samples=n_neg_samples, tf_seed=seed)
with g.as_default():
saver = v1.train.Saver()
with v1.Session(graph=g) as sess:
sess.run(v1.global_variables_initializer())
if checkpoint_dir is not None and load_prev:
saver.restore(
sess,
tf.train.latest_checkpoint(checkpoint_dir))
for epoch in range(1, n_epochs+1):
for batch_n, j in enumerate(range(0, n_samples, batch_size), 1):
w, c = word_indices[j:j+batch_size], context_indices[j:j+batch_size]
feed = {'w:0':w, 'c:0':c, 'learning_rate:0':learning_rate,
'l1_penalty:0':l1_penalty, 'l2_penalty:0':l2_penalty}
_ = sess.run('train_op', feed_dict=feed)
if print_reports and batch_n in report_at_batches:
loss = sess.run('regularized_loss:0', feed_dict=feed)
print(str(datetime.now())+':', 'Epoch %d, batch %d: loss %.4f' % (epoch+prev_epochs, batch_n, loss))
if print_reports:
loss = self._loss(sess, saver, word_indices, context_indices,
regularize=True, l1_penalty=l1_penalty, l2_penalty=l2_penalty,
use_batches=True, batch_size=batch_size, n_loss_batches=n_loss_batches)
print(str(datetime.now())+':', 'Epoch %d: loss %.4f' % (epoch+prev_epochs, loss))
if checkpoint_dir is not None:
saver.save(sess, os.path.join(checkpoint_dir, 'skip_gram_'+str(self.emb_length)), global_step=epoch+prev_epochs)
if epoch < n_epochs:
new_order = random.permutation(n_samples)
word_indices = word_indices[new_order]
context_indices = context_indices[new_order]
def embed(self, word_indices, checkpoint_dir=None, seed=None):
g = self.build_graph(self.vocab_length, self.emb_length, tf_seed=seed)
with g.as_default():
saver = v1.train.Saver()
with v1.Session(graph=g) as sess:
sess.run(v1.global_variables_initializer())
if checkpoint_dir is None:
v1.logging.warn('No checkpoint selected. Embedding matrix will be randomly initialized')
else:
saver.restore(
sess,
tf.train.latest_checkpoint(checkpoint_dir))
feed={'w:0': word_indices}
return sess.run('w_emb:0', feed)