forked from youngseon/vae-paraphrase
-
Notifications
You must be signed in to change notification settings - Fork 0
/
model.py
279 lines (190 loc) · 12.5 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
import tensorflow as tf
import numpy as np
from tensorflow.python.layers import core
from tensorflow.python.ops import array_ops
import data
def build_cell(cell_hidden, keep_prob):
cell = tf.nn.rnn_cell.LSTMCell(cell_hidden)
return tf.contrib.rnn.DropoutWrapper(cell, output_keep_prob=keep_prob)
def build_decoder_cell(cell_hidden, keep_prob):
multi_cell = tf.nn.rnn_cell.MultiRNNCell([build_cell(cell_hidden, keep_prob) for n in range(2)])
return multi_cell
class Encoder():
def encode(self, embedding, inputs, lengths, cell_hidden, keep_prob, dtype, initial_state=(None, None), scope='Encoder', reuse=False):
print("build_encoder/", scope)
init_state_fw, init_state_bw =initial_state
with tf.variable_scope(scope, reuse=reuse):
cell_fw = build_cell(cell_hidden, keep_prob=keep_prob)
cell_bw = build_cell(cell_hidden, keep_prob=keep_prob)
final_input = tf.nn.embedding_lookup(embedding, inputs)
encoder_outputs, encoder_state = tf.nn.bidirectional_dynamic_rnn(cell_fw, cell_bw, final_input,
sequence_length=lengths,
dtype=dtype,
initial_state_fw=init_state_fw,
initial_state_bw=init_state_bw)
return encoder_outputs, encoder_state
class Decoder():
def __init__(self, attention_dim):
self.decoder_cell = None
self.att_mech = None
self.attention_dim = attention_dim
def build_decoder_cell(self, n_hidden, keep_prob, scope="DecoderCell", reuse=False):
with tf.variable_scope(scope, reuse=reuse):
self.decoder_cell = build_decoder_cell(n_hidden, keep_prob=keep_prob)
return self.decoder_cell
def build_att_mechanism(self, seq_input, seq_len, beam_width, scope='Attention', reuse=False):
assert(self.decoder_cell != None)
with tf.variable_scope(scope, reuse=reuse):
tiled_seq_input = tf.contrib.seq2seq.tile_batch(seq_input, multiplier=beam_width)
tiled_seq_len = tf.contrib.seq2seq.tile_batch(seq_len, multiplier=beam_width)
self.att_mech = tf.contrib.seq2seq.BahdanauAttention(self.attention_dim, tiled_seq_input, memory_sequence_length=tiled_seq_len)
self.decoder_cell = tf.contrib.seq2seq.AttentionWrapper(self.decoder_cell, self.att_mech)
return self.decoder_cell
def decode(self, embedding, beam, decoder_inputs, decoder_lengths, init_state, projection_layer, latents, latent_dim, scope='Decoder', reuse=False):
print("build_decoder/", scope)
decoder_emb_inp = tf.nn.embedding_lookup(embedding, decoder_inputs)
max_len = tf.shape(decoder_inputs)[1]
batch_size = tf.shape(decoder_inputs)[0]
latents = tf.expand_dims(latents, axis=1)
latents = tf.tile(latents, [1,max_len,1])
latents = tf.reshape(latents, [batch_size, max_len, latent_dim])
decoder_final_inp = tf.concat([decoder_emb_inp, latents], axis=2)
with tf.variable_scope(scope, reuse=reuse):
if (self.att_mech is not None):
decoder_initial_state = self.decoder_cell.zero_state(batch_size, dtype=tf.float32)
init_state = decoder_initial_state.clone(cell_state=init_state)
# build training helper
training_helper = tf.contrib.seq2seq.TrainingHelper(decoder_final_inp, decoder_lengths)
# output layer
training_decoder = tf.contrib.seq2seq.BasicDecoder(self.decoder_cell, training_helper, init_state, projection_layer)
outputs, _, _ = tf.contrib.seq2seq.dynamic_decode(training_decoder)
# logits for the loss calculation
logits = outputs.rnn_output
sample_ids = outputs.sample_id
return logits, sample_ids
def predict(self, embedding, beam, batch_size, init_state, projection_layer, latents, latent_dim, max_decoder_length=15, beam_width=10, scope='Decoder', reuse=False):
print("build_predictor/", scope)
# build predition helper
start_tokens = tf.tile([data._SOS_ID], [batch_size])
end_token = data._EOS_ID
tiled_latents = tf.expand_dims(latents, axis=1)
tiled_latents = tf.tile(tiled_latents, [1,beam,1])
tiled_latents = tf.reshape(tiled_latents, [batch_size, beam, latent_dim])
init_state = tf.contrib.seq2seq.tile_batch(init_state, multiplier=beam)
def emb_fn(input_tokens, tiled_latents=tiled_latents, embedding_decoder=embedding):
emb_inp = tf.nn.embedding_lookup(embedding, input_tokens)
concat_inp = array_ops.concat([emb_inp, tiled_latents], 2)
return concat_inp
with tf.variable_scope(scope, reuse=True):
if (self.att_mech is not None):
decoder_initial_state = self.decoder_cell.zero_state(batch_size * beam, dtype=tf.float32)
init_state = decoder_initial_state.clone(cell_state=init_state)
# build predition helper
start_tokens = tf.tile([data._SOS_ID], [batch_size])
end_token = data._EOS_ID
predicting_decoder = tf.contrib.seq2seq.BeamSearchDecoder(cell=self.decoder_cell, embedding=emb_fn, start_tokens=start_tokens, end_token=end_token,
initial_state=init_state, beam_width=beam_width, output_layer=projection_layer)
predict_outputs, _, _ = tf.contrib.seq2seq.dynamic_decode(predicting_decoder, maximum_iterations=max_decoder_length)
return predict_outputs
class Model():
def __init__(self, config, vocab_size, pretrained_emb):
self.encoder = Encoder()
self.decoder = Decoder(attention_dim=config["attention_dim"])
self.annealing = True
self.initializer = tf.contrib.layers.xavier_initializer()
self.vocab_size = vocab_size
self.embedding_size = config["emb_size"]
self.pretrained_embedding = pretrained_emb
self.n_hidden = config["hidden"]
self.learning_rate = config["learning_rate"]
self.step_size = config["step_size"]
self.annealing_pivot = config["annealing_pivot"]
self.latent_dim = config["latent_dim"]
self.beam_size = config["beam_size"]
self.reconst_weight = 100.0
self.dtype = tf.float32
self.embedding = self.get_embedding()
self.global_step = tf.Variable(0, name='global_step', trainable=False)
self.epsilon = 0.000001
self.mode = "SVAE"
def define_kl_updators(self):
g_step_float = tf.cast(self.global_step, dtype=self.dtype)
reg = tf.constant(-self.annealing_pivot, dtype=self.dtype)
nom = tf.constant(1000, dtype=self.dtype)
step_term = tf.nn.tanh(tf.div(tf.add(g_step_float, reg), nom))
kl_rate = tf.div(tf.add(step_term, tf.constant(1.0, dtype=self.dtype)), tf.constant(2.0, dtype=self.dtype))
return kl_rate
def get_embedding(self):
with tf.variable_scope('embedding'):
embedding = tf.get_variable(name="embedding", shape=[self.vocab_size, self.embedding_size],
initializer=tf.constant_initializer(np.array(self.pretrained_embedding)), trainable=False, dtype=self.dtype)
return embedding
def get_loss(self, label, predict, max_length, decoder_lengths, dtype):
label_length = tf.shape(label)[1]
logit_length = tf.shape(predict)[1]
pad_size = label_length - logit_length
predict = tf.pad(predict, [[0, 0], [0, pad_size], [0, 0]], constant_values=data._PAD_ID)
max_decoder_length = max_length
masks = tf.sequence_mask(lengths=decoder_lengths,
maxlen=max_decoder_length, dtype=dtype, name='masks')
reconst_cost = tf.contrib.seq2seq.sequence_loss(logits=predict,
targets=label,
weights=masks,
average_across_timesteps=True,
average_across_batch=True)
return reconst_cost
def build_encode_latent(self, output_states, batch_size, keep_prob):
state_fw, state_bw = output_states
init_state = [state_fw.c, state_bw.c, state_fw.h, state_bw.h]
vector = tf.concat(axis=1, values=init_state)
with tf.variable_scope('encode_latent'):
vector = tf.reshape(vector, [batch_size, 4 *self.n_hidden])
sample = tf.contrib.layers.fully_connected(vector, self.latent_dim * 2, activation_fn=tf.nn.tanh)
sample = tf.nn.dropout(sample, keep_prob=keep_prob)
mu = tf.contrib.layers.fully_connected(sample, self.latent_dim, activation_fn=tf.nn.tanh)
logvar = tf.contrib.layers.fully_connected(sample, self.latent_dim, activation_fn=tf.nn.softplus)
return mu, logvar
def add_gaussian_noise(self, mu, logvar, scope='kl_sample'):
with tf.variable_scope(scope):
q_z = tf.distributions.Normal(mu, logvar)
z = q_z.sample()
p_z = tf.distributions.Normal(tf.zeros_like(z), tf.ones_like(z))
kl = tf.distributions.kl_divergence(q_z, p_z, allow_nan_stats=True)
kl_cost = tf.reduce_mean(tf.reduce_sum(kl, axis=-1))
return z, kl_cost
def build_train_ops(self, data_cap):
self.kl_rate = kl_rate = self.define_kl_updators()
self.keep_prob = keep_prob = data_cap.keep_prob
batch_size = tf.shape(data_cap.source_inputs)[0]
# encode sentences
source_output, source_last_state = self.encoder.encode(self.embedding, data_cap.source_inputs, data_cap.source_lengths,
self.n_hidden, keep_prob, self.dtype, scope='SourceEncoder')
reference_output, reference_last_state = self.encoder.encode(self.embedding, data_cap.reference_inputs, data_cap.reference_lengths,
self.n_hidden, keep_prob, self.dtype, initial_state=source_last_state, scope='SourceEncoder', reuse=True)
mu, sig = self.build_encode_latent(reference_last_state, batch_size, keep_prob)
self.not_sampled_latent = mu
latent, kl_cost = self.add_gaussian_noise(mu, sig)
self.sampled_latent = latent
self.given_latent = data_cap.latent_variable
self.kl_cost = kl_cost
projection_layer = core.Dense(self.vocab_size, name='output_projection')
# attention
self.decoder.build_decoder_cell(self.n_hidden, keep_prob)
source_output_concat = tf.concat([source_output[0], source_output[1]], axis=2)
self.decoder.build_att_mechanism(source_output_concat, data_cap.source_lengths, data_cap.beam)
# decode
logits, self.sample_ids = self.decoder.decode(self.embedding, data_cap.beam,
data_cap.decoder_inputs, data_cap.decoder_lengths,
source_last_state, projection_layer, self.sampled_latent, self.latent_dim)
self.predict_op = predict_op = self.decoder.predict(self.embedding, data_cap.beam,
batch_size, source_last_state, projection_layer, self.given_latent, self.latent_dim,
max_decoder_length=self.step_size, beam_width=self.beam_size, reuse=True)
max_length = tf.shape(data_cap.decoder_inputs)[1]
# loss and ops
sequence_loss = self.get_loss(data_cap.targets, logits, max_length, data_cap.decoder_lengths, self.dtype)
self.loss = (self.reconst_weight*sequence_loss) + (kl_rate * kl_cost)
params = tf.trainable_variables()
gradients = tf.gradients(self.loss, params)
clipped_gradients, _ = tf.clip_by_global_norm(gradients, 5.0)
self.train_op = train_op = tf.train.AdamOptimizer(learning_rate=self.learning_rate).apply_gradients(zip(clipped_gradients, params), global_step=self.global_step)
return train_op, predict_op