-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathbidaf.py
199 lines (159 loc) · 9.11 KB
/
bidaf.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
# -*- coding: utf-8 -*-
"""
Created on Tue Feb 27 20:20:44 2018
@author: Peter
"""
import tensorflow as tf
class BiDAF(object):
def __init__(self, embeddings, quelen, conlen, embedding_size, hidden_size, keep_prob, stddev=0.1):
self.embeddings = embeddings
self.que_time_step = quelen
self.con_time_step = conlen
# self.embedding_size = embedding_size
self.embedding_size = hidden_size
self.hidden_size = hidden_size
self.keep_prob = keep_prob
self.stddev = stddev
self.x = tf.placeholder(tf.int32, [None, self.con_time_step])
self.q = tf.placeholder(tf.int32, [None, self.que_time_step])
self.y1 = tf.placeholder(tf.int32, [None, None])
self.y2 = tf.placeholder(tf.int32, [None, None])
self.lr = tf.placeholder(tf.float32)
self.X, self.Q = self.embedding_layer(self.x, self.q, self.con_time_step, self.que_time_step, self.embeddings)
self.H, self.U = self.contextual_embedding_layer(self.X, self.Q, self.con_time_step, self.que_time_step, self.hidden_size, self.embedding_size, self.keep_prob)
self.G = self.attention_flow_layer(self.H, self.U)
self.M = self.modeling_layer(self.G, self.con_time_step, self.que_time_step, self.hidden_size, self.embedding_size, self.keep_prob)
self.logit_s, self.logit_e = self.output_layer(self.G, self.M, self.con_time_step, self.hidden_size, self.embedding_size, self.keep_prob)
self.train, self.loss = self.optimize(self.logit_s, self.logit_e, self.y1, self.y2, self.lr)
self.acc_s = self.compute_accuracy(self.logit_s, self.y1)
self.acc_e = self.compute_accuracy(self.logit_e, self.y2)
def bidirectional_LSTM(self, X, time_step, hidden_size, embedding_size, keep_prob, output_dim=2, num_layers=1):
inputs = X
for _ in range(num_layers):
with tf.variable_scope(None, default_name="bidirectional-rnn"):
lstm_fw_cell = tf.contrib.rnn.BasicLSTMCell(hidden_size)
lstm_bw_cell = tf.contrib.rnn.BasicLSTMCell(hidden_size)
fw_drop = tf.contrib.rnn.DropoutWrapper(lstm_fw_cell, output_keep_prob=keep_prob)
bw_drop = tf.contrib.rnn.DropoutWrapper(lstm_bw_cell, output_keep_prob=keep_prob)
outputs, _ = tf.nn.bidirectional_dynamic_rnn(fw_drop, bw_drop, inputs, dtype=tf.float32)
concat_outputs = tf.concat(outputs, 2)
inputs = concat_outputs
return concat_outputs
'''
W = tf.Variable(tf.random_normal([2*hidden_size,embedding_size*output_dim],stddev=self.stddev))
b = tf.Variable(tf.constant(0.0,shape=[embedding_size*output_dim]))
reshaped_output = tf.reshape(concat_outputs,[-1,2*hidden_size]) #(batch_size*time_step,2*hidden_size)
reshaped_output = tf.matmul(reshaped_output,W) + b
Y = tf.reshape(reshaped_output,[-1,time_step,embedding_size*output_dim])
#(batch_size,time_step,output_dim*d)
return Y
'''
def embedding_layer(self, x, q, con_time_step, que_time_step, embeddings):
embeddings = tf.constant(embeddings, dtype=tf.float32)
embed_x = tf.nn.embedding_lookup(embeddings, x)
embed_q = tf.nn.embedding_lookup(embeddings, q)
return embed_x, embed_q
def higyway_network(self, inp, size, layer_size=1):
"""Highway Network (cf. http://arxiv.org/abs/1505.00387).
t = sigmoid(Wy + b)
z = t * g(Wy + b) + (1 - t) * y
where g is nonlinearity, t is transform gate, and (1 - t) is carry gate.
"""
out = inp
with tf.variable_scope('highway', reuse=tf.reuse):
for idx in range(layer_size):
out = tf.tanh(tf.layers.dense(inp, size, scope='dense1'))
transform_gate = tf.sigmoid(tf.layers.dense(inp, size, scope='dense2'))
carry_gate = 1. - transform_gate
out = transform_gate * out + carry_gate * inp
return out
def contextual_embedding_layer(self, X, Q, con_time_step, que_time_step, hidden_size, embedding_size, keep_prob):
with tf.variable_scope('Contextual_embeding_layer'):
with tf.variable_scope('context'):
H = self.bidirectional_LSTM(X, con_time_step, hidden_size, embedding_size, keep_prob, 2)
with tf.variable_scope('question'):
U = self.bidirectional_LSTM(Q, que_time_step, hidden_size, embedding_size, keep_prob, 2)
return H, U
def attention_flow_layer(self, H, U):
with tf.variable_scope('Attention_flow_layer'):
# B = H.get_shape().as_list()[0] # batch_size
T = H.get_shape().as_list()[1] # length of context
J = U.get_shape().as_list()[1] # length of question
D = H.get_shape().as_list()[-1] # dimension
H_trans = tf.transpose(H, [0, 2, 1]) # (B,2d,T)
U_trans = tf.transpose(U, [0, 2, 1]) # (B,2d,J)
# H:(B,T,D)
HH = tf.expand_dims(H, 2) # (B,T,1,D)
HH = tf.tile(HH, [1, 1, J, 1]) # (B,T,J,D)
# U:(B,J,D)
UU = tf.expand_dims(U, 1) # (B,1,J,D)
UU = tf.tile(UU, [1, T, 1, 1])
H_mul_U = tf.multiply(HH, UU) # (B,T,J,D)
H_mul_U = tf.reshape(H_mul_U, [-1, T * J, D]) # (B,T*J,D)
HHH = tf.reshape(HH, [-1, T * J, D]) # (B,T*J,D)
# UUU = tf.tile(U,[1,T,1]) # (B,T*J,D)
UUU = tf.reshape(UU, [-1, T * J, D])
concat = tf.concat([HHH, UUU, H_mul_U], 2) # (B,T*J,3D) 不换行并列连接
w_s = tf.Variable(tf.random_normal([3 * D, 1], stddev=self.stddev))
reshape = tf.reshape(concat, [-1, 3 * D]) # (B*T*J,3*D)
alpha = tf.matmul(reshape, w_s) # (B*T*J,1)
S = tf.reshape(alpha, [-1, T, J]) # (B,T,J)
a = tf.nn.softmax(S, -1)
a_trans = tf.transpose(a, [0, 2, 1]) # (B,J,T)
attended_U = tf.matmul(U_trans, a_trans) # (B,2d,T)
# 先在column上使用maximum function，再softmax
b = tf.nn.softmax(tf.reduce_max(S, 2), -1)
# (B,T) column上的维度被压缩了
b = tf.expand_dims(b, -1) # (B,T,1)
attended_H = tf.matmul(H_trans, b) # (B,2d,1)
attended_H = tf.tile(attended_H, [1, 1, T])
h_el_mul_att_u = tf.multiply(H_trans, attended_U) # (B,2d,T)
h_el_mul_att_h = tf.multiply(H_trans, attended_H) # (B,2d,T)
G = tf.concat([H_trans, attended_U, h_el_mul_att_u,
h_el_mul_att_h], axis=1) # (B,8d,T)
return G
def modeling_layer(self, G, con_time_step, que_time_step, hidden_size, embedding_size, keep_prob):
with tf.variable_scope('Modeling_layer'):
# G = tf.unstack(G,con_time_step,axis=2) # ?
G = tf.transpose(G, [0, 2, 1])
M = self.bidirectional_LSTM(G, con_time_step, hidden_size, embedding_size, keep_prob, 2, num_layers=2)
# (?,con_time_step,2d)
M = tf.transpose(M, [0, 2, 1])
# (?,2d,con_time_step)
return M
def output_layer(self, G, M, con_time_step, hidden_size, embedding_size, keep_prob):
with tf.variable_scope('Output_layer'):
w_p1 = tf.Variable(tf.random_normal([10 * embedding_size, 1], stddev=self.stddev)) # (10d,1)
G_M = tf.concat([G, M], axis=1) # (?,10d,con_time_step)
G_M = tf.transpose(G_M, [0, 2, 1]) # (?,con_time_step,10d)
G_M = tf.reshape(G_M, [-1, 10 * embedding_size])
p1 = tf.matmul(G_M, w_p1) # (?*con_time_step,1)
# (?,con_time_step)
p1 = tf.reshape(p1, [-1, M.get_shape().as_list()[-1]])
# M = tf.unstack(M,con_time_step,axis=2)
M2 = tf.transpose(M, [0, 2, 1])
M2 = self.bidirectional_LSTM(M2, con_time_step, hidden_size, embedding_size, keep_prob, 2) # (?,con_time_step,2d)
M2 = tf.transpose(M2, [0, 2, 1])
G_M2 = tf.concat([G, M2], axis=1)
# M = tf.stack(M,axis=2)
w_p2 = tf.Variable(tf.random_normal([10 * embedding_size, 1], stddev=self.stddev)) # (10d,1)
G_M2 = tf.transpose(G_M2, [0, 2, 1])
G_M2 = tf.reshape(G_M2, [-1, 10 * embedding_size])
p2 = tf.matmul(G_M2, w_p2)
# (?,con_time_step)
p2 = tf.reshape(p2, [-1, M.get_shape().as_list()[-1]])
return p1, p2
def optimize(self, logit_s, logit_e, y1, y2, lr):
loss1 = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logit_s, labels=y1))
loss2 = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=logit_e, labels=y2))
loss = loss1 + loss2
tvars = tf.trainable_variables()
grads, _ = tf.clip_by_global_norm(tf.gradients(loss, tvars), 5)
optimizer = tf.train.AdamOptimizer(learning_rate=lr)
train = optimizer.apply_gradients(zip(grads, tvars))
return train, loss
def compute_accuracy(self, logits, labels):
correct_prediction = tf.equal(
tf.argmax(logits, 1), tf.argmax(labels, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
return accuracy