-
Notifications
You must be signed in to change notification settings - Fork 1
/
model.py
238 lines (220 loc) · 12.9 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
# -*- coding: utf-8 -*-
# @Time : 2019/3/19 16:21
# @Author : Alan
# @Email : [email protected]
# @File : model.py
# @Software: PyCharm
import tensorflow as tf
from model_utils import *
from tensorflow.contrib import rnn
class MAN(object):
def __init__(self, config):
self.ques_len = config.ques_length
self.ans_len = config.ans_length
self.hidden_size = config.hidden_size
self.output_size = config.output_size
self.rnn_size = config.rnn_size
self.learning_rate = config.learning_rate
self.optimizer = config.optimizer
self.l2_lambda = config.l2_lambda
self.clip_value = config.clip_value
self.embeddings = config.embeddings
self.window_sizes = config.window_sizes
self.n_filters = config.n_filters
self.margin = config.margin
self.num_steps = config.num_steps
self.layer_size = config.layer_size
self._placeholder_init_pointwise()
self._initialize_weights()
self.q_a_cosine, self.q_aneg_cosine = self._build(self.embeddings)
# 损失和精确度
self.total_loss, self.accu = self._add_loss_op(self.q_a_cosine, self.q_aneg_cosine, self.l2_lambda)
# 训练节点
self.train_op = self._add_train_op(self.total_loss)
def _placeholder_init_pointwise(self):
self._ques = tf.placeholder(tf.int32, [None, self.ques_len], name='ques_point')
self._ans = tf.placeholder(tf.int32, [None, self.ans_len], name='ans_point')
self._ans_neg = tf.placeholder(tf.int32, [None, self.ans_len], name='ans_point')
self._ques_mask = tf.placeholder(tf.int32, [None], 'ques_mask')
self._ans_mask = tf.placeholder(tf.int32, [None], 'ans_mask')
self._ans_neg_mask = tf.placeholder(tf.int32, [None], 'ans_neg_mask')
self.dropout_keep_prob = tf.placeholder(tf.float32, name="dropout_keep_prob")
self.batch_size, self.list_size = tf.shape(self._ans)[0], tf.shape(self._ans)[1]
def _initialize_weights(self):
self.W_q1 = weight_variable('W_q1', [2 * self.rnn_size, 2*self.rnn_size])
self.W_m1 = weight_variable('W_m1', [2 * self.rnn_size, 2 * self.rnn_size])
self.W_s1 = weight_variable('W_s1', [2 * self.rnn_size, 1])
self.W_q2 = weight_variable('W_q2', [2 * self.rnn_size, 2*self.rnn_size])
self.W_m2 = weight_variable('W_m2', [2 * self.rnn_size, 2 * self.rnn_size])
self.W_s2 = weight_variable('W_s2', [2 * self.rnn_size, 1])
self.W_q3 = weight_variable('W_q3', [2 * self.rnn_size, 2*self.rnn_size])
self.W_m3 = weight_variable('W_m3', [2 * self.rnn_size, 2 * self.rnn_size])
self.W_s3 = weight_variable('W_s3', [2 * self.rnn_size, 1])
def _bilstm_layer(self, inputs, rnn_size, seq_len, layer_size, batch_size, keep_prob, scope, reuse=False):
"""
双向LSTM
"""
with tf.variable_scope(scope, reuse=reuse):
def cell():
lstm_cell = rnn.LSTMCell(num_units=rnn_size)
lstm_cell = rnn.DropoutWrapper(lstm_cell, output_keep_prob=keep_prob) ##
return lstm_cell
cell_bw = cell_fw = tf.nn.rnn_cell.MultiRNNCell([cell() for _ in range(layer_size)])
cell_fw_initial = cell_fw.zero_state(batch_size, tf.float32)
cell_bw_initial = cell_bw.zero_state(batch_size, tf.float32)
output = tf.nn.bidirectional_dynamic_rnn(cell_fw=cell_fw,
cell_bw=cell_bw,
inputs=inputs,
initial_state_fw=cell_fw_initial,
initial_state_bw=cell_bw_initial,
sequence_length=seq_len,
dtype=tf.float32)
return output
def _multihop_layer(self, q):
with tf.variable_scope("MultihopAttention_layer", reuse=False):
'''
主要是为了更新o_q(k)
1. o_q[0] = sum(h_q(t)) / l
2. m_q[0] = sum(h_q(t)) / l
3. M = tanh(W_q(k).h_q(t))
4. N = tanh(W_m(k).m_q(k-1))
5. s_t = tanh(W_q(k).h_q(t)) * tanh(W_m(k).m_q(k-1))
6. w = w_s(k).s_t(k)
7. alpha_t = softmax(w_s(k).s_t(k))
8. o_q[1] = sum( alpha_t * h_q(t))
'''
# num_steps = 1的情况
m_q = [None] * self.num_steps
m_q[0] = tf.reduce_mean(q, axis=1) # (bz, 2hz)
o_q = [None] * (self.num_steps + 1)
o_q[0] = tf.reduce_mean(q, axis=1) # (bz, 2hz)
M = tf.tanh(multiply_3_2(q, self.W_q1)) # (bz, len, 2hz)
N = tf.expand_dims(tf.tanh(tf.matmul(m_q[0], self.W_m1)), axis=1) # (bz, 1, 2hz)
s_t = tf.multiply(M, N) # (bz, len, 2hz)
w = multiply_3_2(s_t, self.W_s1) # (bz, len, 1)
alpha_t = tf.nn.softmax(w, axis=1) # (bz, len, 1)
o_q[1] = tf.reduce_sum(tf.multiply(q, alpha_t), axis=1) # (bz, 2hz)
if self.num_steps > 1: # num_steps = 2的情况
m_q[1] = m_q[0] + o_q[1] # (bz, 2hz)
M = tf.tanh(multiply_3_2(q, self.W_q2)) # (bz, len, 2hz)
N = tf.expand_dims(tf.tanh(tf.matmul(m_q[1], self.W_m2)), axis=1) # (bz, 1, 2hz)
s_t = tf.multiply(M, N) # (bz, len, 2hz)
w = multiply_3_2(s_t, self.W_s2) # (bz, len, 1)
alpha_t = tf.nn.softmax(w, axis=1) # (bz, len, 1)
o_q[2] = tf.reduce_sum(tf.multiply(q, alpha_t), axis=1) # (bz, 2hz)
if self.num_steps > 2: # num_steps = 3的情况
m_q[2] = m_q[1] + o_q[2] # (bz, 2hz)
M = tf.tanh(multiply_3_2(q, self.W_q3)) # (bz, len, 2hz)
N = tf.expand_dims(tf.tanh(tf.matmul(m_q[2], self.W_m3)), axis=1) # (bz, 1, 2hz)
s_t = tf.multiply(M, N) # (bz, len, 2hz)
w = multiply_3_2(s_t, self.W_s3) # (bz, len, 1)
alpha_t = tf.nn.softmax(w, axis=1) # (bz, len, 1)
o_q[3] = tf.reduce_sum(tf.multiply(q, alpha_t), axis=1) # (bz, 2hz)
return o_q
def _sequential_layer(self, inputs1, inputs2, ans_mask, rnn_size, batch_size, seq_len, dropout_keep_prob, index,
isreuse = False):
"""
论文地址：
.代表矩阵乘法
1. Gama = Gama_1 = j *W.h_i
2. outputs_n = n_i = GRU(n_i, gama_i)
3. n_ = n_i = concat(n_i, n_i)
4. alpha = alpha_i = softmax(1(T).n_i)
5. partial_bilinear_output = O = sum(alpha * h_)
6. pred = a = argmax(M_a.O) 在本模型中，这一步取消
"""
j = tf.expand_dims(inputs1, axis=1) # (bz, 1, 2hz)
h = inputs2 # (bz, len, 2hz)
# W_h = tf.matmul(h, self.W)
Gama = tf.multiply(j, h) # (bz, len, 2hz)
print_shape('Gama', Gama)
outputs_n, finalState_n = self._bilstm_layer(Gama, rnn_size, seq_len, 1, batch_size, 1.0,
'sequentional_attention_{}'.format(index), reuse=isreuse)
n_ = tf.concat(outputs_n, axis=2) # (bz, len, 2hz)
alpha = tf.nn.softmax(tf.reduce_sum(n_, axis=2), axis=-1) # (bz, len)
# 将句子padding部分权重消除
# alpha = tf.expand_dims(tf.multiply(alpha, ans_mask), axis=-1)
alpha = tf.expand_dims(alpha, axis=-1)
partial_bilinear_output = tf.reduce_sum(tf.multiply(alpha, h), axis=1) # (bz, 2hz)
print_shape('SequentialAttention_output', partial_bilinear_output)
return partial_bilinear_output
def _match_layer(self, q, a_pos, a_neg, p_mask, n_mask, rnn_size, batch_size, p_seq_len, n_seq_len, dropout_keep_prob):
for i in range(self.num_steps+1):
pos = self._sequential_layer(q[i], a_pos, p_mask, rnn_size, batch_size, p_seq_len, dropout_keep_prob, i+1,
isreuse=False)
neg = self._sequential_layer(q[i], a_neg, n_mask, rnn_size, batch_size, n_seq_len, dropout_keep_prob, i+1,
isreuse=True)
q_ = tf.nn.l2_normalize(q[i], dim=1)
as_pos = tf.nn.l2_normalize(pos, dim=1)
as_neg = tf.nn.l2_normalize(neg, dim=1)
q_pos_cosine = tf.reduce_sum(tf.multiply(q_, as_pos), 1)
q_neg_cosine = tf.reduce_sum(tf.multiply(q_, as_neg), 1)
sim_pos = q_pos_cosine if i == 0 else sim_pos + q_pos_cosine
sim_neg = q_neg_cosine if i == 0 else sim_neg + q_neg_cosine
return sim_pos, sim_neg
def _build(self, embeddings, rnn_type = 'lstm'):
self.Embedding = tf.Variable(tf.to_float(embeddings), trainable=False, name='Embedding')
self.q_embed = tf.nn.dropout(tf.nn.embedding_lookup(self.Embedding, self._ques), keep_prob=self.dropout_keep_prob)
self.a_embed = tf.nn.dropout(tf.nn.embedding_lookup(self.Embedding, self._ans), keep_prob=self.dropout_keep_prob)
self.a_neg_embed = tf.nn.dropout(tf.nn.embedding_lookup(self.Embedding, self._ans_neg), keep_prob=self.dropout_keep_prob)
q_mask = tf.sequence_mask(self._ques_mask, self.ques_len, dtype=tf.float32)
a_mask = tf.sequence_mask(self._ans_mask, self.ans_len, dtype=tf.float32)
a_neg_mask = tf.sequence_mask(self._ans_neg_mask, self.ans_len, dtype=tf.float32)
# 上下文编码
if rnn_type == 'lstm':
q_outputs, q_final_state = self._bilstm_layer(self.q_embed, self.rnn_size, self._ques_mask, self.layer_size,
self.batch_size, self.dropout_keep_prob, 'lstm')
a_outputs, a_final_state = self._bilstm_layer(self.a_embed, self.rnn_size, self._ans_mask, self.layer_size,
self.batch_size, self.dropout_keep_prob, 'lstm',
reuse=True)
a_neg_outputs, a_neg_final_state = self._bilstm_layer(self.a_embed, self.rnn_size, self._ans_neg_mask,
self.layer_size, self.batch_size,
self.dropout_keep_prob, 'lstm', reuse=True)
elif rnn_type == 'gru':
q_outputs, q_final_state = self._bilstm_layer(self.q_embed, self.rnn_size, self._ques_mask, self.layer_size,
self.batch_size, self.dropout_keep_prob, 'gru')
a_outputs, a_final_state = self._bilstm_layer(self.a_embed, self.rnn_size, self._ans_mask, self.layer_size,
self.batch_size, self.dropout_keep_prob, 'gru',
reuse=True)
a_neg_outputs, a_neg_final_state = self._bilstm_layer(self.a_embed, self.rnn_size, self._ans_neg_mask,
self.layer_size, self.batch_size,
self.dropout_keep_prob, 'gru', reuse=True)
rnn_q = tf.concat(q_outputs, axis=-1)
rnn_a = tf.concat(a_outputs, axis=-1)
rnn_a_neg = tf.concat(a_neg_outputs, axis=-1)
o_q = self._multihop_layer(rnn_q)
sim_pos, sim_neg = self._match_layer(o_q, rnn_a, rnn_a_neg, a_mask, a_neg_mask, self.rnn_size, self.batch_size,
self._ans_mask, self._ans_neg_mask, self.dropout_keep_prob)
return sim_pos, sim_neg
def _margin_loss(self, pos_sim, neg_sim):
original_loss = self.margin - pos_sim + neg_sim
l = tf.maximum(tf.zeros_like(original_loss), original_loss)
loss = tf.reduce_sum(l)
return loss, l
def _add_loss_op(self, p_sim, n_sim, l2_lambda=0.0001):
"""
损失节点
"""
loss, l = self._margin_loss(p_sim, n_sim)
accu = tf.reduce_mean(tf.cast(tf.equal(0., l), tf.float32))
reg_losses = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
l2_loss = sum(reg_losses) * l2_lambda
pairwise_loss = loss + l2_loss
tf.summary.scalar('pairwise_loss', pairwise_loss)
self.summary_op = tf.summary.merge_all()
return pairwise_loss, accu
def _add_train_op(self, loss):
"""
训练节点
"""
with tf.name_scope('train_op'):
# 记录训练步骤
self.global_step = tf.Variable(0, name='global_step', trainable=False)
opt = tf.train.AdamOptimizer(self.learning_rate)
# train_op = opt.minimize(loss, self.global_step)
gradients, v = zip(*opt.compute_gradients(loss))
clip_gradients = gradients
if self.clip_value is not None:
clip_gradients, _ = tf.clip_by_global_norm(gradients, self.clip_value)
train_op = opt.apply_gradients(zip(clip_gradients, v), global_step= self.global_step)
return train_op