-
Notifications
You must be signed in to change notification settings - Fork 5
/
train_dssm.py
170 lines (147 loc) · 6.48 KB
/
train_dssm.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
#!/usr/bin/python
# -*- coding:utf-8 -*-
## @objection : try to implement the DSSM using tensorflow ##
## @reference : Learning Deep Structured Semantic Models for Web Search using Clickthrough Data
## @date : 2017-01-23
## @author : yujianmin
from __future__ import print_function
from __future__ import absolute_import
import tensorflow as tf
## data-batch get from this function ##
def get_query_doc_input(step): ## query_batch, doc_batch -> array ##
query_batch = np.ndarray()
doc_batch = np.ndarray()
## model struct ##
## input : 30k word-hash vector ##
## layer-1 : 300 hidden nodes ##
## layer-2 : 300 hidden nodes ##
## output : 128 output nodes ##
## input variables ##
input_node_num = 30000
layer1_node_num = 300
layer2_node_num = 300
output_node_num = 128
## notice : 英文可以用word-hash来限制输入维度的数量26个字母的n-letter组合，但对中文较为勉强，还是用tf-idf算了 ##
## notice : paper-author用的是query-title做的试验 ## 我们这里也用query-title 对来做输入 ##
## 先用稀疏占位符 构造输入数据的接收变量 ##
## 稀疏矩阵 ##
query_input = tf.sparse_placeholder(shape = [None, input_node_num], dtype = tf.float32, name = 'query_batch')
doc_input = tf.sparse_placeholder(shape = [None, input_node_num], dtype = tf.float32, name = 'doc_batch')
## middle variables ##
## layer-1 ## output ##
max_min_threshold = tf.sqrt(6*(input_node_num + output_node_num))
layer1_w = tf.Variable(tf.random_uniform(
shape = [input_node_num, layer1_node_num],
minval = -max_min_threshold,
maxval = +max_min_threshold),
name = 'layer1_w',
dtype = tf.float32)
layer1_b = tf.Variable(tf.random_uniform(
shape = [layer1_node_num],
minval = -max_min_threshold,
maxval = +max_min_threshold),
name = 'layer1_b',
dtype = tf.float32)
query_1_output = tf.nn.relu(
tf.sparse_tensor_dense_matmul(query_input, layer1_w) + layer1_b
)
doc_1_output = tf.nn.relu(
tf.sparse_tensor_dense_matmul(doc_input, layer1_w) + layer1_b
)
## layer-2 ## output ##
layer2_w = tf.Variable(tf.random_uniform(
shape = [layer1_node_num, layer2_node_num],
minval = -max_min_threshold,
maxval = +max_min_threshold),
name = 'layer2_w',
dtype = tf.float32)
layer2_b = tf.Variable(tf.random_uniform(
shape = [layer2_node_num],
minval = -max_min_threshold,
maxval = +max_min_threshold),
name = 'layer2_b',
dtype = tf.float32)
query_2_output = tf.nn.relu(
tf.matmul(query_1_output, layer2_w) + layer2_b
)
doc_2_output = tf.nn.relu(
tf.matmul(doc_1_output, layer2_w) + layer2_b
)
## layer-3 and output ##
layer3_w = tf.Variable(tf.random_uniform(
shape = [layer2_node_num, output_node_num],
minval = -max_min_threshold,
maxval = +max_min_threshold),
name = 'layer3_w',
dtype = tf.float32)
layer3_b = tf.Variable(tf.random_uniform(
shape = [output_node_num],
minval = -max_min_threshold,
maxval = +max_min_threshold),
name = 'layer3_b',
dtype = tf.float32)
query_output = tf.nn.relu(
tf.matmul(query_2_output, layer3_w) + layer3_b
)
doc_output = tf.nn.relu(
tf.matmul(doc_2_output, layer3_w) + layer3_b
)
## optimal-function ##
# loss = - tf.sum( tf.log(P(D|Q))) min
# P(D|Q) = tf.nn.softmax(similar_matrix(Q&D))
## once a batch ：shape = m * input_node_num
# [ input ] [ output ]
# [ sample-1 : query|doc ] [ f(sample-1) ]
# [ sampel-2 : query|doc ] == > [ f(sample-2) ]
# [ sample-3 : query|doc ] [ f(sampel-3) ]
# [ sample-4 : query|doc ] [ f(sampel-4) ]
## current time ##
## Q&D 的相似度 矩阵，用于计算 softmax ## P(D|Q) = R(D, Q) / sum( R(Q, Di)) ## 所有文档中，当前分布所占概率 ##
## 故相似矩阵如下 ：shape = m * m
# [ R(D1, Q1), R(D2, Q1), R(D3, Q1), ... , R(Dm, Q1) ]
# [ R(D1, Q2), R(D2, Q2), R(D3, Q2), ... , R(Dm, Q2) ]
# [ R(D1, Q3), R(D2, Q3), R(D3, Q3), ... , R(Dm, Q3) ]
# [ R(D1, Q4), R(D2, Q4), R(D3, Q4), ... , R(Dm, Q4) ]
## 其中 R(D, Q) = (query_output * doc_output.T) / (|query_output| * |doc_output|)
# query_input.shape = m * input_node_num
# query_output.shape = m * 128
# doc_input.shape = m * input_node_num
# doc_output.shape = m * 128
## 故计算相似矩阵 如下 ：
# 分子 = tf.matmul(query_output, doc_output, transpose_b=True) shape = m * m
# 分母 = tf.mul(query_norm.tile([1, m]), doc_norm.tile([1, m]).T)
# query_norm = tf.sqrt( tf.reduce_sum(query_output, reduction_indices=0, keep_dim=True))
# doc_norm = tf.sqrt( tf.reduce_sum(doc_output, reduceion_indices=0, keep_dim=True))
similar_fenzi = tf.matmul(query_output, doc_output, transpose_b = True)
query_norm = tf.sqrt(tf.reduce_sum(query_output, reduction_indices = 0, keep_dim = True)
doc_norm = tf.sqrt(tf.reduce_sum(doc_output , reduction_indices = 0, keep_dim = True)
similar_fenmu = tf.mul(query_norm.tile([1, m]), tf.transpose(doc_norm.tile([1, m])))
similar_matrix= tf.truediv(similar_fenzi, similar_fenmu)
prob = tf.nn.softmax(similar_matrix) ## P(D|Q) ##
## 取出对角线的概率来，这些是成对的query-doc的条件概率 ##
Batch_num= tf.shape(query_input)[0]
one_diag = tf.diag(tf.ones(shape = [Batch_num]))
pair_pro = tf.mul(prob, one_diag) ## element-mul ## 这样就取出了query-doc对的条件概率 ## 非对角位置都是 0 ##
loss = -tf.reduce_mean(tf.log(pair_pro), [0,1]) ## 对那些成对的query-doc, max{sum(log(P(D|Q)))}，即min{-sum}
learn_step = 0.001
train_step = tf.train.GradientDescentOptimizer(learn_step).minimize(loss)
## 训练效果评估 ##
accuracy_bool = tf.equal(
tf.argmax(prob, dimension=1),
tf.constant(value = range(Batch_num), dtype = tf.int64)) ## 匹配y_real; y_predict ##
accuracy_ratio = tf.reduce_mean(tf.cast(accuracy_bool, dtype = tf.int64))
## train the model ##
train_max_num = 100
with tf.Session() as sess:
sess.run(tf.initialize_all_variables())
for step in range(train_max_num):
cur_query_input, cur_doc_input = get_query_doc_input(step)
sess.run(train_step, feed_dict = {
query_input : cur_query_input,
doc_input : cur_doc_input})
train_accuracy = sess.run(accuracy_ratio, feed_dict = {
query_input : cur_query_input,
doc_input : cur_doc_input})
print ('step : ', step, ', its train accracy is : ', train_accuracy)
## 测试集效果评估 ##
sess.close()