-
Notifications
You must be signed in to change notification settings - Fork 3
/
nnqlm.py
226 lines (166 loc) · 7.69 KB
/
nnqlm.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
#coding=utf8
from __future__ import division
import numpy as np
import pandas as pd
import cPickle as Pickle
import os, time, sys
import ConfigParser
from sklearn.utils import shuffle
from sklearn import linear_model, metrics
import theano
import theano.tensor as T
import theano.tensor.shared_randomstreams
from theano.tensor.signal import downsample
from util import batch_num, mrr_metric, map_metric, share_x, share_y
from nnlayer import HiddenLayer, LogisticRegression, CNNLayer, Density, Density_Dot
cf = ConfigParser.ConfigParser()
cf.read('config.conf')
_, _, trainpara, testpara, gobal, _, _, args = cf.sections()
if __name__ == '__main__':
trainfile = cf.get(trainpara, 'trainfile')
testfile = cf.get(testpara, 'testfile')
df_train= pd.read_csv(trainfile, header=None,sep="\t",names=['qid','aid',"question","answer","flag"],quoting =3)
df_test= pd.read_csv(testfile, header=None,sep="\t",names=['qid','aid',"question","answer","flag"],quoting =3)
learning_rate = cf.getfloat(args, 'learning_rate')
batch_size = cf.getint(args, 'batch_size')
filter_row = cf.getint(args, 'filter_row')
filter_num = cf.getint(args, 'filter_num')
reg_rate = cf.getfloat(args, 'reg_rate')
itera_number = cf.getint(args, 'itera_number')
print 'learning_rate is: ', learning_rate
print 'batch_size is: ', batch_size
print 'step is: ', filter_row
print 'feature map is: ', filter_num
print 'reg_rate is', reg_rate
logfile = cf.get(gobal, 'log')
qlen = cf.getint(gobal, 'qlen')
alen = cf.getint(gobal, 'alen')
vocab = cf.get(gobal, 'vocab')
log = open(logfile, 'a')
pen_rate = 0
trainlabel = share_y(cf.get(trainpara, 'label'), batch_size)
trainquestion = share_x(cf.get(trainpara, 'question_index'), batch_size)
trainanswer = share_x(cf.get(trainpara, 'answer_index'), batch_size)
trainqlens = share_x(cf.get(trainpara, 'qlens'), batch_size)
trainalens = share_x(cf.get(trainpara, 'alens'), batch_size)
testlabel = share_y(cf.get(testpara, 'label'), batch_size)
testquestion = share_x(cf.get(testpara, 'question_index'), batch_size)
testanswer = share_x(cf.get(testpara, 'answer_index'), batch_size)
testqlens = share_x(cf.get(testpara, 'qlens'), batch_size)
testalens = share_x(cf.get(testpara, 'alens'), batch_size)
print 'data sets have been shared'
embeddings = Pickle.load(open(vocab, 'rb'))
embedding_ndim = len(embeddings[0])
filter_col = filter_row
train_sample_num = len(df_train)
test_sample_num = len(df_test)
train_batch_num = batch_num(df_train, batch_size)
test_batch_num = batch_num(df_test, batch_size)
channel_num = 1
class_num = 2
image_shape = (batch_size, channel_num, embedding_ndim, embedding_ndim)
filter_shape = (filter_num, channel_num, filter_row, filter_col)
rng = np.random.RandomState(23455)
index = T.lscalar('index')
x_q = T.matrix('x_q')
x_a = T.matrix('x_a')
y = T.ivector('y')
len_qs = T.vector('len_qs')
len_as = T.vector('len_as')
Embeddings = theano.shared(value = embeddings, name = 'Embeddings', borrow=True)
layer0_q = Embeddings[T.cast(x_q.flatten(),dtype="int32")].\
reshape((batch_size, qlen, embedding_ndim))
layer0_a = Embeddings[T.cast(x_a.flatten(),dtype="int32")].\
reshape((batch_size, alen, embedding_ndim))
layer1_q = Density(rng, layer0_q, len_qs, qlen)
layer1_a = Density(rng, layer0_a, len_as, alen)
layer2 = Density_Dot(layer1_q.output, layer1_a.output)
layer3 = CNNLayer(rng, input = layer2.output.reshape(image_shape), \
filter_shape = filter_shape, image_shape = image_shape)
layer4_q = layer3.output.max(axis = 3).flatten(2)
layer4_a = layer3.output.max(axis = 2).flatten(2)
Logistic_in_num = 2 * int(embedding_ndim - filter_row + 1) * filter_num
Logistic_input = T.concatenate([layer4_q, layer4_a], axis = 1)
layer5 = LogisticRegression(Logistic_input, Logistic_in_num, class_num)
params = layer1_q.params + layer1_a.params + layer3.params + layer5.params
reg_term = (layer1_q.params[0] ** 2).sum() + (layer1_a.params[0] ** 2).sum() \
+ (layer3.params[0] ** 2).sum() + (layer5.params[0] ** 2).sum()
# reg_term is the penalty for the complexity of the model
pen_term = ((layer1_q.params[0] ** 2) ** 0.5).sum() + ((layer1_a.params[0] ** 2)**0.5).sum() - 2
# pen_term is the penalty that the trace of the density matrix is not equal 1
cost = layer5.negative_log_likelihood(y) + reg_rate * reg_term + pen_rate * pen_term
grads = T.grad(cost, params)
updates = [(param_i, param_i - learning_rate * grad_i) for param_i, grad_i in zip(params, grads)]
given_train = {x_q: trainquestion[index * batch_size : (index + 1) * batch_size], \
x_a: trainanswer[index * batch_size : (index + 1) * batch_size], \
len_qs: trainqlens[index * batch_size : (index + 1) * batch_size], \
len_as: trainalens[index * batch_size : (index + 1) * batch_size], \
y: trainlabel[index * batch_size : (index + 1) * batch_size]}
given_test = {x_q: testquestion[index * batch_size : (index + 1) * batch_size], \
x_a: testanswer[index * batch_size : (index + 1) * batch_size], \
len_qs: testqlens[index * batch_size : (index + 1) * batch_size], \
len_as: testalens[index * batch_size : (index + 1) * batch_size], \
y: testlabel[index * batch_size : (index + 1) * batch_size]}
train_model = theano.function([index], cost, updates = updates, givens = given_train, on_unused_input='ignore', allow_input_downcast=True)
train_score_model = theano.function([index], layer5.y_score, givens = given_train, on_unused_input='ignore', allow_input_downcast=True)
test_model = theano.function([index], layer5.y_score, givens = given_test, on_unused_input='ignore', allow_input_downcast=True)
train_model_loss = theano.function([index], cost, givens = given_train, on_unused_input='ignore', allow_input_downcast=True)
test_model_loss = theano.function([index], cost, givens = given_test, on_unused_input='ignore', allow_input_downcast=True)
print "begin to train"
test_mrr = 0.
test_map = 0.
for itera_i in range(itera_number):
print 'The ', itera_i, 'epcho begin'
count_i = 0
train_score = []
for i in shuffle(range(train_batch_num), random_state = 121):
# for i in range(train_batch_num):
train_model(i)
if count_i % 10 == 0:
print count_i, 'batch train data have been trained'
count_i += 1
count_i = 0
for i in range(train_batch_num):
score = train_score_model(i)
for j in score:
train_score.append(j[1])
if count_i % 10 == 0:
print count_i, 'batch train data have been tested, loss is ', train_model_loss(i)
count_i += 1
train_score = train_score[:train_sample_num]
df_train['score'] = train_score
train_flag = df_train['flag']
roc = metrics.roc_auc_score(train_flag, train_score)
print 'The roc of train is: ', roc*100
test_score = []
count_i = 0
for i in range(test_batch_num):
score = test_model(i)
for j in score:
test_score.append(j[1])
if count_i % 10 == 0:
print count_i, 'batch test data have been tested, loss is ', test_model_loss(i)
count_i += 1
test_score = test_score[:test_sample_num]
df_test['score'] = test_score
test_flag = df_test['flag']
roc = metrics.roc_auc_score(test_flag, test_score)
print 'The roc of test is: ', roc*100
test_mrr = df_test.groupby('question').apply(mrr_metric).mean()
test_map = df_test.groupby('question').apply(map_metric).mean()
print test_mrr
print test_map
log.write('learning_rate ' + str(learning_rate) + ', ')
log.write('batch_size ' + str(batch_size) + ', ')
log.write('filter_num ' + str(filter_num) + ', ')
log.write('filter_row ' + str(filter_row) + ', ')
log.write('reg_rate ' + str(reg_rate) + ' :\n')
print 'end train'
print 'the mrr is : ', test_mrr
print 'the map is : ', test_map
log.write('the mrr is : ')
log.write(str(test_mrr) + '\t')
log.write('the map is : ')
log.write(str(test_map) + '\t')
log.write('\n\n')
log.close()