-
Notifications
You must be signed in to change notification settings - Fork 0
/
CNN_on_NLP.py
289 lines (247 loc) · 14.8 KB
/
CNN_on_NLP.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
280
281
282
283
284
285
286
287
288
289
import tensorflow as tf
import numpy as np
import ReadCsv as read_utils
import utils
from tensorflow import graph_util
class CNN(object):
def __init__(self):
self.LENTH_MAX = 20
self.LOAD_SIZE = 4985 ## 由于语料原因，最多只能是4985
self.VOCABULARY_SIZE = 10000
self.EMBEDING_SIZE = 128
self.BATCH_SIZE = 50
self.KEEP_PROB = 0.7
self.LEARNING_RATE = 0.01
self.CLASS_NUM = 3 ## '好评':0, '中评':1, '差评':2
self.MAX_INTERATION = 20
self.GUANLIAN = 5
def interface(self, input, keep_prob):
with tf.variable_scope("Layer_1"): ## 定义命名域
W1 = utils.weight_variable([5, 5, 1, 32], name="W1") ## 构造变量W1
b1 = utils.bias_variable([32], name="b1")
conv_1 = utils.conv2d_basic(input, W1, b1) ## 卷积
relu_1 = tf.nn.relu(conv_1, name="relu_1") ## 使用激活函数relu
## pool_1: size [BATCH_SIZE, 32, 64, 32]
pool_1 = utils.avg_pool_2x2(relu_1) ## 对激活函数输出进行平均卷积
with tf.variable_scope("Layer_2"):
W2 = utils.weight_variable([5, 5, 32, 256], name="W2")
b2 = utils.bias_variable([256], name="b2")
conv_2 = utils.conv2d_basic(pool_1, W2, b2)
relu_2 = tf.nn.relu(conv_2, name='relu_2')
## pool_2: size [BATCH_SIZE, 16, 32, 256]
pool_2 = utils.avg_pool_2x2(relu_2)
# scope.reuse_variables()
with tf.variable_scope("Layer_3"):
W3 = utils.weight_variable([3, 3, 256, 384], name="W3")
b3 = utils.bias_variable([384], name="b3")
conv_3 = utils.conv2d_basic(pool_2, W3, b3)
relu_3 = tf.nn.relu(conv_3, name="relu_3")
## pool_3: size [BATCH_SIZE, 8,16,384]
pool_3 = utils.avg_pool_2x2(relu_3)
with tf.variable_scope("Layer_4"):
W4 = utils.weight_variable([3, 3, 384, 384], name="W4")
b4 = utils.bias_variable([384], name="b4")
conv_4 = utils.conv2d_basic(pool_3, W4, b4)
relu_4 = tf.nn.relu(conv_4, name="relu_4")
## pool_4: size [BATCH_SIZE, 4,8, 384]
pool_4 = utils.avg_pool_2x2(relu_4)
with tf.variable_scope("Layer_5"):
W5 = utils.weight_variable([3, 3, 384, 256], name="W5")
b5 = utils.weight_variable([256], name="b5")
conv_5 = utils.conv2d_basic(pool_4, W5, b5)
relu_5 = tf.nn.relu(conv_5, name="relu_5")
## pool_5: size [BATCH_SIZE, 2, 4, 256]
pool_5 = utils.avg_pool_2x2(relu_5)
with tf.variable_scope("all_link"): ### 全连接层
W_fc1 = utils.weight_variable([1 * 4 * 256, 4096], name="W_fc1")
b_fc1 = utils.bias_variable([4096], name="b_fc1")
pool_5_flag = tf.reshape(pool_5, [-1, 1 * 4 * 256])
## h_fc1: size [BATCH_SIZE, 4096]
h_fc1 = tf.nn.relu(tf.nn.xw_plus_b(pool_5_flag, W_fc1, b_fc1))
h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)
W_fc2 = utils.weight_variable([4096, self.CLASS_NUM], name="W_fc2")
b_fc2 = utils.bias_variable([self.CLASS_NUM], name="b_fc2")
h_fc2 = tf.nn.tanh(tf.nn.xw_plus_b(h_fc1_drop, W_fc2, b_fc2))
# h_fc2_drop = tf.nn.dropout(h_fc2, keep_prob)
return h_fc2
def interface_column(self, input, keep_prob): ## 全行扫描（及卷积核单向移动）网络
lenth = input.shape[2] ## 提取输入矩阵的长度
layers = [] ## 用于统计不同呢的卷积核卷积出来的结果，方便后来的合成
guanlian = self.GUANLIAN-2 ## 关联长度，也就是指卷积核最多框住多少个词，最少2个，最大为：self.GUANLIAN
with tf.variable_scope('Layer_1'):
for i in range(2, self.GUANLIAN):
name = 'conv_'+str(i)
kernels = tf.Variable(tf.random_normal(shape = [i, self.EMBEDING_SIZE, 1, 1]))
# print(i)
# print(kernels.shape)
bias = tf.constant(0.0, shape=[1])
## current.shape = [BATCH_SIZE, LENTH_MAX, 1, 1]
## 横向移动距离为lenth，即矩阵本身长度，保证不会左右平移，只会在列方向移动
current = tf.nn.conv2d(input, kernels, strides = [1, 1, lenth, 1], padding='SAME', name = name)
# 把计算出来的这一层保存起来，便于后面的激活函数预算与全连接
layers.append(current)
## layer_1.shape = [BATCH_SIZE, LENTH_MAX, 1, GUANLIAN]
layer_1 = tf.concat(layers, axis=3)
relu = tf.nn.relu(layer_1, name='relu')
with tf.variable_scope('aver_pooling'):
size = relu.shape
## pool.shape = [BATCH_SIZE, 1, 1, GUANLIAN]
pool = tf.nn.avg_pool(relu, ksize=[1, size[1], size[2], 1], strides=[1, size[1], size[2], 1], padding='SAME')
with tf.variable_scope("all_link"):
h_fc1_drop = tf.nn.dropout(pool, keep_prob)
W_fc1 = utils.weight_variable([guanlian, self.CLASS_NUM], name="W_fc1")
b_fc1 = utils.bias_variable([self.CLASS_NUM], name="b_fc1")
pool_5_flag = tf.reshape(h_fc1_drop, [self.BATCH_SIZE, -1])
# print(relu.shape)
# print(W_fc1.shape)
## h_fc1: size [BATCH_SIZE, 4096]
h_fc1 = tf.nn.relu(tf.nn.xw_plus_b(pool_5_flag, W_fc1, b_fc1))
return h_fc1
def CNNTrain(self, mode, BATCH_SIZE=None, text = None):
"""
Args:
mode = "train" or "use"
"""
if(BATCH_SIZE ==None) & (mode =="train"):
BATCH_SIZE = self.BATCH_SIZE
elif (BATCH_SIZE==None):
BATCH_SIZE = 1
# BATCH_SIZE = self.BATCH_SIZE
text_input = tf.placeholder(
dtype=tf.int32, shape=[BATCH_SIZE, self.LENTH_MAX], name="text_input")
text_input_change = tf.reshape(text_input, [BATCH_SIZE, -1], name="Input") ## 如果有3维向量的时候就需要reshape了，此处没有变化
keep_prob = tf.placeholder(dtype=tf.float32, name="keep_prob")
label = tf.placeholder(dtype=tf.int64, shape=[BATCH_SIZE, ], name="label")
embeding_var = tf.Variable(tf.random_uniform(
shape=[self.VOCABULARY_SIZE, self.EMBEDING_SIZE]), dtype=tf.float32, name='embeding_var')
## batch_embeding: size [BATCH_SIZE, LENTH_MAX, EMBEDING_SIZE]
batch_embeding = tf.nn.embedding_lookup(embeding_var, text_input_change)
batch_embeding_normal = tf.reshape(
batch_embeding, [-1, self.LENTH_MAX, self.EMBEDING_SIZE, 1])
print(batch_embeding_normal)
# output = self.interface(batch_embeding_normal, keep_prob)
output = self.interface_column(batch_embeding_normal, keep_prob)
# print(output)
# print(label)
with tf.variable_scope("loss"):
loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=output, labels=label, name="loss"))
with tf.variable_scope("accuracy"):
# accuracy = tf.reduce_mean(
# tf.cast(tf.equal(tf.round(tf.sigmoid(output)), label), dtype=tf.float32))
accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(tf.sigmoid(output),1), label), dtype = tf.float32))
result = tf.nn.softmax(output, name='result')
# print(result)
with tf.variable_scope("train_op"):
train_op = tf.train.AdadeltaOptimizer(self.LEARNING_RATE).minimize(loss)
saver = tf.train.Saver()
if mode == 'train':
# print("output:")
# print(output)
tf.summary.scalar('loss', loss)
tf.summary.scalar('accuracy', accuracy)
summary_op = tf.summary.merge_all()
accuracy_average = 0
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
train_datareader = read_utils.ReadData(
self.LENTH_MAX, mode="train", load_size=self.LOAD_SIZE, vocabulary_size=self.VOCABULARY_SIZE)
train_datareader.Train_Init()
test_datareader = read_utils.ReadData(
self.LENTH_MAX, mode="test", load_size=self.LOAD_SIZE, vocabulary_size=self.VOCABULARY_SIZE)
test_datareader.Train_Init()
summary_writer = tf.summary.FileWriter("./logs", sess.graph)
for itr in range(self.MAX_INTERATION):
x, y = train_datareader.NextBatch(BATCH_SIZE)
feed_dict = {text_input: x, label: y, keep_prob: self.KEEP_PROB}
sess.run(train_op, feed_dict)
# print(itr)
if itr % 10 == 0:
loss_train, accuracy_train, summary_str = sess.run(
[loss, accuracy, summary_op], feed_dict)
print("Step: %d , Train_loss: %g ,Train_accuracy: %g" %
(itr, loss_train, accuracy_train)) #### 每10次的训练后进行输出的语句
accuracy_average += accuracy_train
summary_writer.add_summary(summary_str, itr)
if itr % 100 == 0:
x_test, y_test = test_datareader.NextBatch(BATCH_SIZE)
loss_test, accuracy_test = sess.run([loss, accuracy], feed_dict={
text_input: x_test, label: y_test, keep_prob: 1})
print("Step: %d , loss_loss: %g ,loss_accuracy: %g" %
(itr, loss_test, accuracy_test)) ### 每100次训练后对测试集进行测试输出的语句
saver.save(sess, "./log/CNN-model.ckpt")
# saver2.save
# tf.get_default_graph()返回当前会话的默认图
# as_graph_def()返回一个图的序列化的GraphDef表示
# 序列化的GraphDef可以导入至另一个图中(使用 import_graph_def())
# 或者使用C++ Session API
graph_def = tf.get_default_graph().as_graph_def()
# graph_util模块的convert_variables_to_constants(
# sess, ----------------变量所在的会话
# input_graph_def, -------------持有需要保存的网络结构的GraphDef对象
# output_node_names, -------------需要保存的节点名称，注意命名域[scope_name/op_name]
# variable_names_whitelist=None, --要转换的变量名(在默认情况下，所有的变量都被转换)。
# variable_names_blacklist=None ---变量名的集合，省略转换为常量
# )
output_graph = graph_util.convert_variables_to_constants(sess, graph_def, ['result'])
# output_graph = graph_util.remove_training_nodes(graph_def, ['all/result'])
# output_graph = graph_util.extract_sub_graph(graph_def, ['all/result'])
with tf.gfile.GFile("./log/combined_model.pb","wb") as f:
f.write(output_graph.SerializeToString())
num = self.MAX_INTERATION/10
accuracy_average = accuracy_average/num
print("accuracy_average: %g "%accuracy_average)
else:
with tf.Session() as sess:
saver.restore(sess, "./log/CNN-model.ckpt")
train_datareader = read_utils.ReadData(
self.LENTH_MAX, mode="train", load_size=self.LOAD_SIZE, vocabulary_size=self.VOCABULARY_SIZE)
x = train_datareader.Test(text)
print(x)
feed_dict = {text_input:x, keep_prob: 1 }
print(result.eval(feed_dict))
def test(self, text):
"""
Args:
text:需要进行预测情感分类的文本
Return：
result：文本分类的3中可能的概率，'好评':0, '中评':1, '差评':2
"""
model_filename = "./log/combined_model.pb"
# 先是新建一个图
graph = tf.Graph()
with graph.as_default(): ## 把新建的图作为默认图
with tf.gfile.FastGFile(model_filename, 'rb') as f:
graph_def = tf.GraphDef() ## 新建GraphDef类对象，用于存储图
graph_def.ParseFromString(f.read()) ## 解析pb文件的GraphDef对象
train_datareader = read_utils.ReadData(
self.LENTH_MAX, mode="train", load_size=self.LOAD_SIZE, vocabulary_size=self.VOCABULARY_SIZE)
textnum = train_datareader.Test(text) ### 对输入文本进行规则化处理，编码
feed_dict = {'text_input':textnum, 'keep_prob':1.0}
result=tf.import_graph_def(graph_def, input_map=feed_dict, return_elements=['result:0']) ##无返回类型的操作，把文件中的graph导入默认graph
# import_graph_def(
# graph_def, --------pb文件解析出来的GraphDef对象
# input_map=None, -----需要映射到图中op的值，这里是用来填放feed_dict的
# return_elements=None,--网络需要返回的值
# name=None, --这个操作的名称
# op_dict=None, --已经弃用的选项
# producer_op_list=None --
# )
with tf.Session(graph = graph) as sess:
last = sess.run(result)
print(last[0][0])
return(last[0][0])
if __name__=="__main__":
# with tf.variable_scope('all') as scope:
# test = CNN()
# print(type(test))
# test.CNNTrain(mode = "train")
# test.test("包装严实物流也很快，下次还会再来")
# test.test("非常感谢京东商城给予的优质的服务，从仓储管理、物流配送等各方面都是做的非常好的。送货及时，配送员也非常的热情，有时候不方便收件的时候，也安排时间另行配送。同时京东商城在售后")
# test.CNNTrain(mode = "test", text="包装严实物流也很快，下次还会再来")
# scope.reuse_variables()
# test.CNNTrain(mode = "test", text="非常感谢京东商城给予的优质的服务，从仓储管理、物流配送等各方面都是做的非常好的。送货及时，配送员也非常的热情，有时候不方便收件的时候，也安排时间另行配送。同时京东商城在售后")
test = CNN()
# test.CNNTrain(mode="train")
test.test("包装严实物流也很快，下次还会再来")
test.test("不好，宝宝都不喜欢喝，二阶段的宝宝就喝，这个浪费了。")