This repository has been archived by the owner on Nov 23, 2022. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 8
/
load_model.py
158 lines (125 loc) · 6.62 KB
/
load_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
import tensorflow as tf
import os
class Load_model:
def __init__(self, **kwargs):
# 模型的参数
self.lstmUnits = 64
self.attentionSize = 128
self.dropOutRate = 1.0
self.learn_rate = 0.0003
self.numClasses = 4
# 数据的参数
self.maxSeqLength = 350
self.numDimensions = 300
self.iterations = kwargs['iter']
self.label = kwargs['label']
self.id = kwargs['id']
self.model_path = './model/label_' + str(self.label) + '/' + str(self.id) + '/'
if os.path.exists('./model/') is False:
os.mkdir('./model/')
if os.path.exists('./model/label_' + str(self.label)) is False:
os.mkdir('./model/label_' + str(self.label))
if os.path.exists(self.model_path) is False:
os.mkdir(self.model_path)
self.graph = self.init_graph()
def init_graph(self):
graph = tf.Graph()
with graph.as_default():
with tf.variable_scope('Inputs'):
labels = tf.placeholder(tf.float32, [None, self.numClasses], name='labels')
input_data = tf.placeholder(tf.float32, [None, self.maxSeqLength, self.numDimensions],
name='input_data')
input_length = tf.placeholder(tf.int32, [None], name='input_length')
with tf.variable_scope('BiLSTM'):
encoder_fw = tf.nn.rnn_cell.BasicLSTMCell(self.lstmUnits)
encoder_fw = tf.nn.rnn_cell.DropoutWrapper(cell=encoder_fw,
output_keep_prob=self.dropOutRate)
encoder_bw = tf.nn.rnn_cell.BasicLSTMCell(self.lstmUnits)
encoder_bw = tf.nn.rnn_cell.DropoutWrapper(cell=encoder_bw,
output_keep_prob=self.dropOutRate)
(
(encoder_fw_output, encoder_bw_output),
(encoder_fw_state, encoder_bw_state)
) = (
tf.nn.bidirectional_dynamic_rnn(cell_fw=encoder_fw,
cell_bw=encoder_bw,
inputs=input_data,
sequence_length=input_length,
dtype=tf.float32)
)
output = tf.concat((encoder_fw_output, encoder_bw_output), 2)
with tf.variable_scope('Attention'):
att_in = tf.expand_dims(output, axis=2)
# 初始化选择方法：https://blog.csdn.net/fanzonghao/article/details/82851327?utm_source=blogxgwz1
# Xavier initialization
# W = tf.Variable(np.random.randn(node_in, node_out).astype('float32'))/np.sqrt(node_in)
# 对于relu激活，He initialization
# W = tf.Variable(np.random.randn(node_in,node_out)) / np.sqrt(node_in/2)
w_att = tf.Variable(
tf.random_normal([1, 1, 2 * self.lstmUnits, self.attentionSize],
stddev=0.1
), name='w_att'
)
b_att = tf.Variable(tf.random_normal([self.attentionSize], stddev=0.1), name='b_att')
u_att = tf.Variable(
tf.random_normal([1, 1, self.attentionSize, 1],
stddev=0.1
), name='u_att'
)
v_att = tf.tanh(
tf.nn.conv2d(att_in, w_att, strides=[1, 1, 1, 1],
padding='SAME')
+ b_att
)
betas = tf.nn.conv2d(v_att, u_att, strides=[1, 1, 1, 1], padding='SAME')
exp_betas = tf.reshape(tf.exp(betas), [-1, self.maxSeqLength])
alphas = exp_betas / tf.reshape(tf.reduce_sum(exp_betas, 1), [-1, 1])
last = tf.reduce_sum(output * tf.reshape(alphas, [-1, self.maxSeqLength, 1]), 1)
with tf.variable_scope('Prediction'):
recall_matrix = tf.Variable(tf.zeros([4, 4], dtype=tf.float64, name='recall_matrix'))
acc_matrix = tf.Variable(tf.zeros([4, 4], dtype=tf.float64, name='acc_matrix'))
weight = tf.Variable(
tf.random_normal([2 * self.lstmUnits, self.numClasses],
stddev=0.1),
name='weight'
)
bias = tf.Variable(tf.random_normal([self.numClasses], stddev=0.1), name='bias')
y_ = (tf.matmul(last, weight) + bias) # 32,4
prediction = tf.argmax(y_, 1)
correctPred = tf.equal(prediction, tf.argmax(labels, 1))
accuracy = tf.reduce_mean(tf.cast(correctPred, tf.float32))
# 计算混淆矩阵
confusion_matrix = tf.confusion_matrix(
labels=tf.argmax(labels, 1), # 需要加一个 1
predictions=prediction,
num_classes=4,
dtype=tf.int32
)
# 横着除是召回率，纵着除是准确率
col = tf.reduce_sum(confusion_matrix, 1)
row = tf.reduce_sum(confusion_matrix, 0)
recall_matrix = tf.add(recall_matrix, confusion_matrix / col)
acc_matrix = tf.add(recall_matrix, confusion_matrix / row)
loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(logits=y_, labels=labels)
)
tf.summary.scalar('loss', loss)
tf.summary.scalar('recall of \'-2\'', recall_matrix[0][0])
tf.summary.scalar('recall of \'-1\'', recall_matrix[1][1])
tf.summary.scalar('recall of \'-0\'', recall_matrix[2][2])
tf.summary.scalar('recall of \'1\'', recall_matrix[3][3])
tf.summary.scalar('accuarcy of \'-2\'', acc_matrix[0][0])
tf.summary.scalar('accuarcy of \'-1\'', acc_matrix[1][1])
tf.summary.scalar('accuarcy of \'0\'', acc_matrix[2][2])
tf.summary.scalar('accuarcy of \'1\'', acc_matrix[3][3])
merged = tf.summary.merge_all()
with tf.variable_scope('Optimise'):
optimizer = tf.train.AdamOptimizer(learning_rate=self.learn_rate).minimize(loss)
return graph
if __name__ == '__main__':
kw = {}
kw['iter'] = 6600
kw['label'] = 1
kw['id'] = 111
model = Load_model(**kw)
graph = model.graph