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
| # encoding:utf-8
"""
@version:1
@author:Chenlini
@file:softmax.py
@time:2017/6/15 11:22
"""
# coding:utf8
from theano import *
from sklearn.datasets import *
from pandas import *
from sklearn.model_selection import *
from numpy import *
class SoftMax:
def __init__(self, Iter=1, learnrate=0.01, alpha=0.1):
self.Iter = Iter
self.learnrate = learnrate
self.alpha = alpha
def training(self, X, Y, typenum, batch_size=500):
##featuresnum是X的纬度，batchesnum是分区的数目
featuresnum = X.shape[1]
batchesnum = int(X.shape[0] / batch_size)
###因为在softMax中theta和x都是全局变量，所以这里应该声明为全局变量
X = theano.shared(np.asarray(X, dtype=float))
###保证Y是整型的
Y = theano.tensor.cast(theano.shared(np.asarray(Y)), 'int32')
#x是一个矩阵，y是一个向量
x = theano.tensor.matrix('x')
y = theano.tensor.ivector('y')
##表示index是整型标量
index = theano.tensor.lscalar()
theta = theano.shared(value=0.001 * zeros((featuresnum, typenum),
dtype=theano.config.floatX),
name='theta', borrow=True)
###hx就是计算softmax()的值
hx = theano.tensor.nnet.softmax(theano.tensor.dot(x, theta))
##这个就是loss
cost = -theano.tensor.mean(theano.tensor.log(hx)[theano.tensor.arange(y.shape[0]), y]) + 0.5 * self.alpha * theano.tensor.sum(theta ** 2)
#grad可以自动计算倒数
g_theta = theano.tensor.grad(cost, theta)
##这里就是用一个结构保存更新的值
updates = [(theta, theta - self.learnrate * g_theta)]
#批量更新
train_model = theano.function(
inputs=[index], outputs=cost, updates=updates, givens={
x: X[index * batch_size: (index + 1) * batch_size],
y: Y[index * batch_size: (index + 1) * batch_size]
}, allow_input_downcast=True
)
lastcostJ = np.inf
stop = False
epoch = 0
costj = []
###总体的循环执行，当达到最大循环次数或者cost不再下降的时候结束
while (epoch < self.Iter) and (not stop):
epoch = epoch + 1
for minibatch_index in range(batchesnum):
costj.append(train_model(minibatch_index))
if np.mean(costj) >= lastcostJ:
print("costJ is increasing !!!")
stop = True
else:
lastcostJ = np.mean(costj)
print(('epoch %i, minibatch %i/%i,averange cost is %f') %
(epoch, minibatch_index + 1, batchesnum, lastcostJ))
self.theta = theta
return self.theta.get_value()
def testing(self, X, Y, batch_size=500):
batchesnum = int(X.shape[0] / batch_size)
X = theano.shared(np.asarray(X, dtype=float))
Y = theano.tensor.cast(theano.shared(np.asarray(Y)), 'int32')
x = theano.tensor.matrix('x')
y = theano.tensor.ivector('y')
index = theano.tensor.lscalar()
hx = theano.tensor.nnet.softmax(theano.tensor.dot(x, self.theta))
###求最后的类别结果和误差
predict = theano.tensor.argmax(hx, axis=1)
errors = theano.tensor.mean(theano.tensor.neq(predict, y))
test_model = theano.function(
inputs=[index], outputs=errors, givens={
x: X[index * batch_size: (index + 1) * batch_size],
y: Y[index * batch_size: (index + 1) * batch_size]
}, allow_input_downcast=True
)
test_losses = []
for minibatch_index in range(batchesnum):
test_losses.append(test_model(minibatch_index))
test_score = np.mean(test_losses)
print(('minibatch %i/%i, test error of model %f %%') %
(minibatch_index + 1, batchesnum, test_score * 100.))
if __name__ == '__main__':
X, y = make_classification(n_samples=80000)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.5)
Testsoftmax = SoftMax()
###typenum=len(set(y)
Testsoftmax.training(X_train, y_train, typenum=len(set(y)))
Testsoftmax.testing(X_test, y_test)
|
