使用感知机训练加法模型

 1 # -*-coding:utf-8-*-
 2 '@author: xijun.gong'
 3 import numpy as np
 4 import random
 5 import math
 6 
 7 
 8 class Perceptron:
 9     def __init__(self, learnRate, maxIter, bit_len):
10         """
11         :param bit_len
12         :param learnRate:
13         :param maxIter:  最大迭代次数
14         """
15         self.learmRate = learnRate;
16         self.weight = None;
17         self.maxIter = maxIter;
18         # produce map
19         self.bit_len = bit_len;
20         self.nummap = None;
21         self.initMap()
22         pass
23 
24     def initMap(self):
25         maxNum = (1 << self.bit_len);  # 该位数下的最大值
26         self.nummap = np.zeros((maxNum, self.bit_len), dtype=np.int);  # include zero
27         for _id in xrange(maxNum):
28             for index in xrange(self.bit_len):
29                 self.nummap[_id][index] = 1 & (_id >> index);
30         pass
31 
32     def initWeight(self):
33         """
34         :return:
35         """
36         self.weight = np.ones(self.bit_len) / self.bit_len;
37 
38     def fit(self, fds, labels):
39         """
40         :param fds: 训练样本集合
41         :param labels:
42         :return:
43         """
44         feature_nums = fds.shape[1]  # 样本中的特征参数数量
45         self.initWeight()
46         for iter in xrange(self.maxIter):
47             print 'train as iter is {} '.format(iter)
48             acc_cnt = 0
49             for _ind, sample in enumerate(fds):
50                 a = self.nummap[int(sample[0])];
51                 b = self.nummap[int(sample[1])];
52                 label_y = sum(self.weight * (a + b))
53                 # 计算var_w 表示倒三角w
54                 print 'the reality:{} , predict {}'.format(labels[_ind], label_y);
55                 if math.fabs(labels[_ind] - label_y) <= 0.000001:
56                     acc_cnt += 1;
57                     continue;
58                 var_w = self.learmRate * (labels[_ind] - label_y) * (a + b)
59                 self.weight += var_w;
60             print 'accuary is {}'.format(acc_cnt / (len(fds) * 1.0))
61             if acc_cnt == len(fds):
62                 np.save('weight.npy', {'weight': self.weight});
63                 return;
64         pass
65 
66     def load(self, path='weight.npy'):
67         return np.load(path)['weight']
68 
69     def predict(self, fd):
70         a = self.nummap[fd[0]];
71         b = self.nummap[fd[1]];
72         return sum(self.weight * (a + b))
73 
74     def predict_prod(self):
75         pass
76 
77 
78 if __name__ == '__main__':
79     import time
80 
81     perceptron = Perceptron(learnRate=0.01, maxIter=2000, bit_len=5);
82     xa = np.arange(31);
83     xb = np.zeros(31);
84     labels = np.zeros(31)
85     for i in xrange(31):
86         xb[i] = random.randint(0, (int(time.time() + 1)) % 31)
87         labels[i] = xb[i] + xa[i]
88     perceptron.fit(np.array([xa, xb]).T, labels)
89     print 'predict is {}'.format(perceptron.predict([24, 13]))

train as iter is 277 
the reality:0.0 , predict 0.0
the reality:16.0 , predict 16.0000005749
the reality:16.0 , predict 15.9999994995
the reality:3.0 , predict 3.00000059084
the reality:18.0 , predict 17.999999818
the reality:15.0 , predict 15.0000000195
the reality:20.0 , predict 19.9999998534
the reality:22.0 , predict 22.0000009642
the reality:10.0 , predict 9.99999911021
the reality:22.0 , predict 21.9999996143
the reality:23.0 , predict 22.9999990943
the reality:17.0 , predict 17.0000000549
the reality:25.0 , predict 24.9999994128
the reality:18.0 , predict 18.0000008934
the reality:20.0 , predict 19.9999998534
the reality:15.0 , predict 15.0000000195
the reality:27.0 , predict 26.999999038
the reality:31.0 , predict 30.9999993919
the reality:25.0 , predict 25.0000003525
the reality:21.0 , predict 20.9999999986
the reality:35.0 , predict 34.9999997457
the reality:29.0 , predict 28.9999993564
the reality:39.0 , predict 38.9999996894
the reality:26.0 , predict 26.0000009079
the reality:31.0 , predict 30.9999993919
the reality:25.0 , predict 24.9999990026
the reality:33.0 , predict 32.9999994273
the reality:32.0 , predict 31.9999999473
the reality:32.0 , predict 31.9999991549
the reality:34.0 , predict 34.0000002657
the reality:33.0 , predict 32.9999994273
accuary is 1.0
predict is 36.9999984312