自适应线性神经网络Adaline

Posted 2022-11-01 thsss

自适应线性神经网络Adaptive linear network， 是神经网络的入门级别网络。

相对于感知器，

1. 采用了f（z）=z的激活函数，属于连续函数。
2. 代价函数为LMS函数，最小均方算法，Least mean square。

 

 实现上，采用随机梯度下降，由于更新的随机性，运行多次结果是不同的。

  1 ‘‘‘
  2 Adaline classifier
  3 
  4 created on 2019.9.14
  5 author: vince
  6 ‘‘‘
  7 import pandas 
  8 import math
  9 import numpy  
 10 import logging
 11 import random
 12 import matplotlib.pyplot as plt
 13 
 14 from sklearn.datasets import load_iris
 15 from sklearn.model_selection import train_test_split
 16 from sklearn.metrics import accuracy_score
 17 
 18 ‘‘‘
 19 Adaline classifier
 20 
 21 Attributes
 22 w: ld-array = weights after training
 23 l: list = number of misclassification during each iteration 
 24 ‘‘‘
 25 class Adaline:
 26     def __init__(self, eta = 0.001, iter_num = 500, batch_size = 1):
 27         ‘‘‘
 28         eta: float = learning rate (between 0.0 and 1.0).
 29         iter_num: int = iteration over the training dataset.
 30         batch_size: int = gradient descent batch number, 
 31             if batch_size == 1, used SGD; 
 32             if batch_size == 0, use BGD; 
 33             else MBGD;
 34         ‘‘‘
 35 
 36         self.eta = eta;
 37         self.iter_num = iter_num;
 38         self.batch_size = batch_size;
 39 
 40     def train(self, X, Y):
 41         ‘‘‘
 42         train training data.
 43         X:array-like, shape=[n_samples, n_features] = Training vectors, 
 44             where n_samples is the number of training samples and 
 45             n_features is the number of features.
 46         Y:array-like, share=[n_samples] = traget values.
 47         ‘‘‘
 48         self.w = numpy.zeros(1 + X.shape[1]);
 49         self.l = numpy.zeros(self.iter_num);
 50         for iter_index  in range(self.iter_num):
 51             for rand_time in range(X.shape[0]): 
 52                 sample_index = random.randint(0, X.shape[0] - 1);
 53                 if (self.activation(X[sample_index]) == Y[sample_index]):
 54                     continue;
 55                 output = self.net_input(X[sample_index]);
 56                 errors = Y[sample_index] - output;
 57                 self.w[0] += self.eta * errors;
 58                 self.w[1:] += self.eta * numpy.dot(errors, X[sample_index]);
 59                 break;
 60             for sample_index in range(X.shape[0]): 
 61                 self.l[iter_index] += (Y[sample_index] - self.net_input(X[sample_index])) ** 2 * 0.5;
 62             logging.info("iter %s: w0(%s), w1(%s), w2(%s), l(%s)" %
 63                     (iter_index, self.w[0], self.w[1], self.w[2], self.l[iter_index]));
 64             if iter_index > 1 and math.fabs(self.l[iter_index - 1] - self.l[iter_index]) < 0.0001: 
 65                 break;
 66 
 67     def activation(self, x):
 68         return numpy.where(self.net_input(x) >= 0.0 , 1 , -1);
 69 
 70     def net_input(self, x): 
 71         return numpy.dot(x, self.w[1:]) + self.w[0];
 72 
 73     def predict(self, x):
 74         return self.activation(x);
 75 
 76 def main():
 77     logging.basicConfig(level = logging.INFO,
 78             format = ‘%(asctime)s %(filename)s[line:%(lineno)d] %(levelname)s %(message)s‘,
 79             datefmt = ‘%a, %d %b %Y %H:%M:%S‘);
 80 
 81     iris = load_iris();
 82 
 83     features = iris.data[:99, [0, 2]];
 84     # normalization
 85     features_std = numpy.copy(features);
 86     for i in range(features.shape[1]):
 87         features_std[:, i] = (features_std[:, i] - features[:, i].mean()) / features[:, i].std();
 88 
 89     labels = numpy.where(iris.target[:99] == 0, -1, 1);
 90 
 91     # 2/3 data from training, 1/3 data for testing
 92     train_features, test_features, train_labels, test_labels = train_test_split(
 93             features_std, labels, test_size = 0.33, random_state = 23323);
 94     
 95     logging.info("train set shape:%s"  % (str(train_features.shape)));
 96 
 97     classifier = Adaline();
 98 
 99     classifier.train(train_features, train_labels);
100         
101     test_predict = numpy.array([]);
102     for feature in test_features:
103         predict_label = classifier.predict(feature);
104         test_predict = numpy.append(test_predict, predict_label);
105 
106     score = accuracy_score(test_labels, test_predict);
107     logging.info("The accruacy score is: %s "% (str(score)));
108 
109     #plot
110     x_min, x_max = train_features[:, 0].min() - 1, train_features[:, 0].max() + 1;
111     y_min, y_max = train_features[:, 1].min() - 1, train_features[:, 1].max() + 1;
112     plt.xlim(x_min, x_max);
113     plt.ylim(y_min, y_max);
114     plt.xlabel("width");
115     plt.ylabel("heigt");
116 
117     plt.scatter(train_features[:, 0], train_features[:, 1], c = train_labels, marker = ‘o‘, s = 10);
118 
119     k = - classifier.w[1] / classifier.w[2];
120     d = - classifier.w[0] / classifier.w[2];
121 
122     plt.plot([x_min, x_max], [k * x_min + d, k * x_max + d], "go-");
123 
124     plt.show();
125     
126 
127 if __name__ == "__main__":
128     main();