机器学习-十大人工智能算法

import numpy as np import random # 激活函数 def activation(activationType, x): if activationType == "sigmoid": return 1 / (1 + np.exp(-x)) elif activationType == "tanh": return np.tanh(x) else: print("暂无其他激活函数，请重新选择") # 求导 def derivativeOfActivation(activationType, y): # y=activation(activation,x) if activationType == "sigmoid": return y * (1 - y) elif activationType == "tanh": return 1 - np.power(y, 2) else: print("暂无其他激活函数导数，请重新选择") def activation_derivative(activationType, x): if activationType == "sigmoid": return sigmoid(x) * (1 - sigmoid(x)) # sigmoid 函数的导数 elif activationType == "tanh": return 1 - np.tanh(x) ** 2 # tanh 函数的导数 else: raise ValueError("Unknown activation type: %s" % (activationType)) def sigmoid(x): return 1 / (1 + np.exp(-x)) # 构建3层BP的弱分类器 # 输入->隐藏层->输出层 class BPNN(object): def __init__(self, iN, wHN, oN, activationType="sigmoid", eta=0.01, n_iter=10): ''' eta:学习率 n_iter:迭代次数 iN:输入层结点数目 wHN:隐藏层结点的数目 oN:输出层结点的数目 ''' self.eta = eta # 0.01 self.activationType = activationType self.n_iter = n_iter self.iN_ = iN; self.wHN_ = wHN; self.oN_ = oN; random.seed(10) self.weightIToH_ = 0.1 * np.random.randn(self.wHN_, self.iN_) # 输入层到隐藏层的权重矩阵hidden*input h*i self.weightHToO_ = 0.1 * np.random.randn(self.oN_, self.wHN_) # 隐藏层到输出层的权重矩阵output*hidden o*h self.b1 = np.zeros((1, self.wHN_)) # 存放输入层到隐藏层的偏移b1,1*hidden 1*h self.b2 = np.zeros((1, self.oN_)) # 存放隐藏层到输出层的偏移b2,1*output 1*o def loss(self, x, y, lamba=0): ''' x:m,n y:m,1 ''' n_train, n_features = x.shape # 前向传播 t = (np.dot(x, self.weightIToH_.T) + self.b1) # 输入层到隐藏层的结果 a1 = activation(self.activationType, t) # 隐藏层的输出结果 a2 = activation(self.activationType, np.dot(a1, self.weightHToO_.T) + self.b2) # 输出层的输出结果 loss = 0.5 * np.sum((a2 - y) ** 2) / n_train loss += 0.5 * lamba * ( np.sum(self.weightIToH_ * self.weightIToH_) + np.sum(self.weightHToO_ * self.weightHToO_)) / n_train # backward后向传播过程 delta2 = (a2 - y) * activation_derivative(self.activationType, np.dot(a1, self.weightHToO_.T) + self.b2) # 输出层的误差 dWeightHToO = np.dot(delta2.T, a1) # 隐藏层到输出层的权重矩阵的梯度 db2 = np.sum(delta2, axis=0) # 隐藏层到输出层的偏移的梯度 delta1 = np.dot(delta2, self.weightHToO_) * activation_derivative(self.activationType, t) # 隐藏层的误差 dWeightIToH = np.dot(delta1.T, x) # 输入层到隐藏层的权重矩阵的梯度 db1 = np.sum(delta1, axis=0) # 输入层到隐藏层的偏移的梯度 dWeightIToH /= n_train dWeightHToO /= n_train db1 /= n_train db2 /= n_train return loss, dWeightIToH, dWeightHToO, db1, db2 def train(self, x, y): loss_list = [] # 存放损失函数 errors_samples = [] # 存放分类错误的样本数据 errors_labels = [] # 存放分类错误的样本数据的正确值的标签 true_samples = [] # 存放分类正确的样本数据 true_labels = [] # 存放分类正确的样本数据的正确值的标签 for i in range(self.n_iter): loss, dw1, dw2, db1, db2 = self.loss(x, y) loss_list.append(loss) self.weightIToH_ += -self.eta * dw1 self.weightHToO_ += -self.eta * dw2 self.b1 += -self.eta * db1 self.b2 += -self.eta * db2 if i % 10 == 0: print("i=%d,loss=%f" % (i, loss)) for i in range(x.shape[0]): if y[i, :] != self.generatePredictYMul(x)[i, :]: errors_samples.append(x[i, :]) errors_labels.append(y[i, :]) else: true_samples.append(x[i, :]) true_labels.append(y[i, :]) return loss_list, errors_samples, true_samples, errors_labels, true_labels def generatePredictYMul(self, x): mulIToH = activation(self.activationType, (np.dot(x, self.weightIToH_.T) + self.b1)) # 隐藏层输出结果m*hidden mulHToO = activation(self.activationType, (np.dot(mulIToH, self.weightHToO_.T) + self.b2)) # 输出层输出结果m*output return np.sign(mulHToO) from sklearn.datasets import load_iris, load_breast_cancer from sklearn.model_selection import train_test_split data = load_breast_cancer() x = data.data y = data.target X_train, X_test, y_train, y_test = train_test_split(x, y, test_size=0.2, random_state=888) print(X_train.shape, X_test.shape) bp = BPNN(30, 100, 1, eta=0.01, n_iter=100) loss_list, errors_samples, true_samples, errors_labels, true_labels = bp.train(X_train, y_train.reshape(y_train.shape[0], 1)) import matplotlib.pyplot as plt plt.plot(loss_list, label='train_loss') plt.title('loss') plt.xlabel('iters') plt.ylabel('loss') plt.legend() plt.show()