利用python实现BP算法

 

import matplotlib.pyplot as plt
import seaborn as sns #要注意的是一旦导入了seaborn，matplotlib的默认作图风格就会被覆盖成seaborn的格式
from sklearn.datasets import make_moons
from sklearn.model_selection import train_test_split
N_SAMPLES = 2000 # 采样点数
TEST_SIZE = 0.3 # 测试数量比率
# 利用工具函数直接生成数据集
X, y = make_moons(n_samples = N_SAMPLES, noise=0.2, random_state=100)
# 将2000 个点按着7:3 分割为训练集和测试集
X_train, X_test, y_train, y_test = train_test_split(X, y,test_size=TEST_SIZE, random_state=42)
print(X.shape, y.shape)
# 绘制数据集的分布，X 为2D 坐标，y 为数据点的标签
def make_plot(X, y, plot_name, file_name=None, XX=None, YY=None, preds=None,dark=False):
    if (dark):
        plt.style.use('dark_background')
    else:
        sns.set_style("whitegrid")
    plt.figure(figsize=(16,12))
    axes = plt.gca()
    axes.set(xlabel="$x_1$", ylabel="$x_2$")
    plt.title(plot_name, fontsize=30)
    plt.subplots_adjust(left=0.20)
    plt.subplots_adjust(right=0.80)
    if(XX is not None and YY is not None and preds is not None):
        plt.contourf(XX, YY, preds.reshape(XX.shape), 25, alpha = 1,cmap=plt.cm.Spectral)
        plt.contour(XX, YY, preds.reshape(XX.shape), levels=[.5],cmap="Greys", vmin=0, vmax=.6)
    # 绘制散点图，根据标签区分颜色
    plt.scatter(X[:, 0], X[:, 1], c=y.ravel(), s=40, cmap=plt.cm.Spectral,edgecolors='none')
    plt.savefig('dataset.svg')
    plt.show()
    plt.close()
# 调用make_plot 函数绘制数据的分布，其中X 为2D 坐标，y 为标签
make_plot(X, y, "Classification Dataset Visualization ")

(2000, 2) (2000,)

class Layer:
    # 全连接网络层
    def __init__(self, n_input, n_neurons, activation=None, weights=None,
                 bias=None):
       
 
        # 通过正态分布初始化网络权值，初始化非常重要，不合适的初始化将导致网络不收敛
        self.weights = weights if weights is not None else np.random.randn(n_input, n_neurons) * np.sqrt(1 / n_neurons)
        self.bias = bias if bias is not None else np.random.rand(n_neurons) *0.1
        self.activation = activation  # 激活函数类型，如’sigmoid’
        self.last_activation = None  # 激活函数的输出值o
        self.error = None  # 用于计算当前层的delta 变量的中间变量
        self.delta = None  # 记录当前层的delta 变量，用于计算梯度
 
    def activate(self, x):
        # 前向传播
        r = np.dot(x, self.weights) + self.bias  # X@W+b
        # 通过激活函数，得到全连接层的输出o
        self.last_activation = self._apply_activation(r)
        return self.last_activation
    # 其中self._apply_activation 实现了不同的激活函数的前向计算过程：
    def _apply_activation(self, r):
 
        # 计算激活函数的输出
        if self.activation is None:
            return r  # 无激活函数，直接返回
        # ReLU 激活函数
        elif self.activation == 'relu':
            return np.maximum(r, 0)
        # tanh
        elif self.activation == 'tanh':
            return np.tanh(r)
        # sigmoid
        elif self.activation == 'sigmoid':
            return 1 / (1 + np.exp(-r))
        return r
 
    # 针对于不同的激活函数，它们的导数计算实现如下：
    def apply_activation_derivative(self, r):
 
        # 计算激活函数的导数
        # 无激活函数，导数为1
        if self.activation is None:
            return np.ones_like(r)
        # ReLU 函数的导数实现
        elif self.activation == 'relu':
            grad = np.array(r, copy=True)
            grad[r > 0] = 1.
            grad[r <= 0] = 0.
            return grad
        # tanh 函数的导数实现
        elif self.activation == 'tanh':
            return 1 - r ** 2
        # Sigmoid 函数的导数实现
        elif self.activation == 'sigmoid':
            return r * (1 - r)
        return r

 

class NeuralNetwork:
    # 神经网络大类
    def __init__(self):
        self._layers = [] # 网络层对象列表
    def add_layer(self, layer):
        # 追加网络层
        self._layers.append(layer)
    # 网络的前向传播只需要循环调用个网络层对象的前向计算函数即可
    def feed_forward(self, X):
        # 前向传播
        for layer in self._layers:
            # 依次通过各个网络层
            X = layer.activate(X)
        return X
 
    #网络模型的反向传播实现稍复杂，需要从最末层开始，计算每层的𝛿变量，根据我们
    #推导的梯度公式，将计算出的𝛿变量存储在Layer类的delta变量中
    # 因此，在backpropagation 函数中，反向计算每层的𝛿变量，并根据梯度公式计算每层参数的梯度值，
    # 按着梯度下降算法完成一次参数的更新。
    def backpropagation(self, X, y, learning_rate):
 
        # 反向传播算法实现
        # 前向计算，得到输出值
        output = self.feed_forward(X)
        for i in reversed(range(len(self._layers))):  # 反向循环
            layer = self._layers[i]  # 得到当前层对象
            # 如果是输出层
            if layer == self._layers[-1]:  # 对于输出层
                layer.error = y - output  # 计算2 分类任务的均方差的导数
            # 关键步骤：计算最后一层的delta，参考输出层的梯度公式
                layer.delta = layer.error * layer.apply_activation_derivative(output)
 
            else:  # 如果是隐藏层
                next_layer = self._layers[i + 1]  # 得到下一层对象
                layer.error = np.dot(next_layer.weights, next_layer.delta)
                # 关键步骤：计算隐藏层的delta，参考隐藏层的梯度公式
                layer.delta = layer.error * layer.apply_activation_derivative(layer.last_activation)
 
 
    # 在反向计算完每层的𝛿变量后，只需要按着式计算每层的梯度，并更新网络参数即可。
    # 由于代码中的delta 计算的是−𝛿，因此更新时使用了加号。
                # 循环更新权值
        for i in range(len(self._layers)):
            layer = self._layers[i]
        # o_i 为上一网络层的输出
            o_i = np.atleast_2d(X if i == 0 else self._layers[i-1].last_activation)
            # 梯度下降算法，delta 是公式中的负数，故这里用加号
            layer.weights += layer.delta * o_i.T * learning_rate

 

def train(self, X_train, X_test, y_train, y_test, learning_rate,max_epochs):
        # 网络训练函数
        # one-hot 编码
        y_onehot = np.zeros((y_train.shape[0], 2))
        y_onehot[np.arange(y_train.shape[0]), y_train] = 1
        mses = []
        accs = []
        for i in range(max_epochs):  # 训练1000 个epoch
            for j in range(len(X_train)):  # 一次训练一个样本
                self.backpropagation(X_train[j], y_onehot[j], learning_rate)
            if i % 10 == 0:
                # 打印出MSE Loss
                mse = np.mean(np.square(y_onehot - self.feed_forward(X_train)))
                mses.append(mse)
                print('Epoch: #%s, MSE: %f' % (i, float(mse)))
                # 统计并打印准确率
                acc = self.accuracy(self.predict(X_test),y_test.flatten())
                accs.append(acc)
                print('Accuracy: %.2f%%' % (acc * 100))

        return mses,accs
 
    def accuracy(self,y_pre,y_true):
        return np.mean((np.argmax(y_pre, axis=1) == y_true))
 
    def predict(self,X_test):
        return self.feed_forward(X_test)

import numpy as np
nn = NeuralNetwork() # 实例化网络类
nn.add_layer(Layer(2, 25, 'sigmoid')) # 隐藏层1, 2=>25
nn.add_layer(Layer(25, 50, 'sigmoid')) # 隐藏层2, 25=>50
nn.add_layer(Layer(50, 25, 'sigmoid')) # 隐藏层3, 50=>25
nn.add_layer(Layer(25, 2, 'sigmoid')) # 输出层, 25=>2
learning_rate = 0.01
max_epochs = 1000
mses,accs = nn.train(X_train, X_test, y_train, y_test, learning_rate,max_epochs)

Epoch: #0, MSE: 0.248026
Accuracy: 49.50%
Epoch: #10, MSE: 0.147105
Accuracy: 81.17%
Epoch: #20, MSE: 0.100478
Accuracy: 87.50%
Epoch: #30, MSE: 0.096396
Accuracy: 88.67%
Epoch: #40, MSE: 0.096130
Accuracy: 89.00%
Epoch: #50, MSE: 0.096095
Accuracy: 89.00%
Epoch: #60, MSE: 0.096077
Accuracy: 89.00%
Epoch: #70, MSE: 0.096060
Accuracy: 88.83%
...
Accuracy: 97.67%
Epoch: #980, MSE: 0.024284
Accuracy: 97.67%
Epoch: #990, MSE: 0.024256
Accuracy: 97.67%

 

import matplotlib
import matplotlib.pyplot as plt


# Data for plotting
t = np.arange(0, 100, 1)


fig, ax = plt.subplots()
ax.plot(t, mses)

ax.set(xlabel='Epoch', ylabel='MSE',
       title='MSE Loss')
ax.grid()

#fig.savefig("test.png")
plt.show()

import matplotlib
import matplotlib.pyplot as plt


# Data for plotting
t = np.arange(0, 100, 1)


fig, ax = plt.subplots()
ax.plot(t, accs)

ax.set(xlabel='Epoch', ylabel='Accuracy',
       title='Accuracy')
ax.grid()

#fig.savefig("test.png")
plt.show()