python实现bp神经网络

以下是一个简单的 Python 实现 BP 神经网络的示例代码：

import numpy as np

class NeuralNetwork:
    def __init__(self, inputs, hidden, outputs):
        self.inputs = inputs
        self.hidden = hidden
        self.outputs = outputs

        self.weights1 = np.random.rand(self.inputs, self.hidden)
        self.weights2 = np.random.rand(self.hidden, self.outputs)

    def sigmoid(self, x):
        return 1 / (1 + np.exp(-x))

    def sigmoid_derivative(self, x):
        return x * (1 - x)

    def feed_forward(self, X):
        self.hidden_sum = np.dot(X, self.weights1)
        self.activated_hidden = self.sigmoid(self.hidden_sum)
        self.output_sum = np.dot(self.activated_hidden, self.weights2)
        self.activated_output = self.sigmoid(self.output_sum)
        return self.activated_output

    def backward(self, X, y, o):
        self.o_error = y - o
        self.o_delta = self.o_error * self.sigmoid_derivative(o)

        self.z2_error = self.o_delta.dot(self.weights2.T)
        self.z2_delta = self.z2_error * self.sigmoid_derivative(self.activated_hidden)

        self.weights1 += X.T.dot(self.z2_delta)
        self.weights2 += self.activated_hidden.T.dot(self.o_delta)

    def train(self, X, y):
        o = self.feed_forward(X)
        self.backward(X, y, o)

    def save_weights(self, model):
        np.save(model, [self.weights1, self.weights2])

    def load_weights(self, model):
        self.weights1, self.weights2 = np.load(model)

X = np.array([[0, 0], [0, 1], [1, 0], [1, 1]])
y = np.array([[0], [1], [1], [0]])

NN = NeuralNetwork(2, 3, 1)

for i in range(10000):
    NN.train(X, y)
    if i % 500 == 0:
        print("Iteration ", i)
        print("Input: \n", X)
        print("Actual Output: \n", y)
        print("Predicted Output: \n", NN.feed_forward(X))
        print("Loss: \n", np.mean(np.square(y - NN.feed_forward(X))))
        print("\n")

NN.save_weights("model.npy")

在这个示例中，我们使用了一个包含 2 个输入、3 个隐藏单元和 1 个输出的神经网络。我们使用 sigmoid 激活函数和均方误差（MSE）作为损失函数。训练时，我们使用反向传播算法来更新权重，并使用 feed_forward 函数来进行前向传播预测。最后，我们使用 save_weights 和 load_weights 函数来保存和加载模型权重。