python实现神经网络

神经网络包含一层输入层，一层隐含层，一层输出层，且都是三维向量
通过梯度下降法实现对神经网络的训练

import numpy as np
X = np.array([[1, 1, 1, 1],
              [1, 1, 0, 1],
              [1, 0, 1, 1]], dtype=float)
Y = np.array([[1, 0, 0], [1, 0, 0], [0, 1, 0]], dtype=int)
l_num = 2   #神经网络层数-1
W = np.array(np.random.randint(0, 1, (l_num, len(X[0])-1, len(X[0]))), dtype=float)
a = 0.3 #学习率
A = np.array(np.random.randint(1, 2, (len(X[0]), l_num + 1)), dtype=float)   #存储所有神经网络结点的值
B = np.array(np.random.randint(0, 1, (l_num+1, l_num)), dtype=float)   #存储dJ/dz,用于反向传播
while True:
    D = np.array(np.random.randint(0, 1, (l_num, len(X[0])-1, len(X[0]))), dtype=float)  #存储dJ/dW,用于梯度下降
    W_T = np.copy(W)
    for i in range(len(X)):
        # 前向传播
        A[:, 0] = X[i]
        k = 1
        while k < 3:
            A[1:, k] = 1 / (1 + np.exp(-(W[k - 1] @ A[:, k - 1])))
            k = k + 1
        # 反向传播
        B[:, 1] = A[1:, 2] - Y[i]
        k = 0
        while k > -1:
            temp = A[1:, k + 1] * (np.array([1.0, 1.0, 1.0]) - A[1:, k + 1])
            B[:, k] = (W[k + 1][:, 1:].T @ B[:, k + 1]) * temp
            k = k - 1
        k = 0
        #计算偏导数
        while k < l_num:
            D[k] = D[k] + B[:, k].reshape(3, 1) @ A[:, k].reshape(1, 4)
            k = k + 1
    D = D / len(X)
    # 梯度下降
    W = W - a * D
    res = abs(W_T - W) < 0.00001
    if res.all():
        break

print('权值:'+str(W))
#测试
A[:, 0] = np.array([1, 0, 1, 0])
k = 1
while k < 3:
    A[1:, k] = 1 / (1 + np.exp(-(W[k - 1] @ A[:, k - 1])))
    k = k + 1
print('预测结果：'+str(A[1:,-1]))