Question

我的3层神经网络（输入 - >隐藏 - >输出）无法正常工作，没有语法错误，但它不能正常工作。为什么这样，我该如何解决？

我的反向传播不正确或者是什么错误？我不明白，一切似乎都是正确的，但显然不是。

我试图调整我的超参数，但它并没有真正做任何事情，我也试图以不同的顺序乘以衍生物，但这也不起作用。

"""
This is a three layer neural network that will predict the xor value of 
an input, 
using gradient descent and backpropagation
"""

import numpy as np

class NeuralNetwork:


    # relu activation function for hidden layer
    # derivative of relu at a specific point
    def relu(self, val, deriv=False):
        if deriv:
            val[val <= 0] = 0
            val[val > 0] = 1
        return np.maximum(0, val)


    # sigmoid actiavation function for output layer
    # derivative of sigmoid at a specific point
    def sigmoid(self, val, deriv=False):
        if deriv:
            return val * (1-val)
        return 1/(1+np.exp(-val))

    # start taining
    def Train(self, X, Y, w_ih, b_h, w_ho, b_o, lr, epochs):
        # training loop
        for i in range(epochs):

            # forward propagte
            # weighted sum of the hidden layer
            Z_h = np.dot(X, w_ih) + b_h

            # prediction of the hidden layer
            P_h = self.relu(Z_h)

            # weighted sum of output layer
            Z_o = np.dot(P_h, w_ho) + b_o

            # prediction of the output layer
            P_o = self.sigmoid(Z_o)

            # backpropagation
            # error for prediction
            # MSE - mean squared error
            E_o = ((P_o - Y)**2)/2

            # partial derivative of error with respect to the parameters
            dE_dPo = (P_o - Y) # 4, 1
            dPo_dZo = self.sigmoid(P_o, True) # 4, 1
            dZo_dwho = P_h # 4, 2
            dZo_dPh = w_ho # 2, 1
            dPh_dZh = self.relu(P_h, True) # 4, 2
            dZh_dwih = X # 4, 2

            # update weights and biases from hidden to output
            delta_w_ho = lr * np.dot(dZo_dwho.T, dE_dPo * dPo_dZo)
            delta_b_o = lr * sum(dE_dPo * dPo_dZo)

            w_ho -= delta_w_ho
            b_o -= delta_b_o

            # update weights and biases from input to hidden
            delta_w_ih = lr * np.dot(np.dot(dZo_dPh, dE_dPo.T * dPo_dZo.T), dPh_dZh * dZh_dwih)
            delta_b_h = lr * sum(np.dot(np.dot(dZo_dPh, dE_dPo.T * dPo_dZo.T), dPh_dZh))

            w_ih -= delta_w_ih
            b_h -= delta_b_h

        print("prediction")
        print(P_o)
        print("MSE for output layer")
        print(E_o)

if __name__ == "__main__":

    nn = NeuralNetwork()

    input_data = np.array([[0, 0],
                           [0, 1],
                           [1, 0],
                           [1, 1]])

    output_data = np.array([[0],
                            [1],
                            [1],
                            [0]])

    # number of neurons in each layer
    n_i = 2
    n_h = 2
    n_o = 1

    # initialize weights and biases
    # seed the numbers
    np.random.seed(0)
    # weights and biases for hidden layer
    w_ih = np.random.random((n_i, n_h))
    b_h = np.random.random((n_h))
    # weights and biases for output layer
    w_ho = np.random.random((n_h, n_o))
    b_o = np.random.random(n_o)

    # hyper parameters
    learning_rate = 0.2
    epochs = 1000

    nn.Train(input_data, output_data, w_ih, b_h, w_ho, b_o, learning_rate, epochs)

为什么我的神经网络预测值接近0.5？

0 个答案: