我想知道为什么我的神经网络不起作用。 我想说的是与此类似的问题,但是我还有一些我不明白的地方...
代码:
import numpy as np
inputs = np.array([
[[0],[0]],
[[1],[0]],
[[0],[1]],
[[1],[1]]
])
expected_output = np.array([
[0],
[1],
[1],
[0]
])
epochs = 100
lr = 0.2
hidden_weights = np.array([
[0.2, 0.3],
[0.4, 0.5]
])
hidden_bias = np.array([[0.3], [0.6]])
output_weights = np.array([[0.6, 0.7]])
output_bias = np.array([[0.5]])
def sigmoid(z):
return 1/(1+np.exp(-z))
def sigmoid_derivative(z):
return z * (1.0-z)
for _ in range(epochs):
for index, input in enumerate(inputs):
hidden_layer_activation = np.dot(hidden_weights, input)
hidden_layer_activation += hidden_bias
hidden_layer_output = sigmoid(hidden_layer_activation)
output_layer_activation = np.dot(output_weights, hidden_layer_output)
output_layer_activation += output_bias
predicted_output = sigmoid(output_layer_activation)
#Backpropagation
output_errors = expected_output[index] - predicted_output
hidden_errors = output_weights.T.dot(output_errors)
d_predicted_output = output_errors * sigmoid_derivative(predicted_output)
d_hidden_layer = hidden_errors * sigmoid_derivative(hidden_layer_output)
output_weights += np.dot(d_predicted_output, hidden_layer_output.T) * lr
hidden_weights += np.dot(d_hidden_layer, input.T) * lr
output_bias += np.sum(d_predicted_output) * lr
hidden_bias += np.sum(d_hidden_layer) * lr
# NOW THE TESTING,I pass 2 input neurons. One with value 1 and value 1
test = np.array([
[[1], [1]]
])
hidden_layer_activation = np.dot(hidden_weights, test[0])
hidden_layer_activation += hidden_bias
hidden_layer_output = sigmoid(hidden_layer_activation)
output_layer_activation = np.dot(output_weights, hidden_layer_output)
output_layer_activation += output_bias
predicted_output = sigmoid(output_layer_activation)
print(predicted_output)
Result : [[0.5]] for inputs 1 and 1
Wanted : [[0]] for inputs 1 and 1
我已经测试了前馈传播,并且工作正常。 错误似乎很好。
我认为更新权重是问题,但是更新权重具有正确的公式。这段代码来自“制作您自己的神经网络,我使用的几乎是同一本书:
self.who += self.lr * numpy.dot((output_errors * final_outputs * (1.0 final_outputs)), numpy.transpose(hidden_outputs))
目前,我当时只转发2个神经元的1个输入,并计算出错误。我希望它保持这种状态,而不是一遍又一遍地转发整个测试数据。
有什么办法可以做到吗? 预先谢谢你:)
答案 0 :(得分:1)
您有一个小的实现错误:
在反向传播中,您评估:
hidden_errors = output_weights.T.dot(output_errors)
,但是必须基于d_predicted_output评估您的隐藏错误,如下所示:
hidden_errors = output_weights.T.dot(d_predicted_output)
此外,您应该降低学习速度,并增加学习次数。 10000个纪元且lr = 0.1对我有用,但是您可以对此进行微调。