看完一个有关神经网络的youtube视频后,我试图从头开始用numpy创建一个视频,但这给我一个错误,那就是尺寸错误。我对矩阵或numpy不太了解,这就是为什么我无法得到答案的原因。如果有人可以帮助我改善我的网络,我将非常高兴。
import numpy as np
class NeuralNetwork:
def __init__(self, input_size, hidden_size, output_size):
self.input_size = input_size
self.hidden_size = hidden_size
self.output_size = output_size
self.w1 = np.random.randn(self.input_size, self.hidden_size)
self.w2 = np.random.randn(self.hidden_size, self.output_size)
def sigmoid(self, n):
return 1 / (1 + np.exp(-n))
def sigmoid_prime(self, n):
return self.sigmoid(n) * (1 - self.sigmoid(n))
def propagation(self, x, y):
# forward
hidden1 = np.dot(x, self.w1)
hidden = self.sigmoid(hidden1)
output1 = np.dot(hidden, self.w2)
output = self.sigmoid(output1)
# backward
output_error = y - output
output_delta = output_error * self.sigmoid_prime(output)
hidden_error = output_delta.dot(self.w2.T)
hidden_delta = hidden_error*self.sigmoid_prime(hidden)
self.w1 += x.T.dot(hidden_delta)
self.w2 += hidden.T.dot(output_delta)
def predict(self, x):
hidden1 = np.dot(x, self.w1)
hidden = self.sigmoid(hidden1)
output1 = np.dot(hidden, self.w2)
output = self.sigmoid(output1)
return(output)
MyNet = NeuralNetwork(2, 5, 1)
for _ in range(500):
MyNet.propagation(
np.array([0, 1]),
np.array([1])
)
print(MyNet.predict( np.array([0, 1])))
答案 0 :(得分:0)
为帮助您提供帮助,请在错误中指出错误发生的位置以及错误中的哪个。
正如我所看到的,您正在尝试执行以下操作(在数学公式中是正确的):
您遇到的问题是,当向量只有一个维时,您试图执行矩阵乘法。为了解决这个问题,您应该简单地使用reshape:
self.w1 + = x.T.reshape(-1, 1).dot(hidden_delta.reshape(1, -1))
self.w2 + = hidden.T.reshape(-1, 1).dot(output_delta.reshape(1, -1))
我建议在使用S形时不要使用MSE错误(|| y - y_pred||^2
),而应使用cross entropy。当您处理概率时,这是更好的方法。
交叉熵定义如下:
# cross_entropy = - y log(x) + (1 - y) log(1 - x)
# dcross_entropy/dx = - y / x + (1 - y) / (1 - x)
output_error = (-y / (output + 1e-10)) + ((1 - y) / ( (1 - output) + 1e-10))
编辑:
import numpy as np
class NeuralNetwork:
def __init__(self, input_size, hidden_size, output_size, lr=0.1):
self.lr = lr
self.input_size = input_size
self.hidden_size = hidden_size
self.output_size = output_size
self.w1 = np.random.randn(self.input_size, self.hidden_size)
self.w2 = np.random.randn(self.hidden_size, self.output_size)
def sigmoid(self, n):
return 1 / (1 + np.exp(-n))
"""def sigmoid_prime(self, n):
return self.sigmoid(n) * (1 - self.sigmoid(n))"""
def propagation(self, x, y):
# forward
hidden1 = np.dot(x, self.w1)
hidden = self.sigmoid(hidden1)
output1 = np.dot(hidden, self.w2)
output = self.sigmoid(output1)
loss = -np.sum(y*np.log(output) + (1 - y)*np.log(1 - output), axis=-1)
print('Loss:', np.mean(loss))
# backward
#output_error = (-y / (output + 1e-10)) + ((1 - y) / ( (1 - output) + 1e-10))
#output_delta = output_error * output * (1 - output)
# simplified
output_delta = - y*(1 - output) + (1 - y)*output
self.w2 += - self.lr*hidden.T.dot(output_delta) / x.shape[0]
hidden_error = output_delta.dot(self.w2.T)
hidden_delta = hidden_error* hidden * (1 - hidden)
self.w1 += - self.lr*x.T.dot(hidden_delta) / x.shape[0]
def predict(self, x):
hidden1 = np.dot(x, self.w1)
hidden = self.sigmoid(hidden1)
output1 = np.dot(hidden, self.w2)
output = self.sigmoid(output1)
return(output)
MyNet = NeuralNetwork(2, 10, 1)
for _ in range(50000):
MyNet.propagation(
np.array([[0, 1], [1, 0], [1, 1], [0, 0]]),
np.array([[1], [1], [0], [0]])
)
print(MyNet.predict( np.array([0, 0]).reshape(1, -1)))
print(MyNet.predict( np.array([0, 1]).reshape(1, -1)))
print(MyNet.predict( np.array([1, 0]).reshape(1, -1)))
print(MyNet.predict( np.array([1, 1]).reshape(1, -1)))