神经网络中反向传播算法(backpropagation)的pytorch实现，pytorch教程中的代码解读以及其他一些疑问与解答

pytorch的官网上有一段教程，是使用python的numpy工具实现一个简单的神经网络的bp算法。下面先贴上自己的代码：

import numpy as np

N,D_in,H,D_out = 4,10,8,5

x = np.random.randn(N,D_in)#4x10
y = np.random.randn(N,D_out)#4x5
#print(x)
#print(y)

w1 = np.random.randn(D_in,H)#10x8
w2 = np.random.randn(H,D_out)#8x5

lr = 1e-6

h = x.dot(w1)#h=x*w1,4x8
h_relu = np.maximum(h,0)#h_relu=relu(h),4x8
y_pred = h_relu.dot(w2)#y_pred=h_relu*w2,4x5
print(h)

loss = np.square(y_pred - y).sum()#loss=sum((y_pred-y)^2)
print("loss: %f" %loss)

#grad_y_pred=d(loss)/d(y_pred),4x5
grad_y_pred = 2.0*(y_pred - y)
#grad_w2=d(loss)/d(w2)=grad_y_pred * d(y_pred)/d(w2),8x4 * 4x5=>8x5
grad_w2 = h_relu.T.dot(grad_y_pred)

#grad_h_relu=d(loss)/d(h_relu)=grad_y_pred * d(y_pred)/d(h_relu),4x5 * 5x8
grad_h_relu = grad_y_pred.dot(w2.T)
#grad_h=(relu())'
grad_h = grad_h_relu.copy()
grad