计算MSELoss的梯度

HDU_Du

已于 2022-10-22 11:13:18 修改

阅读量223

点赞数

文章标签： python

于 2021-11-16 16:03:02 首次发布

本文链接：https://blog.csdn.net/weixin_42656681/article/details/121358275

版权

import torch.nn as nn
x_in = torch.tensor([[0, 1, 2], [3, 4, 5], [6, 7, 8]]).view(1,1,3,3).type(torch.FloatTensor)
x_in = Variable(x_in, requires_grad = True)
print(x_in)

y_out = torch.tensor([[19, 25],[37, 43]]).view(1,1,2,2).type(torch.FloatTensor)
c_core = torch.ones(2,2).view((1,1,2,2)).type(torch.FloatTensor)
c_core = c_core.requires_grad_()

LR=0.01
# 定义一个损失函数
loss_fun = nn.MSELoss(reduce=True, size_average=True)
for i in range(10):
 y_pre = F.conv2d(x_in,c_core)
 loss = loss_fun(y_pre,y_out)
 c_core.retain_grad()
 loss.backward()
 print(c_core.data)
 print(c_core.grad)
 print(c_core.grad_fn)
 c_core = c_core-c_core.grad*LR
 print('the loss is:',loss)
print('c_core: ',c_core)