Python记录 tensor求梯度时为None的错误

今天学习对抗样本，需要对tensor求梯度，

x1, x2, label = data
x1.requires_grad = True
x2.requires_grad = True
if use_gpu:
    x1 = x1.cuda()
    x2 = x2.cuda()
    label = label.cuda()
output1, output2 = net(x1, x2)
# 计算损失
loss = criterion(output1, output2, label)
# 将所有现有的渐变归零
net.zero_grad()
loss.backward()
# 收集grad
x1_grad = x1.grad
x2_grad = x2.grad

求得梯度为None
发现是requires_grad位置不对，应该是tensor放入cuda后，requires_grad 默认改成False

if use_gpu:
    x1 = x1.cuda()
    x2 = x2.cuda()
    label = label.cuda()
x1.requires_grad = True
x2.requires_grad = True 

解决