example.ipynb中进行训练测试loss为nan

[SilentMoebuta]

前面代码不变

model(**batch).loss

改为

for i in range(10):
    output = model(**batch)
    loss = output.loss
    loss.backward()
    optimizer.step()
    lr_scheduler.step()
    optimizer.zero_grad()
    print(loss.detach().float())

输出

tensor(3.2207, device='cuda:0')
tensor(nan, device='cuda:0')
tensor(nan, device='cuda:0')
tensor(nan, device='cuda:0')
tensor(nan, device='cuda:0')
tensor(nan, device='cuda:0')
tensor(nan, device='cuda:0')
tensor(nan, device='cuda:0')
tensor(nan, device='cuda:0')
tensor(nan, device='cuda:0')

只有第1个step时loss正常计算，请问这是为啥？
lr设置1e-8

[iridescentee]

I have met the same problem. Did you find the solution?