Feng Kanghui

Did it happen during the training process? Not at the begining of training?

Because the input of this method (BCE) already passed Sigmoid layer, dont need BCE_with_logits .

Did you find the reason why this Error happens? I met the same error. So upset.....

遇到训练过程中loss为nan/inf的情况报错：CUDA error: device-side assert triggered 可以尝试调小batchsize或者学习率来暂时解决这个情况。