Mixloss 出现 nan

[jasdasdf]

大佬，我用的是Mixloss，一运行loss就 nan.
1、LR 我已经设置很小了（0.00001）；
2、没有/0 情况；请问还有可能是什么原因呢？

[huyanxin]

我印象中之前也遇到过Nan，直观感觉是这个loss确实有点问题，和你的数据集和网络初始化有点关系，曾经试过只喂随机数，前几个step正常，然后就突然崩掉了确实也很奇怪。你换成sisnr应该就可以了正常训练，效果也是相差不大的。

[Chen1399]

可以先用sisnr预训练几步,初始化参数,然后在换成mixloss进行训练

[iver56]

I had this problem too. I think the non-differentiability of sqrt at 0 is the culprit here. Adding a very small value before taking sqrt, or somehow avoid passing zeroes to sqrt, can help against NaNs.

[Andong-Li-speech]

大佬，我用的是Mixloss，一运行loss就 nan.
1、LR 我已经设置很小了（0.00001）；
2、没有/0 情况；请问还有可能是什么原因呢？

This may be caused by the compressed coefficient \alpha, i.e., 0.3. You may as well calculate the gradient of the network output and you will get it.

[Wangzhen-kris]

可以先用sisnr预训练几步,初始化参数,然后在换成mixloss进行训练

我尝试先用sisnr预训练了几步，并使用第一次保存下的模型初始化，但是mix-loss依然为nan，请问您具体是怎么做的呢

[Wangzhen-kris]

I had this problem too. I think the non-differentiability of sqrt at 0 is the culprit here. Adding a very small value before taking sqrt, or somehow avoid passing zeroes to sqrt, can help against NaNs.

That works for me. I set the value to 1e-6 and I'm not sure what is appropriate to set the value.

[iver56]

I had this problem too. I think the non-differentiability of sqrt at 0 is the culprit here. Adding a very small value before taking sqrt, or somehow avoid passing zeroes to sqrt, can help against NaNs.

That works for me. I set the value to 1e-6 and I'm not sure what is appropriate to set the value.

I've been using 1e-12. Note that the square root of 1e-6 ends up being a much larger number, namely 0.001