tensorflow-wavenet icon indicating copy to clipboard operation
tensorflow-wavenet copied to clipboard

为什么我的loss降到1.2345之后上涨了现在又3.6578

Open Yaque opened this issue 6 years ago • 2 comments

why?

Yaque avatar Aug 09 '18 02:08 Yaque

现在它又掉下去了变成0.8989

Yaque avatar Aug 09 '18 02:08 Yaque

please use English since github is quite a community that each person no matter the region or language should have access to and communicate in : - )

As for your question, how do you measure your loss. Is it the loss of one batch or a set of sequences? If it is from a batch, it is normal to go from 1.2345 to 3.6578, since you are learning the distribution which allow variations. And the learned distribution is not fully converged, it is possible for some specific cases the distribtion does not fit.

KingStorm avatar Sep 05 '18 03:09 KingStorm