You can find suggestions on http://deeplearning.net/software/theano/tutorial/nan_tutorial.html
On Tue, Aug 23, 2016, Qiuqiang Kong wrote: > Hi all, > > Sorry for cross posting! > > I was running recurrent neural network. I used scan function. In most > of time it works well. But very occasionally Nan occurs (for example after > 50 epochs). *But when I change from mode=FAST_RUN to mode=FAST_COMPILE, the > Nan disappeared!* > > The cost function is like H + H^2 + ... + H^n, > > If using mode=FAST_COMPILE, there is no error. > > If using mode=FAST_RUN, I guess theano uses optimization to speed up > which introduce numerical problem at the same time. > > Is there anyone know how to solve it? > > Thank you very much! > > Best wishes, > > Qiuqiang > > -- > > --- > You received this message because you are subscribed to the Google Groups > "theano-users" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > For more options, visit https://groups.google.com/d/optout. -- Pascal -- --- You received this message because you are subscribed to the Google Groups "theano-users" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. For more options, visit https://groups.google.com/d/optout.
