Closed chrisjbaik closed 8 years ago
well... so Cross-entropy, according to wikipedia, is a sum of a probability times a log SO.. log(x) = NaN when x < 0 right?
@arquinn you're absolutely right. I'm gonna take a cue from this page:
http://stackoverflow.com/questions/33712178/tensorflow-nan-bug
It's still not really changing much though; have to figure out what's going wrong.
This problem is fixed, but we're still not getting super awesome results. :(
@zainahamid @arquinn @paivaspol
Currently the training isn't working. Two things are suspect:
sum(row) % 10
, and then put it in a one-hot vector form.NaN
. how come?