@karpathy @fchollet @sedielem @_mayfa @j26774 comparing char-rnn models, very different results for keras & lasagne!pic.twitter.com/dkynepPH9e
-
@sedielem fwiw, i noticed disagreement about rmsprop divisor. lasagne&keras use sqrt(a+1e-6). torch's sqrt(a)+1e-8 seems better?
we can definitely consider changing that (unless it creates numerical stability issues on GPU, but that should prolly be fine)
-
-
it seems battle-tested in torch, but i'd be curious to see a thorough justification. this is all i've seen http://sifter.org/~simon/journal/20150420.html …
-
interesting article, thank you. I will do some testing and update the Keras optimizers.
End of conversation
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.