Today I learned that having a bug in your framework's dropout implementation can make you waste a lot of compute time.
-
@fchollet By that, I mean having a layer that returns sequences as the output layer, and computing the cost as the mean MSE over timesteps.
@seaandsailor Doesn't sequence to sequence training with MSE do that already? You can set your targets to be (samples, time, features).
-
-
@fchollet Does it? I never tried, gonna check it out now. Thanks! -
@seaandsailor@fchollet I tried that, it does. But I wish there was a way to pass a mask parameter to the cost function, for dif leng sizes
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.