Issue & Suggestion: Dropout specification

(Using the internal 0.11.x release provided by @robertl)
Suggestion

I checked with some current keras examples and it seems that the dropout it specified as a % to drop - not as currently labelled % to keep (I trust the implementation does in fact reflect the inverted spec!)

Would it make sense to align with keras and change to a % to drop

Issue

The box is labelled probability to keep - when dropout is enabled the default value is 0 - it should perhaps be something that is a more sensible default such as 0.9

I note that the box border turns red if an invalid value is entered - which is nice except that the mouseover help text says for (invalid) 1.1 “the two nearest valid values are 1 and 2” (and for 2.1 entered “2 and 3”). Whilst values [0,1] are accepted, the validation for the message seems to be thinking in terms of integers/

That said, I it would be fun to play with super-definite outcomes (i.e. p>1) just to see what happens :wink:

We went back and fort on having it to keep or to drop a while ago, seems we might have picked the wrong side of the coin :sweat_smile:
It would for sure make sense to keep it in line with Keras, but we might not fix that right away as we are hammering out some other issues and features.

The default setting is something we can easily change to 0.9 though. I think we even have a story for that somewhere on the board. Will see if we can get that prioritized a bit more.

That said, I it would be fun to play with super-definite outcomes (i.e. p>1) just to see what happens

Haha agreed, I don’t think TensorFlow lets us do that though.