Closed matt-gardner closed 7 years ago
Turns out the error was because tensorflow couldn't compute a gradient for variables when there was a switch
in the loss function. This is actually kind of troubling, because it might mean all of our models and layers that use switch
might be broken... Replacing the switch
with an addition fixed things, though, so this is now ready to review.
@haniesedghi, FYI.
This isn't working yet; I'm running into a funny error with actually using this loss function. Once I figure out that error, though, this should be good to use.