How goes keras calculate gradients for custom loss functions?

[Nick Frosst]

So i see that keras can use custom loss functions by simply pacing a function with model.compile(loss=_your_loss_, optimizer=sgd) but how does it optimize this? It doesn't take a derivative version of the function so how does it get the derivative? 
thanks :)

cheers
_nick

[Klemen Grm]

If you look at the source file for builtin objective functions ( https://github.com/fchollet/keras/blob/master/keras/objectives.py ), notice they're all implemented as Theano functions, which enables automatic gradient calculation. This must also be the case for any custom objective function you implement yourself.