added classification switcher into DBN

[snurkabill]

it could be done for all network architectures. it's simple ... maybe dummy :)

[mihaelacr]

Hi! Thanks for the contribution!

Code looks good. I think that it might be more general though to do something similar to the 'activation functions' module, by adding a parameter to the DBN class that specifies the cost function to be used. This will allow to easily have different cost functions both for classification and regression.

What do you think? I would add a parameter cost_function to the DBN constructor and a cost_functions.py module which wraps over the theano functions, just like activation_functions does, using a similar API (cost(x,y)). Note that the wrappers are needed so that someone that does not know theano can use pydeeplearn.

[snurkabill]

I agree, problem might be that I don't know anything about theano :) trying to code something right now