Cleaning up: groundhog.cost_layers

[rizar]

This issue is a part of my initiative to leave only useful/used code in this main fork of Grounhog.

The following classes from groundhog.layers.cost_layers:

• LinearLayer
• SigmoidLayer

are not used in any groundhog-based project I know. May I just delete them?

[kyunghyuncho]

I think we should leave both of them. If anyone tries to train some of our models on non-text datasets, either linear or sigmoid output will be most widely used.

[rizar]

Let me try to convince you:

• Having them in this library means that we are sure they work correctly. How can we be sure, when nobody is using them?
• As you have already acknowledged back in Montreal, having projection and non-linearity in output layers was a design flow. Instead of these two there should have been MSELayer and CrossEntropyLayer. If people start using these current layers, the cost of fixing this design flow will increase.
• A major refactoring of groundhog is coming. We will have to take care of every component, ensure that it keeps working properly after refactoring. Do you want to waste time on components nobody is using at the moment? It will be a way easier to rewrite them later from scratch, without this ugly copy-paste, than to carry them carefully in the future.

[kyunghyuncho]

Go ahead with removing those. Just make sure to leave their names in commit comments so that we can search for them later, if we need it.