Feature/pre lu kaimin

[beckstev]

This branch changes:

• Implementation: PReLU activation function
• Implementation: he_normal weight initializer
• Implementation: ModelCheckpoint
• Implementation: Seperation of model and history to be able to save the history at every time
• Implementation: ManualStopping the model via Keyboard Interrupt
• Moved DataGenerator to cpu
• changes training default parameter
• Defined own PReLU class (inherit from original class) to add __name__ parameter, which is needed to save the traning model properly. The current implementation of PReLU does not include a __name__ attribute, although it is needed - strange.