Learning to Generate Parameters of ConvNets for Unseen Image Data