Better batch building

[rohansaw]

We are currently using random batching mainly. We also have an implementation for a class-sampler in which for each class a certain amount of images are sampled. However, the results seem to overfit fast when using this sampler. I don't know why, exactly, this should be revisited, since in my opinion the class sampler makes more sense to exploit the data.