Get rid of requirement for augmentation factor

[gngdb]

Had to have an augmentation factor to know how to build the empty array before iterating over images. Unfortunately, that means it has to be written into the run settings by hand. There's no point in doing this though, because we could have a function to estimate by running the processing function over a test image (just a np.zeros 2D array) and seeing how many images result. This would also be more robust, because it wouldn't rely on people calculating the augmentation factor correctly.

[scottclowe]

However, if we were using a subsample of a very large augmentation space (sampled probabilistically) we would definitely need to know the augmentation factor. Without it we wouldn't know when to stop generating training images.

[gngdb]

Yeah, I'm not sure how that works. The docs for Pylearn2 say they support drawing samples from some "infinite stream", but I don't know how they decide when an epoch is done in that case (how can you run through the entire dataset if it's an infinite stream?). So then we'll have to figure out how they do and the augmentation factor may matter or may not. I have no idea.

[gngdb]

Online augmented datasets are now preferred and don't require this.