TODO

* add fully randomized mini-batches (copy in samples)

* modify sampler to take a batch size, and option for truly random batches
 - then remove mini_batch function

* add ability to request items by class, or get a class specific sampler

* request samples from the dataset based on a property, such as the class
 - maybe request from the sampler, so you get the same pipeline and ordering?
 - The instances should go in to same order as was decided when creating the sampler. These for now should be:
    * random
    * sorted
    * iterate through all frames of a sub-sequence (animation), then jump to another animation

* Yale faces dataset (use the extended yale faces B)

* berkeley segmentation dataset (500 real images)
  - write some example code showing how to process a directory of images, sample
    patches, write out a dataset, etc...

* Add a video source

* add a directory of videos source
 - or maybe just a directory iterator source, that then plugs into image or
   video extractor
 - do both random and in order iteration in a directory (or with a given sort
   fn)

* Should each function of pipeline accept a function that returns the sample as parameter rather than the sample itself?

  - we could create much mmore flexible structures

```lua
  function my_pipe_function(sampler)
    local sample = sampler()
    if sample == nil then return nil end

    local function sample_is_awesome(sample)
      if sample.is_blessed then
        return true
      else
        return false
      end
    end

    while not sample_is_awesome(sample) do
       sample = sampler()
    end

    return sample
  end
```