Improving classification performance.

[adamltyson]

General issue to track ideas for how to improve classification performance. Possible areas to investigate include:

Easier

• Look at effects of batch size on processing time. Can we set the most appropriate batch size based on computing resources available?
• Look at effects of multiprocessing. Is there a generally optimal number of workers?
• Improve batch generation. The cube generator does not always return "full" batches (i.e. if there is only one cell candidate in a plane, that batch will only have one set of cuboids).
• Caching - the generator doesn't cache any data, so once planes 0-19 are loaded, the next set of batches will reload 1-20.

More difficult

• How accurate can we get without the background channel?
• Investigate the network itself:
  • What is the optimal network depth?
  • Can the network be trained in a more optimal way?
  • Can we use another, more efficient network architecture?

cc @satyakam7