integrate keras models into training pipeline

[ahundt]

based on:
#1

[BrianOn99]

In mscoco_json2seg_mask.ipynb, there is a variable anns never defined but used once.
@ahundt Are you sure the scripts are working or I miss something?

[ahundt]

@BrianOn99 don't use the ipynb, that's from changes that aren't mine. Also, note that the coco code puts everything into a single layer, while the actual labeling has overlapping data. http://mscoco.org/explore/?id=21079 is an example, and that would need to be resolved to get the results you're hoping for. One simple improvement would be putting tables walls at the end of the z buffer list and moving small stuff like cups plates etc to the front.

Or help the ongoing work in https://github.com/PavlosMelissinos/enet-keras/ for mscoco (also not ready for prime time).

[ahundt]

#25 replaces this.