Can't train on MSCOCOReduced

[PavlosMelissinos]

MSCOCOReduced has a bug that crashes the program during training. Reproduce by running train.py using the following solver.json:

{
  "model_name": "enet",
  "epochs": 100,
  "batch_size": 8,
  "completed_epochs": 0,
  "dh": 256,
  "dw": 256,
  "skip": 0,
  "resize_mode": "stretch",
  "instance_mode": true,
  "dataset_name": "mscoco_reduced"
}

The problem is probably related to MSCOCO assuming that the actual classes are depicted exactly in self._coco. Directly removing the extra classes from self._coco should have worked, however that is not the case for some reason.