Allow training to succeed with negative images

[nwestlake]

Allows training to succeed when there are images present without any ground truth boxes. Previously, this failed due to merging empty matrices, argmax over empty matrices or the lack of an objective.

This supercedes 75b3fd (Assert at least one label for training) which would pick up the lack of an objective.

[MinaRe]

Just put "sudo" before your comment, and if the problem still ,you can change number of batch.

[panfengli]

When output the model result, it will get this error:

Traceback (most recent call last):
  File "./tools/train_net.py", line 92, in <module>
    max_iters=args.max_iters)
  File "/home/hust/fast-rcnn/tools/../lib/fast_rcnn/train.py", line 123, in train_net
    sw.train_model(max_iters)
  File "/home/hust/fast-rcnn/tools/../lib/fast_rcnn/train.py", line 98, in train_model
    self.snapshot()
  File "/home/hust/fast-rcnn/tools/../lib/fast_rcnn/train.py", line 56, in snapshot
    orig_0 = net.params['bbox_pred'][0].data.copy()
KeyError: 'bbox_pred'

[panfengli]

Oh, I find it is another error, since I have renamed bbox_pred to bbox_pred_inria in prototxt