OMM during training in Object detection with batch size>12

I am trying to train the ssd_inception_v2

the training break with the error
the result error:
ResourceExhaustedError (see above for traceback): OOM when allocating tensor with shape[1917,1]

when start,I get the message bellow:

Total memory: 11.90GiB
Free memory: 11.75GiB

Ignoring device specification /device:GPU:0 for node 'prefetch_queue_Dequeue' because the input edge from 'prefetch_queue' is a reference connection and already has a device field set to /device:CPU:0

I wonder what the message mean,I use nvidia-smi to see the gpu GPU-Util is not full,sometimes 0%.and if the batch size>12,run out with OOM.if the batch size<12,then it's fine
Which thing should I fix to get out of the error?

This is related to issue #1390 #2038