I follow ALL the instructions given by the author and want to train the 2560*768 rrc_detector.

I use four GPUs (TiTian X), cuda 8.0.61 and cudnn 6.0.21.

The training speed was normal (160s per 10 iterations) at the first few iterations. However, at the 2500th iteration(after two day), the speed is lower(30min per 10 iterations).At that time, the GPU-Util is ZERO. I don't change any parameters in the examples/car/rrc_kitti_car.py.

Previously, I trained the 1280*256 rrc_detector and didn't have this issue.