Inference is implement on each GPU

[jiaerfei]

According to the code in tools/train_net.py, I adopted the build_evaluator.py.

I meant to implement evaluation after a certain number of iterations, and I had three GPUs. However, in the practical inference process, the test sets were loaded to each GPU. Actually, I want to implement inference on one GPU. Thus, I appreciate that if you can give me some suggestions on this issue.