Fork自 PaddlePaddle / PaddleDetection
fix default number of threads when inference with or without MKLDNN