train dataset: VOC 2012 + VOC 2007
test size: 544
test code: Faster rcnn (not use 07 metric)
test GPU: 12G 2080Ti
test CPU: E5-2678 v3 @ 2.50GHz

Version	Network	Backbone	Initial weight	VOC2007 Test(mAP)	Inference(GPU)	Inference(CPU)	Params
V1	YOLOV3	Darknet53	YOLOV3-608.weights	88.8	30.0ms	255.8ms	248M
V2	YOLOV3	Darknet53	Darknet53_448.weights	83.3	30.0ms	255.8ms	248M
V3	YOLOV3-Lite	MobilenetV2	MobilenetV2_1.0_224.ckpt	79.4	18.9ms	80.9ms	27.3M

Check Strongeryolo-pytorch for pytorch version with channel-pruning.
There is also a MNN Demo for Verson V3.