Awesome

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

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

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