Skip to content

Latest commit

 

History

History
 
 

benchmark

README.md

MobileNet-YOLO Caffe

MobileNet-YOLO benchmark

By using tencent/ncnn framework

ubuntu18 , intel i5-7500

loop_count = 8
num_threads = 4
powersave = 0
      squeezenet  min =   27.47  max =   27.62  avg =   27.52
       mobilenet  min =   21.06  max =   22.03  avg =   21.60
    mobilenet_v2  min =   15.84  max =   16.63  avg =   16.16
      shufflenet  min =    7.65  max =    8.54  avg =    7.88
       googlenet  min =  132.11  max =  134.41  avg =  132.82
        resnet18  min =  197.74  max =  201.52  avg =  198.61
         alexnet  min =   77.14  max =   79.23  avg =   77.58
           vgg16  min = 1130.89  max = 1135.79  avg = 1132.22
  squeezenet-ssd  min =   92.09  max =   93.14  avg =   92.70
   mobilenet-ssd  min =   49.84  max =   50.99  avg =   50.39
mobilenet-yolov3-320  min =   56.83  max =   57.72  avg =   57.34
mobilenet-yolov3-416  min =   94.59  max =   98.39  avg =   95.84

Model download

MobileNet-YOLOv3-Lite-VOC

Darknet-YOLOv3

inference time issue

The most time consuming layer were group convolution , not deconvolution

int8 inference mAP list

https://github.com/eric612/caffe-int8-convert-tools

Network mAP Resolution Data Type Framework Bit
MobileNet-YOLOv3 0.737 416 float caffe 32
MobileNet-YOLOv3 0.729 416 float ncnn 32
MobileNet-YOLOv3 0.634 416 int ncnn 8
MobileNet-YOLOv3 0.668 416 mixed precision : remove conv0 ncnn 8

To do