Skip to content

Monoclinic/ShuffleNet-Series

 
 

Repository files navigation

ShuffleNet Series

ShuffleNet Series by Megvii Research.

Introduction

This repository contains the following ShuffleNet series models:

Trained Models

  • OneDrive download: Link
  • BaiduYun download: Link (extract code: mc24)

Details

ShuffleNetV2+

The following is the comparison between ShuffleNetV2+ and MobileNetV3. Details can be seen in ShuffleNetV2+.

Model FLOPs #Params Top-1 Top-5
ShuffleNetV2+ Large 360M 6.7M 22.9 6.7
MobileNetV3 Large 224/1.25 356M 7.5M 23.4 -
ShuffleNetV2+ Medium 222M 5.6M 24.3 7.4
MobileNetV3 Large 224/1.0 217M 5.4M 24.8 -
ShuffleNetV2+ Small 156M 5.1M 25.9 8.3
MobileNetV3 Large 224/0.75 155M 4.0M 26.7 -

ShuffleNetV2

The following is the comparison between ShuffleNetV2 and MobileNetV2. Details can be seen in ShuffleNetV2.

Model FLOPs #Params Top-1 Top-5
ShuffleNetV2 2.0x 591M 7.4M 25.0 7.6
MobileNetV2 (1.4) 585M 6.9M 25.3 -
ShuffleNetV2 1.5x 299M 3.5M 27.4 9.4
MobileNetV2 300M 3.4M 28.0 -
ShuffleNetV2 1.0x 146M 2.3M 30.6 11.1
ShuffleNetV2 0.5x 41M 1.4M 38.9 17.4

ShuffleNetV2.Large

The following is the comparison between ShuffleNetV2.Large and SENet. Details can be seen in ShuffleNetV2.Large.

Model FLOPs #Params Top-1 Top-5
ShuffleNetV2.Large 12.7G 140.7M 18.56 4.48
SENet 20.7G - 18.68 4.47

ShuffleNetV2.ExLarge

The following is the result of ShuffleNetV2.ExLarge. Details can be seen in ShuffleNetV2.ExLarge.

Model FLOPs #Params Top-1 Top-5
ShuffleNetV2.ExLarge 46.2G 254.7M 15.52 2.9

ShuffleNetV1

The following is the comparison between ShuffleNetV1 and MobileNetV1. Details can be seen in ShuffleNetV1.

Model FLOPs #Params Top-1 Top-5
ShuffleNetV1 2.0x (group=3) 524M 5.4M 25.9 8.6
ShuffleNetV1 2.0x (group=8) 522M 6.5M 27.1 9.2
1.0 MobileNetV1-224 569M 4.2M 29.4 -
ShuffleNetV1 1.5x (group=3) 292M 3.4M 28.4 9.8
ShuffleNetV1 1.5x (group=8) 290M 4.3M 29.0 10.4
0.75 MobileNetV1-224 325M 2.6M 31.6 -
ShuffleNetV1 1.0x (group=3) 138M 1.9M 32.2 12.3
ShuffleNetV1 1.0x (group=8) 138M 2.4M 32.0 13.6
0.5 MobileNetV1-224 149M 1.3M 36.3 -
ShuffleNetV1 0.5x (group=3) 38M 0.7M 42.7 20.0
ShuffleNetV1 0.5x (group=8) 40M 1.0M 41.2 19.0
0.25 MobileNetV1-224 41M 0.5M 49.4 -

OneShot

The following is the comparison between Single Path One-Shot NAS and other NAS counterparts. Details can be seen in OneShot.

Model FLOPs #Params Top-1 Top-5
OneShot 328M 3.4M 25.1 8.0
NASNET-A 564M 5.3M 26.0 8.4
PNASNET 588M 5.1M 25.8 8.1
MnasNet 317M 4.2M 26.0 8.2
DARTS 574M 4.7M 26.7 8.7
FBNet-B 295M 4.5M 25.9 -

DetNAS

The following is the performance of DetNAS on ImageNet, compared with ResNet. Details can be seen in DetNAS.

Model FLOPs #Params Top-1 Top-5 mAP*
300M (VOC, RetinaNet) 300M 3.5M 25.4 8.1 80.1
300M (VOC, FPN) 300M 3.7M 25.9 8.3 81.5
300M (COCO, RetinaNet) 300M 3.7M 26.0 8.4 33.3
300M (COCO, FPN) 300M 3.5M 26.2 8.4 36.4
1.3G (COCO, FPN) 1.3G 10.4M 22.8 6.5 40.0
3.8G (COCO, FPN) 3.8G 29.5M 21.6 6.3 42.0
ResNet50 (COCO, FPN) 3.8G - 23.9 7.1 37.3
ResNet101 (COCO, FPN) 7.6G - 22.6 6.4 40.0

*COCO models are coming soon.

About

No description, website, or topics provided.

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • Python 100.0%