Strongeryolo-pytorch

Introduction

This project is inspired by Stronger-Yolo. I reimplemented with Pytorch and continue improving yolov3 with latest papers.
This project will also try out some model-compression approaches(e.g. channel-pruning).
See reimplementation results in MODELZOO.

Environment

python3.6, pytorch1.2(1.0+ should be ok), ubuntu14/16/18 tested.

Quick Start

All checkpoints as well as converted darknet can be downloaded here.链接 提取码: i3pa
See Usage.md for details.

Improvement with latest papers(Using StrongerV3 as baseline)

All results all tested with 544*544 and threshold 0.1

model	mAP50	mAP75	configs	baseline
baseline(with GIOU)	79.6	43.4	strongerv3.yaml	-
+ kl loss&&varvote	78.9	49.2	strongerv3_kl.yaml	strongerv3.yaml
+ ASFF	80.6	46.6	strongerv3_asff.yaml	strongerv3.yaml
+ All improvement	81.1	53.0	strongerv3_all.yaml	strongerv3.yaml

Note:
1.Set EVAL.varvote=True to enable varvote in KL-loss. According to the paper, kl-loss(and varvote) can strongly boost the performance of mAP75(or higher), but decrease mAP50 slightly.
2.The details(e.g. channel number) of ASFF module is not completely the same as the original implementation.
3.The All version including other small tricks like removing relu in detection head. Check config file for details.

Performance on VOC2007 Test(mAP) after pruning

Model	Backbone	MAP	Flops(G)	Params(M)
strongerv3	Mobilev2	79.6	4.33	6.775
strongerv3-sparsed	Mobilev2	77.4	4.33	6.775
strongerv3-Pruned(30% pruned)	Mobilev2	77.1	3.14	3.36
strongerv2	Darknet53	80.2	49.8	61.6
strongerv2-sparsed	Darknet53	78.1	49.8	61.6
strongerv2-Pruned(20% pruned)	Darknet53	76.8	49.8	45.2

Note:
1.Tuning _C.Prune.sr can get better prune ratio, I picked the official number 0.01.

Deployment on mobile devices(Update 2019-12-14)

check deploy.md for more details.

• pytorch -> tensorflow pb
• pytorch -> MNN(Alibaba)
• pytorch -> NPU(HUAWEI)

Supported backbone

• MobileV2(Pruning suppoted)
• DarkNet(Pruning supported) ...

Reference

Stronger-Yolo
focal-loss
kl-loss ASFF