Skip to content
master
Go to file
Code

Latest commit

 

Git stats

Files

Permalink
Failed to load latest commit information.
Type
Name
Latest commit message
Commit time
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

README.md

PyTorch-YOLO_Nano

A minimal PyTorch implementation of YOLO_Nano

Trick here I have done

Bag of Freebies for Training Object Detection Neural Networks tell us that fixup in object detection can increase the mAP, So I realize it and test in result.

  • Data Augmentation
  • Fixup
  • Cosine lr decay
  • Warm up
  • multi-GPU

Download COCO

$ cd data/
$ bash get_coco_dataset.sh

Module Pipeline

Pipeline

training

bash train.sh  

Better Para:  
   --epochs 120  
   --batch_size 8  
   --model_def ./config/yolo-nano_person.cfg  
   --lr 2.5e-4  
   --fix_up True  
   --lr_policy cosine

Testing

python test.py --data_config ./config/coco_person.data --model_def ./config/yolo-nano_person.cfg --weights_path [checkpoint path]

Result

In this engineer we only train our model using coco-train person class
we compare with yolov-3,yolo-tiny. We got competitive results.

Methods mAP@50 mAP weights FPS Model
yolov3(paper) 74.4 40.3 204.8M 28.6FPS Google Disk
yolov3-tiny(paper) 38.8 15.6 35.4M 45FPS Google Disk
yolo-nano 55.6 27.7 22.0M 40FPS Baidu WebDisk

Baidu WebDisk Key: p2j3

Ablation Result

Augmentation fixup mAP
No No 54.3
Yes No 53.9
No YES 55.6
YES YES 54.8

Inference Result

Pipeline

About

Pytorch implementation of yolo_Nano for pedestrian detection

Resources

Releases

No releases published

Packages

No packages published
You can’t perform that action at this time.