PyTorch-YOLO_Nano

A minimal PyTorch implementation of YOLO_Nano

• Yolo_Nano
• Bag of Freebies for Training Object Detection Neural Networks

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

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