Yolov5-Processing

accomplished

• 2021.12.15
  • change backbone to Ghostnet
  • Finish EagleEye pruning YOLOv5 series
• 2021.12.27
  • change backbone to shufflenetv2
  • change backbone to efficientnetv2
• 2022.05.23
  • Finish network slimming pruning YOLOv5 series
  • Finish network SFP pruning YOLOv5 series
  • Finish network FPGM pruning YOLOv5 series

Requirements

pip install -r requirements.txt

Usage

different backbone

such as ghostnet

python train.py --data data/VisDrone.yaml --imgsz 640 --weights '' --cfg models/yolo_ghostnet.yaml --nosave  --device 0,1,2,3 --sync-bn

You can change depth_multiple and width_multiple to choose different yolov5 verson

prune

EagleEye

1. Normal Training

python train.py --data data/VisDrone.yaml --imgsz 640 --weights yolov5s.pt --cfg models/yolov5s_pruning.yaml --nosave  --device 0,1,2,3 --sync-bn 

2. Search for Optimal Pruning Network

python eagleeye.py --data data/VisDrone.yaml --weight the_first_step_trained_model --cfg models/yolov5_pruning.yaml --path models/yolov5s_pruned.yaml --pruned_weights pruned_weight.pt

3. Fine-tuning

python train.py --data data/VisDrone.yaml --imgsz 640 --weights pruned_weight.pt --cfg models/yolov5s_pruned.yaml  --device 0,1,2,3 --nosave --sync-bn

FPGM / SFP

cd yolov5_fpgm_slimming_sfp

1. soft mask

python train.py --data data/VisDrone.yaml --imgsz 640 --weights yolov5s.pt --cfg models/yolov5s_pruning.
yaml --device 0,1,2,3 --sfp/fpgm --sfp_ratio/fpgm_ratio 0.5 --path models/yolov5s_pruned.yaml

2. Fine_tuning

python train.py --data data/VisDrone.yaml --imgsz 640 --weights pruned_weights.pt --cfg models/yolov5s_fpgm/sfp_pruned.yaml --device 0,1,2,3

network_slimming

cd yolov5_fpgm_slimming_sfp

1. BatchNorm Layer \gamma

python train.py --data data/VisDrone.yaml --imgsz 640 --weights yolov5s.pt --cfg models/yolov5s_pruning.yaml --device 0,1,2,3 -sr 

2. BatchNorm Layer pruning

python prune_slimming.py --weights the_first_step_trained_model --data data/VisDrone.yaml --device 0

3. Fine_tuning

python train.py --data data/VisDrone.yaml --imgsz 640 --weights the_second_step_get_model --cfg models/yolov5s_slimm_pruned.yaml --device 0,1,2,3

Results

Models	mAP@.5	mAP@.5:.95	GFLOPS	Parameters(M)
yolov5s	35.1	19.4	15.9	14.4
yolov5l_Ghostnet	33.1	18.2	42.7	49.4
yolov5l_efficientnetv2	23.3	11.4	35.3	42.8
yolov5L_shufflenetv2	29.0	15.2	38.0	40.2
yolov5s_eagleeye	30.0	15.5	8.6	8.0

TO DO

• backbone: ShuffleNetV2
• backbone: EfficientNetV2
• backbone: SwinTrans
• Prune: Other Algorithms
• Quantization
• Knowledge Distillation