ENOT-AutoDL YOLOv8 optimization on VisDrone dataset
This repository contains models accelerated with ENOT-AutoDL framework.
We trained yolov8s on VisDrone dataset and used it as our baseline.
Also we provide simple python script to measure flops and metrics.
YOLOv8 Small
| Model |
GMACs |
Image Size |
mAP50 |
mAP50-95 |
| YOLOv8 Ultralytics Baseline |
14,28 |
640 |
40,2 |
24,2 |
| YOLOv8n Enot Baseline |
8,57 |
928 |
42,9 |
26,0 |
| YOLOv8s Enot Baseline |
30,03 |
928 |
49,4 |
30,6 |
| YOLOv8s (x2) |
15,01 (x2) |
928 |
48,3 (-1,1) |
29,8 (-0,8) |
| YOLOv8s (x3) |
10,01 (x3) |
928 |
46,0 (-3,4) |
28,3 (-2,3) |
Validation
To validate results, follow this steps:
- Install all required packages:
pip install -r requrements.txt
- Use validation script:
python validate.py enot_neural_architecture_selection_x2/weights/best.pt --imgsz 928
- Use measure_macs script:
python measure_macs.py enot_neural_architecture_selection_x2/weights/best.pt --imgsz 928
