thangkt/PCB-Prune-YOLO-Baseline

PCB-Prune-YOLO YOLOv8n Baseline

YOLOv8n baseline trained on DeepPCB to detect six PCB defect classes: open, short, mousebite, spur, copper, and pin-hole.

Validation results

The best checkpoint was selected on the validation split at epoch 98 of 100.

Metric	Value
Precision	0.96545
Recall	0.97221
mAP@0.5	0.98630
mAP@0.5:0.95	0.78524

These are validation results. The official test split should only be evaluated for the final report.

Training configuration

• Model: YOLOv8n pretrained checkpoint
• Image size: 640
• Global batch size: 128 (64 per GPU)
• Hardware: 2x Tesla T4
• Maximum epochs: 100
• Best epoch: 98
• AMP: enabled
• Seed: 42

Usage

from ultralytics import YOLO

model = YOLO("best.pt")
results = model.predict("pcb.jpg")

The checkpoint is stored as best.pt. args.yaml and results.csv contain the training configuration and epoch history.

Source code: https://github.com/pnthang04/PCB-Prune-YOLO