thangkt/PCB-Prune-YOLO-P10-DepGraph

PCB-Prune-YOLO P10 DepGraph

Fine-tuned structurally pruned YOLOv8n detector for the six DeepPCB defect classes: open, short, mousebite, spur, copper, and pin-hole.

Pipeline

The checkpoint was produced with:

YOLOv8n baseline
→ DepGraph GroupNormPruner sparse training (reg=5e-4, alpha=4, 30 epochs)
→ 10% no-round group-magnitude structured pruning
→ fine-tuning (best epoch 27, stopped epoch 37)

Model selection used the validation split only. The DeepPCB test split was not used to select sparse-training, pruning, or fine-tuning settings.

Validation results

Model	Params	MACs	mAP50	mAP50-95	T4 latency	FPS
YOLOv8n baseline	3,012,018	4.0733G	0.98630	0.78524	8.289 ms	120.64
P10 before fine-tuning	2,415,613	3.2328G	0.00243	0.00035	10.127 ms	98.75
P10 after fine-tuning	2,415,613	3.2328G	0.98124	0.76318	9.719 ms	102.89

Relative to the baseline, the fine-tuned P10 checkpoint has 19.80% fewer parameters and 20.63% fewer MACs, with a 2.21-point mAP50-95 drop. It does not provide a batch-1 speedup on Tesla T4: measured latency is 17.25% higher.

The sparse regularizer changed gradients in every sparse-training epoch without introducing non-finite values. However, mean and median group norm moved only -0.0042% and -0.0181%, and the measured near-zero fraction remained zero. This checkpoint therefore demonstrates post-pruning accuracy recovery, not strong group sparsification or deployment acceleration.

Loading

Structured pruning changes the architecture. Clone and install the project so the serialized PrunableC2f class is available, then load the complete model:

git clone https://github.com/pnthang04/PCB-Prune-YOLO.git
cd PCB-Prune-YOLO
pip install -e . --no-deps

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

checkpoint = hf_hub_download(
    repo_id="thangkt/PCB-Prune-YOLO-P10-DepGraph",
    filename="best.pt",
)
model = YOLO(checkpoint)
results = model.predict("pcb.jpg", imgsz=640)

Artifacts

• best.pt: complete fine-tuned structurally pruned model.
• depgraph_sparse_reg5e4.yaml: sparse-training configuration.
• summary.json / summary.csv: comparison and provenance.
• metrics_val.json / metrics_val.csv: validation metrics including classes.
• benchmark.json / benchmark.csv: synchronized batch-1 Tesla T4 benchmark.

Project repository: https://github.com/pnthang04/PCB-Prune-YOLO