PCB-Prune-YOLO-P10-Direct

PCB-Prune-YOLO P10 Direct

Validation-selected YOLOv8n checkpoint produced by direct DepGraph structured pruning followed by matched fine-tuning on DeepPCB. This model does not use sparse learning or knowledge distillation.

Validation results

At seed 42 this direct P10 control exceeded the matched sparse-learning P10 by 1.42 mAP50-95 percentage points. This is a single-seed observation and the DeepPCB test split was not used for model selection.

Compression and Tesla T4 benchmark

Input size is 640. Latency uses 50 warm-up and 200 synchronized CUDA iterations.

Training configuration

• Direct local group-magnitude pruning, ratio 0.10, one step, no channel rounding
• AdamW, lr0=0.001, lrf=0.01, momentum 0.9, weight decay 0.0005
• 50 epochs, batch 64, patience 10, seed 42, AMP and deterministic mode
• Six classes: open, short, mousebite, spur, copper, pin-hole

Loading

Structured pruning changes the serialized architecture. Install the project so the PrunableC2f class is importable before loading:

from ultralytics import YOLO

model = YOLO("best.pt")
results = model("pcb.jpg", imgsz=640)

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

The checkpoint was verified by loading in a new process and running CUDA inference with decoded output shape [1, 10, 8400].