thangkt/PCB-Prune-YOLO-P40-A8-Direct

PCB-Prune-YOLO P40-A8 Direct

YOLOv8n checkpoint produced by DepGraph structured pruning at a hardware latency-first target ratio of 0.40 (round_to=8, "A8" candidate), then fine-tuned with plain AdamW. This is the standard-fine-tune control for the matched P40-A8 knowledge-distillation checkpoint, thangkt/PCB-Prune-YOLO-P40-A8-KD. This model does not use sparse learning or knowledge distillation.

Validation results

Precision	Recall	mAP50	mAP50-95
0.929	0.888	0.950	0.701

An initial 50-epoch fine-tune reached only mAP50-95 0.634; both branches were still improving at the final epoch (not converged), so training was extended to 100 epochs with cosine LR annealing, after which both plateau. The matched knowledge-distillation checkpoint reaches mAP50-95 0.712 under an otherwise identical fine-tune recipe (+1.1 percentage points). Versus P30 direct (0.75030), this checkpoint is 4.93 points lower for 37.8% fewer parameters and 42.85% fewer MACs. The DeepPCB test split was not used for model selection.

Compression and Tesla T4 benchmark

Parameters	MACs	Size	Latency batch 1 (PyTorch)	FPS
903,466	1.1212G	1.959 MiB	7.625 ms	131.15

Versus baseline (3,012,018 params, 4.0733G MACs): -70.00% parameters, -72.47% MACs. Input size is 640. Latency uses 50 warm-up and 200 synchronized CUDA iterations. On the same hardware-latency gate that selected this architecture, a same-session rebuilt TensorRT FP16 engine measured 1.422 ms forward (50 warm-up / 200 iterations, batch 1), about 1.21x faster than a same-session baseline TensorRT engine (1.716 ms); TensorRT engines are not included in this repository.

Training configuration

• DepGraph local group-magnitude pruning, target ratio 0.40, round_to=8
• AdamW, lr0=0.001, lrf=0.01, momentum 0.9, weight decay 0.0005, cosine LR
• 100 epochs, batch 64, patience 20, 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].