PCB-Prune-YOLO-P40-A8-KD

PCB-Prune-YOLO P40-A8 KD

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 Ultralytics' native knowledge distillation, teacher = thangkt/PCB-Prune-YOLO-Baseline. This is the strongest-compression checkpoint in the project so far. Its matched standard-fine-tune control (identical hyperparameters, no distillation) is thangkt/PCB-Prune-YOLO-P40-A8-Direct.

Validation results

An initial 50-epoch fine-tune reached only mAP50-95 0.660, with the best validation epoch landing on the very last epoch (not converged), so training was extended to 100 epochs with cosine LR annealing, after which the curve plateaus. Under an otherwise identical fine-tune recipe, this checkpoint beats the matched standard fine-tune by +1.1 mAP50-95 percentage points (0.712 vs 0.701). A dis sweep (3.0 and 10.0) confirmed the default dis=6.0 used here was already near-optimal (0.711 / 0.712 / 0.709, within noise). Versus P30 direct (0.75030, -51.77%/-51.83% params/MACs), this checkpoint is now only 3.83 points lower for 37.8% fewer parameters and 42.85% fewer MACs — a substantially more competitive trade-off than the initial 50-epoch result. 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; PyTorch batch-1 latency on a shared cloud GPU varies roughly ±5% between runs of the identical checkpoint, so treat single-sample figures as approximate. Fine-tuning does not change channel counts, so a same-session rebuilt TensorRT FP16 engine for this checkpoint measured 1.497 ms on first measurement; averaging 4 repeated measurements of the same unmodified engine gives 1.494 ms versus a same-session baseline average of 1.702 ms — about 1.14x (≈14%) faster than baseline, 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
• Knowledge distillation: Ultralytics 8.4.115 native distill_model + dis=6.0 (framework default, confirmed near-optimal by a 3.0/10.0 sweep) — a score-weighted feature L2 loss between teacher and student at the Detect head's input layers, on top of the normal detection loss. Teacher checkpoint: baseline best.pt, frozen.
• 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].