Proof run / Manufacturing vision

PCB defect detection, fine‑tuned through RailCompute.

This run shows a practical six-class object detector packaged with weights, ONNX export, inference code, metrics, training notes, and output examples. It is a workflow demonstration, not a state-of-the-art benchmark claim.

Watch training run View GitHub Book a design partner call

RailCompute run package

TaskObject detection

Classes6 defects

BaseYOLO11n

OutputPT + ONNX

Precision 0.9774

Recall 0.9896

mAP@50 0.9903

mAP@50-95 0.5951

Sample output

Before and after predictions on PCB inspection images.

Six PCB before and after detection outputs for missing hole, short, spur, spurious copper, mouse bite, and open circuit defects — Six sample detections from the trained model: missing hole, short, spur, spurious copper, mouse bite, and open circuit.

What RailCompute handled

From dataset to packaged model artifact.

The workflow handled dataset preparation, model-training planning, GPU-backed training, validation/test evaluation, and delivery of model files and documentation.

Data preparation

YOLO-format PCB inspection data, class mapping, validation/test split disclosure, and dataset attribution.

Fine-tuning run

YOLO11n base model, 80 epochs, 640px image size, batch 16, seed 42, trained on an RTX A4000.

Evaluation package

Final metrics, per-class localization notes, curves, confusion matrices, speed breakdown, and limitation notes.

Training setup

Small model, focused dataset, clear artifacts.

The run used a public six-class PCB defect dataset in YOLO object detection format. The repository includes PyTorch weights, ONNX export, an inference script, metrics JSON, training config, results CSV, and documentation.

Base modelyolo11n.pt

GPURTX A4000 / 16 GB

Epochs80

Image size640

Batch16

Splits6,370 / 802 / 829

Full training video

Watch the fully automated training run we did on RailCompute.

Artifacts

Open repo, inspectable outputs.

The GitHub repository includes the inference script, model card, dataset disclosure, checksums, weights, metrics, plots, and RailCompute training notes.

View repository Book a call Back to demos

Disclosure

This proof run is an automated training showcase, not a production-certified quality-control system or SOTA benchmark submission. Validate on your own cameras, lighting, PCB types, defect definitions, and acceptance thresholds before operational use.