Deployment Tools

This directory contains executable deployment helpers. The tools do not train models. They take trained checkpoints and produce deployment artifacts plus small manifests that record what was exported.

YOLO Export

export_yolo.py wraps the Ultralytics export API. It is used for PVEL-AD and PV-Multi-Defect detection checkpoints.

Input:

Input	Meaning
--model	A trained YOLO .pt checkpoint, usually from outputs/detection/<run>/weights/best.pt.
--format	Export target, such as onnx, engine, or torchscript.
--imgsz	The image size used by the exported graph. It should match the training and validation setting unless there is a reason to change it.
--device	Runtime device passed to Ultralytics. Use cpu, mps, or a CUDA device depending on the machine.

Command:

python3 deployment/export_yolo.py \
  --model outputs/detection/pvel_ad_yolo11l/weights/best.pt \
  --format onnx \
  --imgsz 640 \
  --name pvel_ad_yolo11l_onnx

Output:

Output	Meaning
Exported model file	Created by Ultralytics next to the checkpoint or in the backend-specific location it reports.
outputs/deployment/yolo/<name>.json	Manifest with checkpoint path, export format, image size, precision flags, and exported path.

The export is normal when Ultralytics prints an exported file path, the manifest exists, and the exported model can be loaded by the target backend. For ONNX, the next check should be ONNX Runtime inference on a fixed image set.

ELPV ONNX Export

export_elpv.py exports the torchvision ResNet-18 or Swin-T checkpoints produced by experiments/elpv/run_torchvision.py.

Input:

Input	Meaning
--config	The ELPV config used to define model name, task, and image size.
--checkpoint	The trained best.pt state dictionary from outputs/elpv/<name>/best.pt.
--output	Optional explicit ONNX output path.
--opset	ONNX opset version. The default is 17.

Command:

python3 deployment/export_elpv.py \
  --config configs/elpv/resnet18_binary.yaml \
  --checkpoint outputs/elpv/elpv_resnet18_binary/best.pt

Output:

Output	Meaning
outputs/deployment/elpv/<name>/model.onnx	ONNX model with dynamic batch dimension.
outputs/deployment/elpv/<name>/manifest.json	Export metadata and expected input/output contract.

The export is normal when model.onnx and manifest.json are created. For binary classification, the ONNX output is two logits. For regression, the ONNX output is one value. The input is an RGB tensor resized to the configured image size and normalized with mean and standard deviation of 0.5 for each channel.

What Should Be Compared After Export

Deployment is not complete just because a file was created. A useful deployment result should compare:

Check	Why it matters
PyTorch vs ONNX output difference	Confirms preprocessing, graph export, and output interpretation are consistent.
Latency average and P95	Shows both typical speed and tail latency.
FPS	Makes throughput comparable across models.
Detection mAP or classification F1/AUC after export	Confirms speed did not hide a quality regression.