CLI Reference

All commands are available through the yolonas entry point.

Global options

Flag	Description
--version	Show version and exit
--verbose / -v	Enable debug logging
--quiet / -q	Suppress logging output

yolonas detect

Run object detection on images or video.

yolonas detect --source image.jpg --model yolo_nas_s --conf 0.25
yolonas detect --source images/ --output results/
yolonas detect --source video.mp4 --skip-frames 2

Option	Default	Description
--source	required	Image file, directory, or video path
--model	yolo_nas_s	Model variant (s/m/l)
--conf	0.25	Confidence threshold
--iou	0.7	NMS IoU threshold
--device	cuda	Device (cuda or cpu)
--output	results	Output directory
--input-size	640	Model input size
--skip-frames	0	Process every N-th frame (video)
--codec	mp4v	Video output codec

yolonas track

Track objects across a video with Deep HM-SORT. See the tracking guide for what the algorithm does and what the defaults assume about your footage.

yolonas track --source match.mp4 --classes 0
yolonas track --source match.mp4 --fusion min                  # the Deep-EIoU baseline
yolonas track --source match.mp4 --keep-all-tracks             # the paper's memory: unlimited
yolonas track --source match.mp4 --no-appearance               # motion-only, as a control

Option	Default	Description
--source	required	Video file path
--model	yolo_nas_s	Model variant (s/m/l)
--weights	—	Custom checkpoint; --model then selects the architecture
--conf	--track-low	Detection threshold. Must stay at or below --track-low, or the low-score association round never sees anything
--iou	0.7	NMS IoU threshold
--classes	all	Comma-separated class ids to track, filtered before association
--appearance / --no-appearance	on	Associate on per-object embeddings as well as motion
--fusion	harmonic	harmonic is Deep HM-SORT; min is Deep-EIoU's original
--track-high	0.6	Score at or above which a detection enters the first round
--track-low	0.4	Score below which a detection is ignored
--new-track	0.5	Lowest score that may start a track
--expansion	0.3	Box growth for the first association round
--max-lost-seconds	2.0	How long a track may go unmatched, in seconds of video — converted with the clip's own frame rate
--keep-all-tracks	off	Never drop a track. The paper's setting, and a closed-environment assumption
--class-aware	off	Refuse to associate across classes
--output	results	Output directory
--codec	mp4v	Video output codec
--show-fps	off	Burn the per-frame time into the output

Boxes are coloured by track id, not by class, so an ID-swap shows as a colour change. The summary reports unique_ids: far above the true object count means ids are fragmenting.

yolonas train

Train a YOLO-NAS model.

yolonas train --data /path/to/dataset --format yolo --epochs 100

Option	Default	Description
--data	required	Path to dataset root
--model	yolo_nas_s	Model variant
--format	yolo	Dataset format (yolo/coco)
--epochs	300	Training epochs
--batch-size	32	Batch size per GPU
--lr	2e-4	Learning rate
--device	cuda	Device
--output	runs/train	Output directory
--resume	None	Checkpoint to resume from
--input-size	640	Input size
--workers	8	DataLoader workers
--pretrained/--no-pretrained	True	Use COCO pretrained weights

yolonas export

Export model to ONNX or OpenVINO format.

yolonas export --model yolo_nas_s --format onnx
yolonas export --model yolo_nas_s --format openvino --target frigate

# Feature embeddings, alone or beside the detections
yolonas export --model yolo_nas_s --target embedding --output embedding.onnx
yolonas export --model yolo_nas_s --target combined --output combined.onnx

# Detections + a vector per detection, NMS inside the graph
yolonas export --model yolo_nas_s --target objects --output objects.onnx

Option	Default	Description
--model	yolo_nas_s	Model variant
--format	onnx	Export format (onnx/openvino)
--output	auto	Output file path
--input-size	640	Model input size
--opset	18	ONNX opset version
--checkpoint	None	Custom checkpoint path
--target	generic	Export target (generic/frigate/embedding/combined/objects)
--embed-layers	c5	Feature maps to pool, comma-separated (embedding/combined)
--embed-pooling	avg	Spatial pooling, avg or max (embedding/combined)
--normalize / --no-normalize	normalize	L2-normalize the embedding

The embedding and combined targets take a second input, valid_region — see the export guide.

yolonas eval

Evaluate model on COCO dataset.

yolonas eval --data /path/to/coco --split val2017

Option	Default	Description
--data	required	Path to COCO dataset root
--model	yolo_nas_s	Model variant
--split	val2017	Split name
--batch-size	32	Batch size
--device	cuda	Device
--conf	0.001	Confidence threshold
--iou	0.65	NMS IoU threshold