infer_ensemble

Run inference with multiple checkpoints and fuse the detections.

Overview

infer_ensemble runs inference on a single image using multiple YOLO checkpoints, then fuses the detections using class-aware NMS-style deduplication. Returns the combined detection set.

Input Schema

Parameter	Type	Required	Default	Description
checkpoints	list[string]	Yes	—	List of checkpoint paths or names.
image_path	string	Yes	—	Path to the input image.
fusion_method	string	No	"wbf"	Detection fusion method: wbf (weighted box fusion).
weights	list[float]	No	null	Per-checkpoint weights for fusion.
imgsz	integer	No	640	Inference image size.
conf	float	No	0.25	Confidence threshold.
device	string	No	"auto"	Inference device.

Output Schema

Field	Type	Description
checkpoints	list[string]	Checkpoints used.
image_path	string	Input image path.
fusion_method	string	Fusion method applied.
detections	list[object]	Fused detection list.
detection_count	integer	Number of fused detections.

Examples

CLI

curl -X POST http://127.0.0.1:7823/tools/infer_ensemble \
  -H "Authorization: Bearer $(cat ~/.fovux/auth.token)" \
  -H "Content-Type: application/json" \
  -d '{"checkpoints": ["yolov8n.pt", "yolov8s.pt"], "image_path": "/data/test.jpg"}'

Python

from fovux.tools.infer_ensemble import infer_ensemble
result = infer_ensemble(["yolov8n.pt", "yolov8s.pt"], "/data/test.jpg", fusion_method="wbf")

Notes & Limits

• Detections are deduplicated using IoU > 0.5 for same-class boxes.
• Higher-confidence detections are kept when overlapping boxes are found.
• Processing time scales linearly with the number of checkpoints.

Failure Modes

• Checkpoint resolution errors for any invalid checkpoint.
• File not found if the image path does not exist.