File	Size	Purpose
`yolo26n_seg_liquor_v1.pte`	11 MB	ExecuTorch — primary, for on-device inference
`yolo26n_seg_liquor_v1.onnx`	11 MB	ONNX — debugging, alt runtime
`yolo26n_seg_liquor_v1.metadata.yaml`	—	Ultralytics model card

Training

Base: yolo26n-seg.pt (Ultralytics pretrained on COCO)
Dataset: 11,626 images (9300 train / 1163 val / 1163 test), remapped to a single liquor_bottle class. Sources: Lamar liquor-data (CC BY 4.0), Bottles_All (Roboflow), PNU Spirit Detection, and composite shelf scenes built on SKU-110K backgrounds (training-only).
Hardware: 1× NVIDIA H100 80GB SXM5 via RunPod
Epochs: 100 (no early-stop, completed in ~109 min)
imgsz: 640
batch: 32, optimizer: SGD, lr0: 0.001, cos_lr: True
mosaic: 1.0, close_mosaic: 10

Metrics (test split, n=1163)

Metric	Value
Box mAP50	0.725
Box mAP50-95	0.503
Seg mAP50	0.679
Seg mAP50-95	0.420
Inference (H100)	3.9 ms/img

Usage (ExecuTorch / React Native)

import { useInstanceSegmentation } from 'react-native-executorch';

const model = await InstanceSegmentationModule.fromCustomModel({
  modelSource: 'https://huggingface.co/Rockenfels/yolo26n-seg-liquor/resolve/main/yolo26n_seg_liquor_v1.pte',
  classNameToIdx: { LIQUOR_BOTTLE: 0 },
});

Usage (Ultralytics / Python)

python

from ultralytics import YOLO

model = YOLO('yolo26n_seg_liquor_v1.onnx')
results = model('shelf.jpg')

License

AGPL-3.0 — inherits from Ultralytics YOLO26. Source code derived from or linking against this model must be released under a compatible copyleft license, or an Ultralytics enterprise license must be obtained. See https://ultralytics.com/license.