Model Benchmark Script

This folder contains benchmark_models.py, which compares a local ONNX/TFLite model against a Roboflow YOLO model on a YOLO-format test set.

The ONNX model used here was created with One AI from One Ware.

A full walkthrough of this benchmark and the dataset is available in the demo: Lemons.

Roboflow API key (required)

Before running, set your API key:

export ROBOFLOW_API_KEY=your_api_key_here

Run

python benchmark_models.py --model lemons.onnx --dataset yolo_dataset

Parameters

• --model Path to local .onnx or .tflite model.
• --yolo-model Roboflow model ID (default: limeline/1).
• --dataset YOLO dataset root (expects images/test and labels/test).
• --max-images Limit number of test images (0 = all).
• --output-json Output path for benchmark results JSON.
• --iou-threshold IoU threshold for TP/FP/FN matching.
• --class-agnostic-metrics Ignore class IDs during matching.
• --onnx-class-offset Offset added to ONNX predicted classes (default: -1).
• --onnx-box-format ONNX box format: xywh or xyxy.
• --onnx-coords ONNX coordinate scale: auto, pixels, or normalized.

License

YOLO baseline models in this benchmark were trained using Roboflow and YOLO tooling.

Some compontents used in the evaluation pipeline are licensed under APGL-3.0. To comply with this license, the full benchmark script used for evaluation is included in this repository.

This repository only contains the evaluation script and does not redistribute Roboflow or YOLO source code.