modelport_onnx

ONNX Runtime engine for ModelPort. Runs the onnx variants of modelport.json bundles through flutter_onnxruntime.

await ModelPortFlutter.init(adapters: [OnnxAdapter()]);
final classifier = await ImageClassifier.load(
  'https://github.com/ayanparvaiz/modelport/releases/download/zoo-v1/mobilenet_v3_small.json',
);

• Tensor details come from the manifest. ONNX Runtime cannot report input and output names, types, or shapes on iOS and macOS. The manifest has them, so the same code works everywhere.
• fp32, fp16, and int8 variants. All three MobileNetV3 variants passed their golden checks on an ARMv8.0 phone, including fp16.
• Native memory is freed after every run. Input and output tensors are disposed as soon as the results are copied out.

Setup

Android: add this line to android/app/proguard-rules.pro for release builds:

-keep class ai.onnxruntime.** { *; }

macOS: set the deployment target to 14.0, which flutter_onnxruntime requires.

Choosing a variant

await ImageClassifier.load(location, variantId: 'onnx-int8');

Without variantId, the first variant listed in the manifest that this engine can run is used.

Speed

MobileNetV3 Small on an OPPO CPH1937 (Snapdragon 665): 77 ms per run. ExecuTorch ran the same model in 17 ms, so consider modelport_executorch too. ONNX Runtime adds about 29 MB to an arm64 APK.

Example

The example app classifies a photo with each variant and runs the golden check.

License

Apache-2.0