mindspore_lite_flutter
A Flutter plugin for on-device AI inference using Huawei MindSpore Lite 1.7.0.
Built for the AgriChain project — crop disease detection and yield prediction for smallholder farmers in Africa.
Features
- 🧠 Run
.msMindSpore Lite models on-device (no internet required) - 📷 Image classification from camera or file path
- ⚡ ARM64 optimised — runs on any Android device with ARM64 architecture
- 🌽 Pre-trained models for maize and coffee disease detection included in AgriChain
Platform Support
| Android | iOS |
|---|---|
| ✅ ARM64 | ❌ Not yet |
iOS support is planned. MindSpore Lite has an iOS runtime but integration is not yet complete.
Installation
dependencies:
mindspore_lite_flutter:
git:
url: https://github.com/Mas-aaron/agri-chain.git
path: mindspore_lite_flutter
Or once published to pub.dev:
dependencies:
mindspore_lite_flutter: ^1.0.0
Usage
1. Add your .ms model to Flutter assets
# pubspec.yaml
flutter:
assets:
- assets/models/maize_model.ms
- assets/models/maize_labels.txt
2. Initialize the model
import 'package:mindspore_lite_flutter/mindspore_lite_flutter.dart';
final success = await MindsporeLiteFlutter.initModel(
'assets/models/maize_model.ms',
);
3. Run inference with labels (recommended)
final results = await MindsporeLiteFlutter.predictImageWithLabels(
imagePath: imageFile.path,
labels: ['Healthy', 'Northern Leaf Blight', 'Common Rust', 'Gray Leaf Spot'],
);
// results is sorted by confidence:
// {'Healthy': 0.92, 'Common Rust': 0.05, ...}
final topLabel = results.keys.first;
final confidence = results.values.first;
print('$topLabel: ${(confidence * 100).toStringAsFixed(1)}%');
4. Or run raw inference and apply softmax yourself
final result = await MindsporeLiteFlutter.predictImage(imageFile.path);
// result contains:
// {
// 'predictions': [0.92, 0.05, 0.03], // raw logits per class
// 'inferenceMs': 45, // inference time in ms
// }
5. Dispose when done
await MindsporeLiteFlutter.disposeModel();
API Reference
| Method | Description |
|---|---|
initModel(String modelPath) |
Load a .ms model from Flutter assets |
predictImage(String imagePath) |
Run inference, returns raw logits + inference time |
predictImageWithLabels({imagePath, labels}) |
Run inference, applies softmax, returns Map<String, double> sorted by confidence |
disposeModel() |
Release native resources |
Model Format
Models must be in MindSpore Lite .ms format. To convert from .mindir:
# Using MindSpore Lite converter
converter_lite \
--fmk=MINDIR \
--modelFile=model.mindir \
--outputFile=model.ms \
--targetDevice=ARM64
Training Your Own Models
See the AgriChain training notebook for a complete example using MobileNetV2 on Huawei ModelArts.
Requirements
- Android API 21+ (Android 5.0)
- ARM64-v8a architecture
- MindSpore Lite 1.7.0 AAR (bundled)
License
MIT — see LICENSE
Credits
Built by the AgriChain team. Uses Huawei MindSpore Lite runtime.