Lamda-Native-Framework-based-App-with-Knowledge-Distillation-based-Rice-Leaf-Disease-Classifier

A complete mobile image classification app built with Lambda Native framework, featuring knowledge distillation and GPU-accelerated training. The app can classify images into 5 custom categories with real-time inference on mobile devices.

🚀 Features

Knowledge Distillation: Teacher-student training for optimized mobile performance

GPU-Accelerated Training: Fast model training using PyTorch with CUDA

Cross-Platform Mobile App: Built with Lambda Native (Scheme) for Android/iOS

Model Quantization: ONNX and TFLite formats for efficient mobile inference

Real-time Classification: Camera integration and gallery support

Customizable Classes: Train with your own 5 image categories

📋 Prerequisites

Hardware NVIDIA GPU (for training)

Android/iOS device (for deployment)

8GB+ RAM recommended

Software Ubuntu 20.04+ (development)

Python 3.9

CUDA 11.8+

Android SDK (for mobile builds)

🛠️ Installation & Setup

1. Clone the Repository bash git clone https://github.com/yourusername/mobilenet-v4-classifier.git cd mobilenet-v4-classifier
2. Environment Setup bash

Create and activate conda environment
conda env create -f environment.yml
conda activate mobilenet-v4-gpu

Or use the setup script

chmod +x setup_gpu.sh
./setup_gpu.sh

3. Verify GPU Setup

bash

python test_gpu_setup.py

📁 Dataset Preparation

Directory Structure
text
project/
├── train_data/
│   ├── class_0/     # Training images for class 0
│   ├── class_1/     # Training images for class 1
│   ├── class_2/     # Training images for class 2
│   ├── class_3/     # Training images for class 3
│   └── class_4/     # Training images for class 4
└── val_data/
    ├── class_0/     # Validation images for class 0
    ├── class_1/     # Validation images for class 1
    ├── class_2/     # Validation images for class 2
    ├── class_3/     # Validation images for class 3
    └── class_4/     # Validation images for class 4
Supported Image Formats
JPEG (.jpg, .jpeg)
PNG (.png)
Recommended size: 224x224 pixels
Minimum: 100 images per class for training

Setup Directories

Auto-create directory structure

chmod +x setup_directories.sh
./setup_directories.sh

Verify structure

python check_dataset.py

🧠 Model Training

1. Start Training with Knowledge Distillation

bash

python train_mobilenet_gpu.py

Training Process Teacher Model: ResNet50 (high accuracy)

Student Model: MobileNet V2 (optimized for mobile)

Knowledge Distillation: Transfers learning from teacher to student

Training Time: ~30-60 minutes on RTX 3060

Expected Accuracy: 75-85% on validation set

2. Monitor Training

Monitor GPU usage (separate terminal)

python gpu_monitor.py

Check training logs

tail -f training_logs.txt

3. Model Conversion

After training, models are automatically converted to:

ONNX format: mobilenet_model.onnx

TFLite format: model_quantized.tflite (quantized for mobile)

Class names: class_names.json

📱 Mobile App Development

Lambda Native Setup

bash

Install Lambda Native dependencies

chmod +x setup_lambda_native.sh
./setup_lambda_native.sh

Verify installation

cd ~/lambdanative
./configure
App Structure
text
lambdanative/
├── apps/
│   └── mobilenetclassifier/
│       ├── main.scm              # Main application logic
│       ├── manifest.scm          # App configuration
│       ├── ui/
│       │   └── layout.scm        # UI definitions
│       └── plugins/
│           ├── mobilenet_model.onnx
│           ├── model_quantized.tflite
│           └── class_names.json

Build Mobile App

bash

Navigate to Lambda Native

cd ~/lambdanative

Create app (if not already created)

./createapp mobilenetclassifier

Copy model files

cp mobilenet_model.onnx apps/mobilenetclassifier/plugins/
cp class_names.json apps/mobilenetclassifier/plugins/
cp model_quantized.tflite apps/mobilenetclassifier/plugins/

Build for different platforms

make android    # Build Android APK
make ios        # Build iOS app (requires macOS)
make desktop    # Build for desktop testing

🎯 App Features

User Interface Camera Integration: Take photos directly in the app

Gallery Access: Choose images from device gallery

Real-time Results: Instant classification with confidence scores

Clean Design: Simple and intuitive user interface

Classification Capabilities 5 Custom Classes: Configured during training

Confidence Scores: Percentage-based predictions

Fast Inference: Optimized for mobile devices

Offline Operation: No internet connection required

🚀 Deployment

Android Deployment bash

Build release APK

cd ~/lambdanative
make android

Install on connected device

adb install -r apps/mobilenetclassifier/mobilenetclassifier.apk

Or distribute the APK file

ls apps/mobilenetclassifier/mobilenetclassifier.apk

iOS Deployment bash

Build for iOS (requires macOS and Xcode)

make ios

Deploy via Xcode or TestFlight

Desktop Testing bash

Test on desktop before mobile deployment

make desktop ./apps/mobilenetclassifier/mobilenetclassifier

🔧 Customization

Modify Classes Edit class_names.json or retrain with new datasets:

json

["cats", "dogs", "birds", "cars", "flowers"]

Adjust Model Parameters Modify train_mobilenet_gpu.py:

python
config = {
    'training': {
        'epochs': 50,           # Training iterations
        'lr': 0.001,           # Learning rate
        'temperature': 3,      # Distillation temperature
        'alpha': 0.7           # Distillation weight
    }
}

UI Customization

Edit ui/layout.scm to modify the app interface.

📊 Performance Optimization

For Better Accuracy Increase training dataset size

Use data augmentation

Train for more epochs

Adjust knowledge distillation parameters

For Faster Inference Use TFLite with quantization

Reduce model complexity

Optimize image preprocessing

Use GPU delegation on supported devices

🐛 Troubleshooting

Common Issues GPU Not Detected

bash

nvidia-smi  # Check GPU status
python test_gpu_setup.py

Training Accuracy Low

Check dataset balance

Verify image quality and labels

Adjust learning rate

Mobile App Build Fails

bash

make clean
./configure
make android
Model Conversion Errors

bash

python convert_to_tflite.py  # Manual conversion
Logs and Debugging

bash

Check training logs

tail -f training_logs.txt

Mobile app debugging

adb logcat | grep -i lambdanative

Desktop testing

./apps/mobilenetclassifier/mobilenetclassifier --debug

📈 Results and Metrics Typical Performance Training Accuracy: 75-85%

Inference Time: <100ms on mobile

Model Size: 8-15MB (quantized)

Supported Devices: Android 8.0+, iOS 12+

Evaluation bash

Generate evaluation report

python evaluate_model.py

Test on sample images

python test_inference.py --image test_image.jpg

🤝 Contributing Fork the repository

Create a feature branch (git checkout -b feature/amazing-feature)

Commit your changes (git commit -m 'Add amazing feature')

Push to the branch (git push origin feature/amazing-feature)

Open a Pull Request

📝 License

This project is licensed under the MIT License - see the LICENSE file for details.

Copyright (c) 2025, A. N. M. Sajedul Alam

🙏 Acknowledgments

Lambda Native framework developers

PyTorch and TensorFlow teams

NVIDIA for CUDA support

MobileNet and ResNet research teams