MobileNetV2 Android Deployment

1. Project Objective

This project aims to deploy a pre-trained MobileNetV2 deep learning model for image classification on an Android device. The model is converted from PyTorch to TFLite and integrated into an Android application, enabling users to classify images using their mobile phones.

2. Technical Structure

2.1 Software Stack

• Model Training & Conversion: PyTorch → ONNX → TensorFlow → TFLite
• Mobile App Development: Android Studio (Kotlin-based)
• Model Deployment: TensorFlow Lite (TFLite) on Android
• Version Control: Git, GitHub

2.2 Hardware

• Device: Android smartphone (Samsung tested, API Level 31+)

2.3 Programming Languages

• Python (Model training & conversion)
• Kotlin (Android application development)

3. How to Use the Files

3.1 Python Files (Model Training & Conversion)

• MobileNetV2.py: Loads and tests MobileNetV2 in PyTorch.
• torch_to_tflite.py: Converts PyTorch model to TFLite. This is the only file you need to run after installed requirements files.
• imagenet_classes.txt: Class labels for classification.
• tensorflow_requirements.txt: Dependencies for TensorFlow.
• pytorch_requirements.txt: Dependencies for PyTorch.
• typing_extensions_requirements.txt: Dependencies for typing_extensions.

3.2 Android Files (App Implementation)

Create an new project in Android studio and replace or create below files.

• app/src/main/java/com/example/mobilenetv2/MainActivity.kt: Main application logic.
• app/src/main/res/layout/activity_main.xml: UI layout.
• app/src/main/AndroidManifest.xml: App permissions and configuration.
• app/build.gradle.kts: Dependencies for TensorFlow Lite.
• assets/mobilenet_v2.tflite: Pre-trained classification model.
• assets/imagenet_classes.txt: Labels used for inference.

Set phone to USB debug model and connect with device. Sync, build and Run.

4. Troubleshooting Experience

4.1 Handling Dependency Conflicts in Python

Issue: TensorFlow and PyTorch dependencies caused conflicts with typing-extensions.

Solution: Split dependencies into separate files and install them sequentially.

pip install -r pytorch_requirements.txt  # Install PyTorch dependencies
pip install -r tensorflow_requirements.txt  # Install TensorFlow dependencies
pip install -r typing_extensions_requirements.txt  # Fix conflicts

4.2 Fixing Android Forecasting Issues

Issue: The same image predicted correctly in Python, but incorrectly in Android.

Troubleshooting Process:

1. Verified that the TFLite model outputs correctly in Python.
2. Compared input values before inference in Python and Android.
3. Found that channel ordering was incorrect in Android.

Step	Python Implementation	Android Implementation	Match?
Image Resizing	image.resize((224, 224), Image.BILINEAR)	Bitmap.createScaledBitmap(bitmap, 224, 224, true)	✅ Yes
Convert Float	np.array(image, dtype=np.float32) / 255.0	((pixel shr X and 0xFF) / 255.0f)	✅ Yes
Standardization	(image - mean) / std	((pixel_component / 255.0f) - mean) / std	✅ Yes
Reorder Channels	np.transpose(image, (2, 0, 1))	❌ Not implemented in Android	❌ No

Solution: Explicitly reorder channels from HWC → CHW format in Android before passing input to the model.

5. Personal Blog

For a detailed step-by-step breakdown, visit: wenyang.xyz/article/mlAndroid

6. Next Step

The Android app's loading speed is reasonable but has noticeable delays, leaving room for improvement.