megalodon-ml/medgemma_kaggle

MedGemma Mobile (Kaggle Edition) 🏥📱

Optimized Medical VLM for Mobile Devices

This repository contains the TensorFlow Lite (TFLite) quantized models for running MedGemma on mobile devices (Android/iOS). These models are optimized for efficiency while retaining medical reasoning capabilities.

🔗 GitHub Repository: medgemma-mobile (Run the demo & source code)

📦 Models Included

Component	Filename	Description	Size
Text Model	models/text/medgemma_4b_mobile_int8_q8_ekv2048.tflite	4B param LLM, INT8/Int4 quantization, 2048 context window.	~3.9 GB
Vision	models/vision/medsiglip_vision_448.tflite	SigLIP vision encoder (448x448 input).	~1.6 GB
Projector	models/projector/multimodal_projector_448.tflite	Linear projection layer.	~11 MB
Tokenizer	tokenizers/medgemma_4b/	Tokenizer files (SPM).	-

🚀 Quick Start (Python)

To run these models, clone the GitHub repository and use the provided demo script.

# Clone the demo code
git clone https://github.com/itikelabhaskar/medgemma_kaggle
cd medgemma_kaggle

# Install dependencies
pip install -r requirements.txt

# Run the demo (Download models from this repo first!)
python demo/run_demo.py --image demo/images/image1.jpg

📱 Android Integration

These models are compatible with ai-edge-litert (TensorFlow Lite Runtime) for Android.

1. Vision Model: Input [1, 3, 448, 448], Output [1, 1024, 1152]
2. Projector: Input [1, 1024, 1152], Output [1, 1024, 2560]
3. Text Model: Input [1, seq, 2560] (embeddings) or Token IDs.

See the GitHub repo for implementation details.

⚠️ Requirements

• RAM: Minimum 8GB RAM (Unified) recommended for the 4B Text Model.
• OS: Linux/WSL recommended for execution (Windows TFLite has file mapping limits >2GB).

Credits

• Based on Google Gemma 2 and SigLIP.
• Fine-tuned for medical VQA tasks.