README.md

metadata

title: AI Fitness Coach - Fine-Tuned Personas
emoji: 🏋️
colorFrom: blue
colorTo: purple
sdk: gradio
sdk_version: 4.44.0
app_file: app.py
pinned: false
license: mit

🏋️ AI Fitness Coach with Fine-Tuned Personas

An intelligent fitness coaching system that analyzes exercise form and provides personalized feedback through fine-tuned language models.

✨ Features

• 3D Pose Estimation: Uses PoseFormer for accurate body pose tracking
• Real-time Scoring: Analyzes form and provides detailed breakdown by body part
• 4 Fine-Tuned Personas: Each with unique coaching style
  • 🔥 Hype Beast: Energetic, motivational, hype-focused
  • 📊 Data Scientist: Analytical, metrics-driven, detailed
  • 💪 No-Nonsense Pro: Direct, tough-love, results-oriented
  • 🧘 Mindful Aligner: Calm, body-awareness focused, supportive

🔬 Model Details

Training Pipeline

• Base Model: DistilGPT-2 (82M parameters)
• Training Data: 320 synthetic examples (80 per persona)
• Data Generation: GPT-4o API with persona-specific prompts
• Fine-Tuning: 1000 steps per persona, FP16 mixed precision
• Training Time: ~15 minutes per persona on RTX GPU

Architecture

User Video → MediaPipe Pose Detection → PoseFormer 3D Estimation 
→ Scoring Algorithm → Fine-Tuned Persona Model → Feedback

🚀 How It Works

1. Upload Video: Record yourself doing an exercise
2. Select Persona: Choose your preferred coaching style
3. Get Analysis: Receive detailed scoring and personalized feedback
4. Improve: Follow the coach's recommendations

📊 Performance Metrics

• Pose Detection Accuracy: 95%+ on clear videos
• Processing Time: ~3-5 seconds per video
• Model Size: ~250MB per persona (compressed)
• Inference Speed: <1 second on GPU, ~2-3 seconds on CPU

🛠️ Technology Stack

• Frontend: Gradio
• Backend: Python, PyTorch
• Pose Estimation: MediaPipe + PoseFormer
• LLM: Fine-tuned DistilGPT-2
• Training: Hugging Face Transformers, Accelerate

📝 Citation

If you use this project, please cite:

@software{ai_fitness_coach_2024,
  title = {AI Fitness Coach with Fine-Tuned Personas},
  author = {Your Name},
  year = {2024},
  url = {https://huggingface.co/spaces/your-username/ai-fitness-coach}
}

📄 License

MIT License - See LICENSE file for details

🙏 Acknowledgments

• PoseFormer for 3D pose estimation
• Hugging Face for model hosting and training tools
• OpenAI GPT-4o for synthetic training data generation
• MediaPipe for 2D pose detection