Spaces:
Sleeping
Sleeping
metadata
marp: true
theme: gaia
_class: lead
paginate: true
backgroundColor: null
backgroundImage: url('https://marp.app/assets/hero-background.svg')
BlossomTune 🌸 Orchestrating Federated Learning with Flower & Gradio
A Technical Overview
The Challenge: Operational Complexity in FL
- Participant Onboarding & Configuration
- Infrastructure Management & Monitoring
- Experiment Coordination & Scaling
Our Solution: BlossomTune
A web-based orchestrator for the entire FL lifecycle.
Core Technologies:
- Flower: The FL Framework
- Gradio: The Interactive Web UI
- Hugging Face: User Authentication & ML Models
Key Features
- Centralized Federation Control (Admin)
- Streamlined Participant Onboarding
- Live System Monitoring
System Architecture
Codebase Deep Dive: Structure & Quality
- Modular & Decoupled Structure (blossomtune_gradio vs. flower_apps)
- Centralized Configuration (config.py)
- Clear UI/Backend Separation (ui/ package)
- High Code Quality (Enforced by ruff & pre-commit hooks)
The Participant Journey
- Request Access (Login & Submit Email)
- Activate (Verify with Email Code)
- Admin Review (Request appears in Admin Panel)
- Approval & Configuration (Admin assigns Partition ID)
- Connect (User receives connection details)
The Admin Experience
- One-Click Infrastructure Management
- Controlled Experiment Execution
- Intuitive Request Management
Example FL App: quickstart_huggingface
- Task: Sentiment Analysis (IMDB Dataset)
- Model: bert-tiny (Lightweight Transformer)
- Federation: Client reads partition-id from orchestrator's configuration.
Conclusion & Future Work
Conclusion: A high-quality, robust orchestrator that solves key operational challenges in FL.
Future Work:
- Enhanced monitoring with visualizations & metrics
- Support for dynamic selection of multiple Flower Apps
- Granular control over Runner configurations
- Integration with other authentication providers
Q&A