Flower: a friendly federated learning framework
Flower: a friendly federated learning framework. - Javier Fernandez-Marques (Flower Labs)
FL moves the training to the data
- Cross-device FL
Massively distributed taking the training to smart devices.
- Cross-silo FL
Institutions collaborate training an ML model while preserving customers/patients data.
- Intra-organisation FL
A single organisation, one or multiple-locations, generating large volumes of data.
Flower community is incredibly active
- Fast adoption in open source projects.
- Involved in shaping the core framework, documentation, examples, tutorials.
- Rich community interactions.
- Used to build the latest research.
Flower Framework
A unified approach to Federated Learning, Analytics and Evaluation.
Python, Android, iOS, C++, cross-silo, cross-device, horizontal, vertical, differential privacy, secure aggregation, SMPC, simulation, deployment? Yes.
- Flower API is lightweight.
- Minimal, but complete, FL training with ~20 lines of extra code.
- Works out of the box, easy to adapt to your setting.
- Extend Flower’s strategies via custom callables.
- Develop your own strategies entirely.
Custom Strategies
A Flower Strategy is at the core of a FL pipeline: samples clients, sends models to clients, aggregates models, etc.
Implement your own client sampling mechanism.
For example: resampling clients in consecutive rounds. In what other settings we might want to customize how client sampling is performed?
- Biased sampling.
- Sample client based on CO2 emissions.
- Client-specific fit() instructions.
Control the periodicity at which certain sub-stages take place within the strategy.
For example: how often the global model is evaluated by the server. When is this particularly useful?
- Cross-device FL with lightweight models for clients but large validation set on the server.
- Event-based global evaluation.
Federated Fine-tuning of a ViT
- State of the art Vision architectures.
- Fast fine-tuning of a ImageNet-pretrained ViT on a single GPU.
On-device Federated Fine-tuning of Open AI’s Whisper
- Finetune Whisper with Flower for keyword spotting applications.
- Runs even on Raspberry Pi.
LLM-FlowerTune: Federated LLM Fine-tuning with Flower
- Finetune an LLM with Flower as a general purpose Chat assistant.
- Highly optimized: LoRA + PEFT
- Runs on a single GPU.
Flower Datasets
- Keep dataset downloading and partitioning simple.
- 9 built-in partitioning methods.
- Interface with 120K+ HF Datasets.
- Great for research and prototyping!
- Open source.
