Skip to content

Latest commit

 

History

History
125 lines (95 loc) · 4.58 KB

README.md

File metadata and controls

125 lines (95 loc) · 4.58 KB
Raw

XAIN Federated Learning

This is the main source code repository for xain-fl.

To get started, take a look at the documentation: https://xain-fl.readthedocs.io/en/latest

For developers, see CONTRIBUTING.md

Spike: XAIN FL Framework

At the time of this writing (xain-fl v0.8) the xain fl framework is composed of the:

  • platform: contains the coordinator and other services (written in rust)
  • sdk: python sdk that contains the participant code. This is also where we define the actual machine learning training tasks. The documentation and tutorial show how it works.

Federation learning session flow

  1. A coordinator is setup with the session parameters: rounds, min_clients, and participants_ratio.
  2. Participants connect to the coordinator.
  3. The coordinator waits until at least min_clients are connected to the coordinator.
  4. Once the coordinator has enough participants connected it starts a round.
  5. The coordinator starts by selecting a subset of the participants (depending of the participants_ratio setting)
  6. At the beginning of a round each participant requests the global model (the weights of the current model)
  7. With the model each participant executes the machine learning task on their own data
  8. After each participant finishes their machine learning task they upload their update model to the coordinator.
  9. After the coordinator receives the updated models from all the participants that are part of the round it runs the aggregation to compute the new global model.
  10. The coordinator starts a new round and repeats steps 5 to 9

Note: Although I refer only the coordinator and participants in this flow, the xain fl framework actually splits the coordinator into two services (coordinator and aggregator)

Running the platform

  1. Start the platform using docker-compose
$ docker-compose -f docker/docker-compose.yml up

This will start both the coordinator which is responsible for managing the participants and coordinating the federated learning session, and the aggregator which is responsible for aggregation the individual models of each participant into one global model.

It will also start a variety of others services useful for development, debugging and monitoring:

  • swagger ui: The REST API specification for the coordinator. In the top explore bar of swagger you can also type ./aggregator.yml to see the REST API specification for the aggregator.
  • grafana: The default username and password for grafana is admin. The most helpful dashboard is the Coordinator Metrics dashboard that shows information about the state of the federated learning session.

Changing the coordinator parameters

The coordinator service in the docker-compose file uses the configs/docker-dev-coordinator.toml configuration file. This file can be used to change the behavior of the coordinator.

The important settings are under the federated_learning section:

  • rounds: the number of rounds the session will run
  • participants_ratio: The ratio of connected participants that will be selected for each round.
  • min_clients: The minimum number of participants that need to participate in each round (the coordinator will wait until enough participants are connected before starting a round)

Running the python example

  1. Install example and dependencies (recommend doing this inside a virtual environment)
$ pip install python/sdk
$ pip install python/client_examples/keras_house_prices

  1. Download the dataset from Kaggle (account required)

  2. Prepare the data

$ cd python/client_examples/keras_house_prices
$ mkdir data
$ cd data
$ unzip house-prices-advanced-regression-techniques.zip
$ split-data --data-directory data --number-of-participants 10
$ cd ..
  1. Run the python example
$ ./run.sh 10

Resources