Home

Awesome

PyGrid logo

Tests

PyGrid is a peer-to-peer network of data owners and data scientists who can collectively train AI models using PySyft. PyGrid is also the central server for conducting both model-centric and data-centric federated learning.

You may control PyGrid via our user-interface, PyGrid Admin.

Architecture

PyGrid platform is composed by three different components.

Use Cases

Federated Learning

Simply put, federated learning is machine learning where the data and the model are initially located in two different locations. The model must travel to the data in order for training to take place in a privacy-preserving manner. Depending on what you're looking to accomplish, there are two types of federated learning that you can perform with the help of PyGrid.

Model-centric FL

Model-centric FL is when the model is hosted in PyGrid. This is really useful when you have data located at an "edge device" like a person's mobile phone or web browser. Since the data is private, we should respect that and leave it on the device. The following workflow will take place:

  1. The device will request to train a model
  2. The model and a training plan may be sent to that device
  3. The training will take place with private data on the device itself
  4. Once training is completed, a "diff" is generated between the new and the original state of the model
  5. The diff is reported back to PyGrid and it's averaged into the model

This takes place potentially with hundreds, or thousands of devices simultaneously. For model-centric federated learning, you only need to run a Domain. Networks and Workers are irrelevant for this specific use-case.

Note: For posterity sake, we previously used to refer to this process as "static federated learning".

Cycled MCFL

Data-centric FL

Data-centric FL is the same problem as model-centric FL, but from the opposite perspective. The most likely scenario for data-centric FL is where a person or organization has data they want to protect in PyGrid (instead of hosting the model, they host data). This would allow a data scientist who is not the data owner, to make requests for training or inference against that data. The following workflow will take place:

  1. A data scientist searches for data they would like to train on (they can search either an individual Domain, or a Network of Domains)
  2. Once the data has been found, they may write a training plan and optionally pre-train a model
  3. The training plan and model are sent to the PyGrid Domain in the form of a job request
  4. The PyGrid Domain will gather the appropriate data from its database and send the data, the model, and the training plan to a Worker for processing
  5. The Worker performs the plan on the model using the data
  6. The result is returned to the Domain
  7. The result is returned to the data scientist

For the last step, we're working on adding the capability for privacy budget tracking to be applied that will allow a data owner to "sign off" on whether or not a trained model should be released.

Note: For posterity sake, we previously used to refer to this process as "dynamic federated learning".

Domain-only data-centric FL

Technically speaking, it isn't required to run a Network when performing data-centric federated learning. Alternatively, as a data owner, you may opt to only run a Domain, but participate in a Network hosted by someone else. The Network host will not have access to your data.

Domain-only DCFL

Network-based data-centric FL

Many times you will wat to use a Network to allow multiple Domains to be connected together. As a data owner, it's not strictly necessary to own and operate mulitple Domains. PyGrid doesn't prescribe one way to organize Domains and Networks, but we expose these applications to allow you and various related stakeholders to make the correct decision about your infrastructure needs.

Network-based DCFL

Local Setup

Currently, we suggest two ways to run PyGrid locally: Docker and manually running from source. With Docker, we can organize all the services we'd like to use and then boot them all in one command. With manually running from source, we have to run them as separate tasks.

Docker

To install Docker, just follow the docker documentation.

1. Setting the your hostfile

Before start the grid platform locally using Docker, we need to set up the domain names used by the bridge network. In order to use these Domains from outside of the containers context, you should add the following domain names on your /etc/hosts

127.0.0.1 network
127.0.0.1 alice
127.0.0.1 bob
127.0.0.1 charlie
127.0.0.1 dan

Note that you're not restricted to running 4 nodes and a network. You could instead run just a single node if you'd like - this is often all you need for model-centric federated learning. For the sake of our example, we'll use the network running 4 nodes underneath but you're welcome to modify it to your needs.

2. Run Docker Images

The latest PyGrid Network and Node images are also available on the Docker Hub.

3. Optional - Build your own images

If you want to build your own custom images, you may do so using the following command for the Node:

docker build ./apps/domain --file ./apps/domain/Dockerfile --tag openmined/grid-domain:mybuildname

Or for the Worker:

docker build ./apps/worker --file ./apps/worker/Dockerfile --tag openmined/grid-worker:mybuildname

Or for the Network:

docker build ./apps/network --file ./apps/network/Dockerfile --tag openmined/grid-network:mybuildname

Manual Start

Running a Node

Installation

First install poetry and run poetry install in apps/node

To start the PyGrid Node manually, run:

cd apps/node
./run.sh --id bob --port 5000 --start_local_db

You can pass the arguments or use environment variables to set the network configs.

Arguments

Environment Variables

Running a Network

To start the PyGrid Network manually, run:

cd apps/network
./run.sh --port 7000 --start_local_db

You can pass the arguments or use environment variables to set the network configs.

Arguments

Environment Variables

Deployment & CLI

Please check the instruction for deployment and CLI here.

Contributing

If you're interested in contributing, check out our Contributor Guidelines.

Support

For support in using this library, please join the #support Slack channel. Click here to join our Slack community!

License

Apache License 2.0