fl-network 1.0.2

Extensible network framework library for federated learning system

Extensible network framework for federated learning system

FL-Network is a simple Network architecture for federated learning. It is suitable for parameter server mode and P2P mode of federated learning.FL-Network uses Redis as multi-terminal interaction messaging middleware, which supports the construction and simulation of federated learning networks in standalone and distributed environments. Researchers can use FL-Network to quickly build an extensible federated learning system. Through The API of FL-Network, multi-terminal message transmission, such as weight(FedAvg), gradient(FedSGD) etc., can be performed without considering the underlying Network. In this way, researchers can focus more on the implementation of federated learning algorithm on FL-Network. At the same time, there is no need to install complex federated learning framework, which is beneficial for researchers to learn the specific implementation process of federated learning algorithm.

Installation for FL-Network

You can install Fl-Network with pip or install manually.

With pip

~$ pip install fl-network

Manually

Download the source distribution or built distribution in release.

~$ tar -zxvf ./fl-network-0.0.1.tar.gz
~$ cd fl-network-0.0.1
~$ python setup.py install

or use built distribution.

~$ pip install fl_network-0.0.1-py3-none-any.whl # wheel
~$ pip install fl_network-0.0.1-py3.8.egg # egg

Installation for Redis

Fl-network uses Redis as multi-terminal interaction messaging middleware. You can install redis on WSL2(Windows Subsystem for Linux), Linux and MacOS. If you prefer to install redis on Windows, please refer to here.

Quick start with demo

Before getting start to program your FL algorithm, you must launch redis server on isolated server or localhost and export ip and port of redis. (default 127.0.0.1:6379)

from FLNetwork.paramServer import ParamServer
from FLNetwork.device import Device

if __name__ == '__main__':

    # setup()
    # setup FL device: dev_01 locally (ip=127.0.0.1, port=6379) 
    dev_01 = Device(name="dev_01",
                    server="param_sys", peers=["dev_02"],
                    args_from_ser=["weight", "grad"], args_from_peer=[],
                    ip="127.0.0.1", port=6379)

    # setup FL device: dev_02 locally (ip=127.0.0.1, port=6379) 
    dev_02 = Device(name="dev_02",
                    server="param_sys", peers=["dev_01"],
                    args_from_ser=["weight", "grad"], args_from_peer=[],
                    ip="127.0.0.1", port=6379)

    # setup FL parameters server: param_sys locally (ip=127.0.0.1, port=6379) 
    server = ParamServer(name="param_sys", 
                         devices=["dev_01", "dev_02"], args_from_dev=["weight", "grad"],
                         ip="127.0.0.1", port=6379)

The setup parameters Illustration

Class	Parameter	Illustration
Device	name	The device name is used as the identifier
-	server	The server name is used as the identifier
-	peers	The list of other peers
-	args_from_ser	The predefined args that will be received from servere
-	args_from_peer	The predefined args that will be received from peer
-	ip	Ip of redis
-	port	Port of redis
ParamServer	name	The server name is used as the identifier
-	devices	The list of devices
-	args_from_dev	The predefined args that will be received from devices
-	ip	Ip of redis
-	port	Port of redis

API

Parameters Server

---

Define a new arg from device:

server.set_from_arg(arg, device) # arg: arg name (str), device: target device name (str)

Send a arg to device by predefined arg channel above:

server.send(arg, value) # arg: arg name (str), value: value of arg (Any)

Read arg from device:

server.read(dev, arg) # dev: device name (str), arg: name of arg to read (str)

Get arg channel (to be received):

channel = server.get_channel() # return channel dict

Set global arg (anyone can read):

server.set(global_arg, value) # global_arg: global arg name (str), value: value of arg (Any)

Get global arg:

arg = server.get(global_arg) # global_arg: global arg name (str)

Device

---

Define a new arg from peer or server:

dev_01.set_from_arg(arg, dev_ser, is_server=False) # arg: arg name (str), dev_ser: target device/server name (str)

Send a arg to device/server by predefined arg channel above:

dev_01.send(arg, value) # arg: arg name (str), value: value of arg (Any)

Read arg from device:

dev_01.read_dev(dev, arg) # dev: device name (str), arg: name of arg to read (str)

Read arg from server:

dev_01.read_ser(ser, arg) # dev: server name (str), arg: name of arg to read (str)

Get arg channel (to be received):

channel = dev_01.get_channel() # return channel dict

Set global arg (anyone can read):

dev_01.set(global_arg, value) # global_arg: global arg name (str), value: value of arg (Any)

Get global arg:

arg = dev_01.get(global_arg) # global_arg: global arg name (str)

Working Example

# 1. dev_01 define a new arg from server
dev_01.set_from_arg("init_weight", server.name)
# 2. parameters server initialize model
import torch
weight = torch.zeros((1000, 1000))
server.send("init_weight", weight)
# 3. device receive model weight
init_weight = dev_01.rend_ser(server.name, "init_weight")
# 4. set init_weight as net.parameters() in Pytorch

Contact

Project Investigator: Rostar ([e-mail])