NetFL is a framework for executing Federated Learning (FL) experiments in simulated IoT and Fog/Edge computing environments. It enables the modeling of heterogeneous and resource-constrained scenarios, incorporating factors such as computational disparities among devices, limited bandwidth, latency, packet loss, and diverse network topologies.
Using its native abstractions for tasks, devices, and networks, NetFL allows researchers to configure and execute FL experiments in a declarative and reproducible manner, providing realistic evaluations of algorithms under non-ideal, real-world conditions.
Under the hood, NetFL leverages Fogbed for distributed network emulation and Flower for federated learning orchestration. These libraries provide robust foundations for virtualization and FL training, and NetFL integrates and extends them into a unified framework designed specifically for FL research in IoT and Fog/Edge Computing.
Requirements: Ubuntu 22.04 LTS or later, Python 3.9 or higher.
Refer to the Containernet documentation for further details.
Install Ansible:
sudo apt-get install ansible
Clone the Containernet repository:
git clone https://github.com/containernet/containernet.git
Run the installation playbook:
sudo ansible-playbook -i "localhost," -c local containernet/ansible/install.yml
Create and activate a virtual environment:
python3 -m venv venv
source venv/bin/activate
Note: The virtual environment must be activated before installing or using any Python packages, including Containernet and NetFL.
Install Containernet into the active virtual environment:
pip install containernet/.
While the virtual environment is still active, run:
pip install netfl
NetFL experiments are designed to be modular and declarative, making it easy to set up federated learning scenarios. The steps below describe how to set up and run an experiment using NetFL. The example uses the MNIST dataset. You can find more examples in the examples folder:
The Task class encapsulates the dataset, model, partitioning strategy, and training configuration. You can inherit from it and override methods to specify:
- Dataset information: Source, input/label keys, data types
- Partitioning: How data is split among clients (IID, non-IID, etc.)
- Preprocessing: Any transformations applied to the data
- Model: The model architecture and optimizer
- Aggregation strategy: Federated averaging or custom strategies
- Training configs: Batch size, epochs, number of clients, etc.
After implementing the task, export an
FLTaskclass that extends it for use by NetFL.
from typing import Any
import tensorflow as tf
from keras import models, optimizers
from flwr.server.strategy import Strategy, FedAvg
from netfl.core.task import Task, Dataset, DatasetInfo, DatasetPartitioner, TrainConfigs
from netfl.core.models import cnn3
from netfl.core.partitioners import IidPartitioner
class MNIST(Task):
def dataset_info(self) -> DatasetInfo:
return DatasetInfo(
huggingface_path="ylecun/mnist",
input_key="image",
label_key="label",
input_dtype=tf.float32,
label_dtype=tf.int32,
)
def dataset_partitioner(self) -> DatasetPartitioner:
return IidPartitioner()
def preprocess_dataset(self, dataset: Dataset, training: bool) -> Dataset:
return Dataset(x=tf.divide(dataset.x, 255.0), y=dataset.y)
def model(self) -> models.Model:
return cnn3(
input_shape=(28, 28, 1),
output_classes=10,
optimizer=optimizers.SGD(learning_rate=0.01),
)
def aggregation_strategy(self) -> tuple[type[Strategy], dict[str, Any]]:
return FedAvg, {}
def train_configs(self) -> TrainConfigs:
return TrainConfigs(
batch_size=16,
epochs=2,
num_clients=4,
num_partitions=4,
num_rounds=10,
seed_data=42,
shuffle_data=True,
)
class FLTask(MNIST):
passNetFL uses resource classes to model the infrastructure. You can create heterogeneous environments by varying these parameters, simulating real-world IoT and edge scenarios:
NetworkResource: Defines network links between clusters/devicesDeviceResource: Represents a server or client device (CPU, memory, bandwidth)ClusterResource: Groups devices into clusters (cloud, edge, etc.)
Use FLExperiment to assemble the experiment:
- Create the network, device, and cluster resources
- Instantiate the experiment with a name, task, and cluster resources
- Create server and client devices
- Assign devices to clusters
- Register remote workers (for distributed execution)
- Link clusters with network resources to define topology
When the worker host
cpu_clockis set toBASE_COMPUTE_UNIT, all resourcecpu_clockvalues are interpreted in Docker CPU units (e.g., millicores) instead of GHz.
from netfl.core.experiment import FLExperiment
from netfl.utils.resources import (
WorkerHostResource,
NetworkResource,
DeviceResource,
ClusterResource,
ClusterResourceType,
BASE_COMPUTE_UNIT,
)
from task import FLTask
task = FLTask()
num_clients = task.train_configs().num_clients
worker_host_resource = WorkerHostResource(cpu_clock=BASE_COMPUTE_UNIT)
server_resource = DeviceResource(
name="server",
cpu_cores=1,
cpu_clock=1.0,
memory=1024,
network_resource=NetworkResource(bw=1000),
worker_host_resource=worker_host_resource,
)
client_a_resource = DeviceResource(
name="client_a",
cpu_cores=1,
cpu_clock=0.5,
memory=512,
network_resource=NetworkResource(bw=100),
worker_host_resource=worker_host_resource,
)
client_b_resource = DeviceResource(
name="client_b",
cpu_cores=1,
cpu_clock=0.25,
memory=512,
network_resource=NetworkResource(bw=50),
worker_host_resource=worker_host_resource,
)
cloud_resource = ClusterResource(
name="cloud",
type=ClusterResourceType.CLOUD,
device_resources=[server_resource],
)
edge_0_resource = ClusterResource(
name="edge_0",
type=ClusterResourceType.EDGE,
device_resources=(num_clients // 2) * [client_a_resource],
)
edge_1_resource = ClusterResource(
name="edge_1",
type=ClusterResourceType.EDGE,
device_resources=(num_clients // 2) * [client_b_resource],
)
exp = FLExperiment(
name="mnist-exp",
task=task,
cluster_resources=[cloud_resource, edge_0_resource, edge_1_resource],
)
server = exp.create_server(server_resource)
edge_0_clients = exp.create_clients(client_a_resource, edge_0_resource.num_devices)
edge_1_clients = exp.create_clients(client_b_resource, edge_1_resource.num_devices)
cloud = exp.create_cluster(cloud_resource)
edge_0 = exp.create_cluster(edge_0_resource)
edge_1 = exp.create_cluster(edge_1_resource)
exp.add_to_cluster(server, cloud)
for client in edge_0_clients:
exp.add_to_cluster(client, edge_0)
for client in edge_1_clients:
exp.add_to_cluster(client, edge_1)
worker = exp.register_remote_worker(ip="127.0.0.1", port=5000)
worker.add_cluster(cloud)
worker.add_cluster(edge_0)
worker.add_cluster(edge_1)
worker.create_cluster_link(cloud, edge_0, NetworkResource(bw=10))
worker.create_cluster_link(cloud, edge_1, NetworkResource(bw=20))
try:
exp.start()
except Exception as ex:
print(ex)
finally:
exp.stop()Start the required worker(s), and then run your experiment script. Refer to the Fogbed documentation for detailed instructions on starting workers.
For example:
RunWorker -p=5000
python3 experiment.py
git clone https://github.com/larsid/netfl.git
In the project root directory, create or modify a NetFL Task and name the file task.py. Refer to the examples in the examples folder for guidance on task creation.
Use Docker Compose to set up the infrastructure, including the server and clients:
docker compose up -d
To check the server logs, run:
docker logs server
Training logs are also stored in the logs/ folder within the project root directory.
To stop and remove all running containers, use the following command:
docker compose down
NetFL is licensed under the Apache License 2.0. See LICENSE for details.