easyFL: A Lightning Framework for Federated Learning

This repository is PyTorch implementation for paper Federated Learning with Fair Averaging which is accepted by IJCAI-21 Conference.

Our easyFL is a strong and reusable experimental platform for research on federated learning (FL) algorithm. It is easy for FL-researchers to quickly realize and compare popular centralized federated learning algorithms.

Requirements
QuickStart
Architecture
Remark
Citation
Contacts
References

Requirements

The model is implemented using Python3 with dependencies below:

numpy>=1.17.2
pytorch>=1.3.1
torchvision>=0.4.2
cvxopt>=1.2.0
scipy>=1.3.1
matplotlib>=3.1.1
prettytable>=2.1.0
ujson>=4.0.2

QuickStart

First, run the command below to get the splited dataset MNIST:

# generate the splited dataset
python generate_fedtask.py

Second, run the command below to quickly get a result of the basic algorithm FedAvg on MNIST with a simple CNN:

python main.py --task mnist_client100_dist0_beta0_noise0 --model cnn --method fedavg --num_rounds 20 --num_epochs 5 --learning_rate 0.215 --proportion 0.1 --batch_size 10 --train_rate 1 --eval_interval 1

The result will be stored in ./fedtask/mnist_client100_dist0_beta0_noise0/record/.

Third, run the command below to get a visualization of the result.

# change to the ./utils folder
cd ../utils
# visualize the results
python result_analysis.py

Performance

The rounds necessary for FedAVG to achieve 99% test accuracy on MNIST using CNN with E=5 (reported in [McMahan. et al. 2017] / ours)
Proportion	iid		non-iid
Proportion	B=FULL	B=10	B=FULL	B=10
0.0	387 / 325	50 / 91	1181 / 1021	956 / 771
0.1	339 / 203	18 / 18	1100 / 453	206 / 107
0.2	337 / 207	18 / 19	978 / 525	200 / 95
0.5	164 / 214	18 / 18	1067 / 606	261 / 105
1.0	246 / 267	16 / 18	-- / 737	97 / 90

Accelarating FL Process by Increasing Parallelism For FedAVG on MNIST using CNN (20/100 clients per round)
Num_threads	1	2	5	10	15	20
Mean of time cost per round(s/r)	19.5434	13.5733	9.9935	9.3092	9.2885	8.3867

Reproduced FL Algorithms

Method	Reference	Publication
FedAvg	[McMahan et al., 2017]	AISTATS' 2017
FedProx	[Li et al., 2020]	MLSys' 2020
FedFV	[Wang et al., 2021]	IJCAI' 2021
qFFL	[Li et al., 2019]	ICLR' 2020
AFL	[Mohri et al., 2019]	ICML' 2019
FedMGDA+	[Hu et al., 2020]	pre-print
FedFA	[Huang et al., 2020]	pre-print
SCAFFOLD	[Karimireddy et al., 2020]	ICML' 2020
FedDyn	[Acar et al., 2021]	ICLR' 2021
...

Options

Basic options:

task is to choose the task of splited dataset. Options: name of fedtask (e.g. mnist_client100_dist0_beta0_noise0).
method is to choose the FL algorithm. Options: fedfv, fedavg, fedprox, …
model should be the corresponding model of the dataset. Options: mlp, cnn, resnet18.

Server-side options:

sample decides the way to sample clients in each round. Options: uniform means uniformly, md means choosing with probability.
aggregate decides the way to aggregate clients' model. Options: uniform, weighted_scale, weighted_com
num_rounds is the number of communication rounds.
proportion is the proportion of clients to be selected in each round.
lr_scheduler is the global learning rate scheduler.
learning_rate_decay is the decay rate of the global learning rate.

Client-side options:

num_epochs is the number of local training epochs.
learning_rate is the step size when locally training.
batch_size is the size of one batch data during local training.
optimizer is to choose the optimizer. Options: SGD, Adam.
momentum is the ratio of the momentum item when the optimizer SGD taking each step.

Other options:

seed is the initial random seed.
gpu is the id of the GPU device, -1 for CPU.
eval_interval controls the interval between every two evaluations.
train_rate is the proportion of the train set and validation set in local dataset.
drop controls the dropout of clients after being selected in each communication round according to distribution Beta(drop,1). The larger this term is, the more possible for clients to drop.
num_threads is the number of threads in the clients computing session that aims to accelarate the training process.

Additional hyper-parameters for particular federated algorithms:

mu is the parameter for FedProx.
alpha is the parameter for FedFV.
tau is the parameter for FedFV.
...

Each additional parameter can be defined in ./utils/fflow.read_option

Architecture

We seperate the FL system into four parts: benchmark, fedtask, method and utils.

├─ benchmark
│  ├─ mnist							//mnist dataset
│  │  ├─ data							//data
│  |  └─ model							//the corresponding model
│  ├─ ...
│  └─ generator.py						//fedtask generator implementation
├─ fedtask
│  ├─ mnist_client100_dist0_beta0_noise0//IID(beta=0) MNIST for 100 clients with not predefined noise
│  │  ├─ record							//record of result
│  |  └─ task.json						//the splitted federated dataset (fedtask)
|  └─ ...
├─ method
│  ├─ fedavg.py							//FL algorithm implementation inherit fedbase.py
│  ├─ fedbase.py						//FL algorithm superclass(i.e.,fedavg)
│  ├─ fedfv.py							//our FL algorithm
│  ├─ fedprox.py
|  └─ ...
├─ utils
│  ├─ fflow.py							//option to read, initialize,...
│  ├─ fmodule.py						//model-level operators
│  └─ result_analysis.py				        //to generate the visualization of record
├─ generate_fedtask.py					        //generate fedtask
├─ requirements.txt
└─ main.py

Benchmark

This module is to generate fedtask by partitioning the particular distribution data through generate_fedtask.py. To generate different fedtask, there are three parameters: dist, num_clients , beta. dist denotes the distribution type (e.g. 0 denotes iid and balanced distribution, 1 denotes niid-label-quantity and balanced distribution). num_clients is the number of clients participate in FL system, and beta controls the degree of non-iid for different dist. Each dataset can correspond to differrent models (mlp, cnn, resnet18, …). We refer to [McMahan et al., 2017], [Li et al., 2020], [Li et al., 2021], [Li et al., 2019], [Caldas et al., 2018], [He et al., 2020] when realizing this module. Further details are described in benchmark/README.md.

Fedtask

We define each task as a combination of the federated dataset of a particular distribution and the experimental results on it. The raw dataset is processed into .json file, following LEAF (https://github.com/TalwalkarLab/leaf). The architecture of the .json file is described as below:

"""
{
    'meta':{
        'benchmark': 'mnist',						//dataset name
        'num_clients': 100,							//the number of clients
        'dist': 0,									//the distribution of data
        'beta': 0,									//the parameter of distribution
    },
    'clients': {
        'user0': {
            'dtrain': {'x': [...], 'y': [...]},
            'dvalid': {'x': [...], 'y': [...]},
            'dvol': 600,
        },...,
        'user99': {
            'dtrain': {'x': [...], 'y': [...]},
            'dvalid': {'x': [...], 'y': [...]},
            'dvol': 600,
        },
    },
    'dtest': {'x':[...], 'y':[...]}
}
"""

Run the file ./generate_fedtask.py to get the splited dataset (.json file).

Since the task-specified models are usually orthogonal to the FL algorithms, we don't consider it an important part in this system. And the model and the basic loss function are defined in ./task/dataset_name/model_name.py. Further details are described in fedtask/README.md.

Method

This module is the specific federated learning algorithm implementation. Each method contains two classes: the Server and the Client.

Server

The whole FL system starts with the main.py, which runs server.run() after initialization. Then the server repeat the method iterate() for num_rounds times, which simulates the communication process in FL. In the iterate(), the BaseServer start with sampling clients by select(), and then exchanges model parameters with them by communicate(), and finally aggregate the different models into a new one with aggregate(). Therefore, anyone who wants to customize its own method that specifies some operations on the server-side should rewrite the method iterate() and particular methods mentioned above.

Client

The clients reponse to the server after the server communicate_with() them, who first unpack() the received package and then train the model with their local dataset by train(). After training the model, the clients pack() send package (e.g. parameters, loss, gradient,... ) to the server through reply().

Further details of this module are described in method/README.md.

Utils

Utils is composed of commonly used operations: model-level operation (we convert model layers and parameters to dictionary type and apply it in the whole FL system), the flow controlling of the framework in and the supporting visualization templates to the result. To visualize the results, please run ./utils/result_analysis.py. Further details are described in utils/README.md.

Remark

Since we've made great changes on the latest version, to fully reproduce the reported results in our paper Federated Learning with Fair Averaging, please use another branch easyFL v1.0 of this project.

Citation

Please cite our paper in your publications if this code helps your research.

@article{wang2021federated,
  title={Federated Learning with Fair Averaging},
  author={Wang, Zheng and Fan, Xiaoliang and Qi, Jianzhong and Wen, Chenglu and Wang, Cheng and Yu, Rongshan},
  journal={arXiv preprint arXiv:2104.14937},
  year={2021}
}te