Figure 1: An Example for FedAvg. You can create a scenario using
generate_DATA.py
and run an algorithm using main.py
, clientNAME.py
, and serverNAME.py
.
We expose this user-friendly algorithm library (with an integrated evaluation platform) for beginners who intend to start federated learning (FL) study.
-
34 traditional FL (tFL) or personalized FL (pFL) algorithms, 3 scenarios, and 20 datasets.
-
Some experimental results are avalible here.
-
Refer to this guide to learn how to use it.
-
This library can simulate scenarios using the 4-layer CNN on Cifar100 for 500 clients on one NVIDIA GeForce RTX 3090 GPU card with only 5.08GB GPU memory cost.
-
PFLlib primarily focuses on data (statistical) heterogeneity. For algorithms and an evaluation platform that address both data and model heterogeneity, please refer to our extended project Heterogeneous Federated Learning (HtFL).
-
As we strive to meet diverse user demands, frequent updates to the project may alter default settings and scenario creation codes, affecting experimental results.
-
Closed issues may help you a lot.
-
When submitting pull requests, please provide sufficient instructions and examples in the comment box.
The origin of the statistical heterogeneity phenomenon is the personalization of users, who generate non-IID (not Independent and Identically Distributed) and unbalanced data. With statistical heterogeneity existing in the FL scenario, a myriad of approaches have been proposed to crack this hard nut. In contrast, the personalized FL (pFL) may take advantage of the statistically heterogeneous data to learn the personalized model for each user.
Thanks to @Stonesjtu, this library can also record the GPU memory usage for the model. By using the package opacus, we introduce DP (differential privacy) into this library (please refer to ./system/flcore/clients/clientavg.py
for example). Following FedCG, we also introduce the DLG (Deep Leakage from Gradients) attack and PSNR (Peak Signal-to-Noise Ratio) metric to evaluate the privacy-preserving ability of tFL/pFL algorithms (please refer to ./system/flcore/servers/serveravg.py
for example). Now we can train on some clients and evaluate performance on other new clients by setting args.num_new_clients
in ./system/main.py
. Note that not all the tFL/pFL algorithms support this feature.
Citation
@article{zhang2023pfllib,
title={PFLlib: Personalized Federated Learning Algorithm Library},
author={Zhang, Jianqing and Liu, Yang and Hua, Yang and Wang, Hao and Song, Tao and Xue, Zhengui and Ma, Ruhui and Cao, Jian},
journal={arXiv preprint arXiv:2312.04992},
year={2023}
}
-
FedAvg — Communication-Efficient Learning of Deep Networks from Decentralized Data AISTATS 2017
Update-correction-based tFL
-
SCAFFOLD - SCAFFOLD: Stochastic Controlled Averaging for Federated Learning ICML 2020
Regularization-based tFL
-
FedProx — Federated Optimization in Heterogeneous Networks MLsys 2020
-
FedDyn — Federated Learning Based on Dynamic Regularization ICLR 2021
Model-splitting-based tFL
-
MOON — Model-Contrastive Federated Learning CVPR 2021
Knowledge-distillation-based tFL
-
FedGen — Data-Free Knowledge Distillation for Heterogeneous Federated Learning ICML 2021
-
FedNTD — Preservation of the Global Knowledge by Not-True Distillation in Federated Learning NeurIPS 2022
-
FedMTL (not MOCHA) — Federated multi-task learning NeurIPS 2017
-
FedBN — FedBN: Federated Learning on non-IID Features via Local Batch Normalization ICLR 2021
Meta-learning-based pFL
-
Per-FedAvg — Personalized Federated Learning with Theoretical Guarantees: A Model-Agnostic Meta-Learning Approach NeurIPS 2020
Regularization-based pFL
-
pFedMe — Personalized Federated Learning with Moreau Envelopes NeurIPS 2020
-
Ditto — Ditto: Fair and robust federated learning through personalization ICML 2021
Personalized-aggregation-based pFL
-
APFL — Adaptive Personalized Federated Learning 2020
-
FedFomo — Personalized Federated Learning with First Order Model Optimization ICLR 2021
-
FedAMP — Personalized Cross-Silo Federated Learning on non-IID Data AAAI 2021
-
FedPHP — FedPHP: Federated Personalization with Inherited Private Models ECML PKDD 2021
-
APPLE — Adapt to Adaptation: Learning Personalization for Cross-Silo Federated Learning IJCAI 2022
-
FedALA — FedALA: Adaptive Local Aggregation for Personalized Federated Learning AAAI 2023
Model-splitting-based pFL
-
FedPer — Federated Learning with Personalization Layers 2019
-
LG-FedAvg — Think Locally, Act Globally: Federated Learning with Local and Global Representations 2020
-
FedRep — Exploiting Shared Representations for Personalized Federated Learning ICML 2021
-
FedRoD — On Bridging Generic and Personalized Federated Learning for Image Classification ICLR 2022
-
FedBABU — Fedbabu: Towards enhanced representation for federated image classification ICLR 2022
-
FedGC — Federated Learning for Face Recognition with Gradient Correction AAAI 2022
-
FedCP — FedCP: Separating Feature Information for Personalized Federated Learning via Conditional Policy KDD 2023
-
GPFL — GPFL: Simultaneously Learning Generic and Personalized Feature Information for Personalized Federated Learning ICCV 2023
-
FedGH — FedGH: Heterogeneous Federated Learning with Generalized Global Header ACM MM 2023
-
DBE — Eliminating Domain Bias for Federated Learning in Representation Space NeurIPS 2023
Knowledge-distillation-based pFL
-
FedDistill (FD) — Communication-Efficient On-Device Machine Learning: Federated Distillation and Augmentation under Non-IID Private Data 2018
-
FML — Federated Mutual Learning 2020
-
FedKD — Communication-efficient federated learning via knowledge distillation Nature Communications 2022
-
FedProto — FedProto: Federated Prototype Learning across Heterogeneous Clients AAAI 2022
-
FedPCL (w/o pre-trained models) — Federated learning from pre-trained models: A contrastive learning approach NeurIPS 2022
-
FedPAC — Personalized Federated Learning with Feature Alignment and Classifier Collaboration ICLR 2023
For the label skew scenario, we introduce 14 famous datasets: MNIST, EMNIST, Fashion-MNIST, Cifar10, Cifar100, AG News, Sogou News, Tiny-ImageNet, Country211, Flowers102, GTSRB, Shakespeare, and Stanford Cars, they can be easy split into IID and non-IID version. Since some codes for generating datasets such as splitting are the same for all datasets, we move these codes into ./dataset/utils/dataset_utils.py
. In the non-IID scenario, 2 situations exist. The first one is the pathological non-IID scenario, the second one is the practical non-IID scenario. In the pathological non-IID scenario, for example, the data on each client only contains the specific number of labels (maybe only 2 labels), though the data on all clients contains 10 labels such as the MNIST dataset. In the practical non-IID scenario, Dirichlet distribution is utilized (please refer to this paper for details). We can input balance
for the iid scenario, where the data are uniformly distributed.
For the feature shift scenario, we use 3 datasets that are widely used in Domain Adaptation: Amazon Review (fetch raw data from this site), Digit5 (fetch raw data from this site), and DomainNet.
For the real-world (or IoT) scenario, we also introduce 3 naturally separated datasets: Omniglot (20 clients, 50 labels), HAR (Human Activity Recognition) (30 clients, 6 labels), PAMAP2 (9 clients, 12 labels). For the details of datasets and FL algorithms in IoT, please refer to my FL-IoT repo.
If you need another data set, just write another code to download it and then use the utils.
- MNIST
cd ./dataset # python generate_MNIST.py iid - - # for iid and unbalanced scenario # python generate_MNIST.py iid balance - # for iid and balanced scenario # python generate_MNIST.py noniid - pat # for pathological noniid and unbalanced scenario python generate_MNIST.py noniid - dir # for practical noniid and unbalanced scenario # python generate_MNIST.py noniid - exdir # for Extended Dirichlet strategy
The output of python generate_MNIST.py noniid - dir
Number of classes: 10
Client 0 Size of data: 2630 Labels: [0 1 4 5 7 8 9]
Samples of labels: [(0, 140), (1, 890), (4, 1), (5, 319), (7, 29), (8, 1067), (9, 184)]
--------------------------------------------------
Client 1 Size of data: 499 Labels: [0 2 5 6 8 9]
Samples of labels: [(0, 5), (2, 27), (5, 19), (6, 335), (8, 6), (9, 107)]
--------------------------------------------------
Client 2 Size of data: 1630 Labels: [0 3 6 9]
Samples of labels: [(0, 3), (3, 143), (6, 1461), (9, 23)]
--------------------------------------------------
Show more
Client 3 Size of data: 2541 Labels: [0 4 7 8]
Samples of labels: [(0, 155), (4, 1), (7, 2381), (8, 4)]
--------------------------------------------------
Client 4 Size of data: 1917 Labels: [0 1 3 5 6 8 9]
Samples of labels: [(0, 71), (1, 13), (3, 207), (5, 1129), (6, 6), (8, 40), (9, 451)]
--------------------------------------------------
Client 5 Size of data: 6189 Labels: [1 3 4 8 9]
Samples of labels: [(1, 38), (3, 1), (4, 39), (8, 25), (9, 6086)]
--------------------------------------------------
Client 6 Size of data: 1256 Labels: [1 2 3 6 8 9]
Samples of labels: [(1, 873), (2, 176), (3, 46), (6, 42), (8, 13), (9, 106)]
--------------------------------------------------
Client 7 Size of data: 1269 Labels: [1 2 3 5 7 8]
Samples of labels: [(1, 21), (2, 5), (3, 11), (5, 787), (7, 4), (8, 441)]
--------------------------------------------------
Client 8 Size of data: 3600 Labels: [0 1]
Samples of labels: [(0, 1), (1, 3599)]
--------------------------------------------------
Client 9 Size of data: 4006 Labels: [0 1 2 4 6]
Samples of labels: [(0, 633), (1, 1997), (2, 89), (4, 519), (6, 768)]
--------------------------------------------------
Client 10 Size of data: 3116 Labels: [0 1 2 3 4 5]
Samples of labels: [(0, 920), (1, 2), (2, 1450), (3, 513), (4, 134), (5, 97)]
--------------------------------------------------
Client 11 Size of data: 3772 Labels: [2 3 5]
Samples of labels: [(2, 159), (3, 3055), (5, 558)]
--------------------------------------------------
Client 12 Size of data: 3613 Labels: [0 1 2 5]
Samples of labels: [(0, 8), (1, 180), (2, 3277), (5, 148)]
--------------------------------------------------
Client 13 Size of data: 2134 Labels: [1 2 4 5 7]
Samples of labels: [(1, 237), (2, 343), (4, 6), (5, 453), (7, 1095)]
--------------------------------------------------
Client 14 Size of data: 5730 Labels: [5 7]
Samples of labels: [(5, 2719), (7, 3011)]
--------------------------------------------------
Client 15 Size of data: 5448 Labels: [0 3 5 6 7 8]
Samples of labels: [(0, 31), (3, 1785), (5, 16), (6, 4), (7, 756), (8, 2856)]
--------------------------------------------------
Client 16 Size of data: 3628 Labels: [0]
Samples of labels: [(0, 3628)]
--------------------------------------------------
Client 17 Size of data: 5653 Labels: [1 2 3 4 5 7 8]
Samples of labels: [(1, 26), (2, 1463), (3, 1379), (4, 335), (5, 60), (7, 17), (8, 2373)]
--------------------------------------------------
Client 18 Size of data: 5266 Labels: [0 5 6]
Samples of labels: [(0, 998), (5, 8), (6, 4260)]
--------------------------------------------------
Client 19 Size of data: 6103 Labels: [0 1 2 3 4 9]
Samples of labels: [(0, 310), (1, 1), (2, 1), (3, 1), (4, 5789), (9, 1)]
--------------------------------------------------
Total number of samples: 70000
The number of train samples: [1972, 374, 1222, 1905, 1437, 4641, 942, 951, 2700, 3004, 2337, 2829, 2709, 1600, 4297, 4086, 2721, 4239, 3949, 4577]
The number of test samples: [658, 125, 408, 636, 480, 1548, 314, 318, 900, 1002, 779, 943, 904, 534, 1433, 1362, 907, 1414, 1317, 1526]
Saving to disk.
Finish generating dataset.
-
for MNIST and Fashion-MNIST
- Mclr_Logistic(1*28*28)
- LeNet()
- DNN(1*28*28, 100) # non-convex
-
for Cifar10, Cifar100 and Tiny-ImageNet
- Mclr_Logistic(3*32*32)
- FedAvgCNN()
- DNN(3*32*32, 100) # non-convex
- ResNet18, AlexNet, MobileNet, GoogleNet, etc.
-
for AG_News and Sogou_News
- LSTM()
- fastText() in Bag of Tricks for Efficient Text Classification
- TextCNN() in Convolutional Neural Networks for Sentence Classification
- TransformerModel() in Attention is all you need
-
for AmazonReview
-
for Omniglot
- FedAvgCNN()
-
for HAR and PAMAP
Install CUDA.
Install conda and activate conda.
conda env create -f env_cuda_latest.yaml # You may need to downgrade the torch using pip to match the CUDA version
-
Create proper environments (see Environments).
-
Download this project to an appropriate location using git.
git clone https://github.com/TsingZ0/PFLlib.git
-
Build evaluation scenarios (see Datasets and scenarios (updating)).
-
Run evaluation:
cd ./system python main.py -data MNIST -m cnn -algo FedAvg -gr 2000 -did 0 # using the MNIST dataset, the FedAvg algorithm, and the 4-layer CNN model
Note: It is preferable to tune algorithm-specific hyper-parameters before using any algorithm on a new machine.
If you need to simulate FL under practical situations, which includes client dropout, slow trainers, slow senders, and network TTL, you can set the following parameters to realize it.
-cdr
: The dropout rate for total clients. The selected clients will randomly drop at each training round.-tsr
and-ssr
: The rates for slow trainers and slow senders among all clients. Once a client is selected as a "slow trainer"/"slow sender", for example, it will always train/send slower than the original one.-tth
: The threshold for network TTL (ms).
It is easy to add new algorithms and datasets to this library.
-
To add a new dataset into this library, all you need to do is write the download code and use the utils which is similar to
./dataset/generate_MNIST.py
(you can also consider it as the template). -
To add a new algorithm, you can utilize the class Server and class Client, which are wrote in
./system/flcore/servers/serverbase.py
and./system/flcore/clients/clientbase.py
, respectively. -
To add a new model, just add it into
./system/flcore/trainmodel/models.py
. -
If you have a new optimizer while training, please add it into
./system/flcore/optimizers/fedoptimizer.py
-
The evaluation platform is also convenient for users to build a new platform for specific applications, such as our FL-IoT and HtFL.
If you are interested in the experimental results (e.g., the accuracy) of the above algorithms, you can find some results in our accepted FL papers (i.e., FedALA, FedCP, GPFL, and DBE) listed as follows that also use this library. Please note that this developing project may not be able to reproduce the results on these papers, since some basic settings may change due to the requests of the community. For example, we previously set shuffle=False
in clientbase.py
@inproceedings{zhang2023fedala,
title={Fedala: Adaptive local aggregation for personalized federated learning},
author={Zhang, Jianqing and Hua, Yang and Wang, Hao and Song, Tao and Xue, Zhengui and Ma, Ruhui and Guan, Haibing},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={37},
number={9},
pages={11237--11244},
year={2023}
}
@inproceedings{Zhang2023fedcp,
author = {Zhang, Jianqing and Hua, Yang and Wang, Hao and Song, Tao and Xue, Zhengui and Ma, Ruhui and Guan, Haibing},
title = {FedCP: Separating Feature Information for Personalized Federated Learning via Conditional Policy},
year = {2023},
booktitle = {Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining}
}
@inproceedings{zhang2023gpfl,
title={GPFL: Simultaneously Learning Global and Personalized Feature Information for Personalized Federated Learning},
author={Zhang, Jianqing and Hua, Yang and Wang, Hao and Song, Tao and Xue, Zhengui and Ma, Ruhui and Cao, Jian and Guan, Haibing},
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
pages={5041--5051},
year={2023}
}
@inproceedings{
zhang2023eliminating,
title={Eliminating Domain Bias for Federated Learning in Representation Space},
author={Jianqing Zhang and Yang Hua and Jian Cao and Hao Wang and Tao Song and Zhengui XUE and Ruhui Ma and Haibing Guan},
booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
year={2023},
url={https://openreview.net/forum?id=nO5i1XdUS0}
}