Personalized Federated Learning with Moreau Envelopes (NeurIPS 2020)

This repository implements all experiments in the paper Personalized Federated Learning with Moreau Envelopes.

Authors: Canh T. Dinh, Nguyen H. Tran, Tuan Dung Nguyen

Full paper: https://arxiv.org/pdf/2006.08848.pdf https://proceedings.neurips.cc/paper/2020/file/f4f1f13c8289ac1b1ee0ff176b56fc60-Paper.pdf

Paper has been accepted by NeurIPS 2020.

This repository does not only implement pFedMe but also FedAvg, and Per-FedAvg algorithms. (Federated Learning using Pytorch)

Software requirements:

• numpy, scipy, torch, Pillow, matplotlib.

• To download the dependencies: pip3 install -r requirements.txt

Dataset: We use 2 datasets: MNIST and Synthetic

• To generate non-idd MNIST Data:

  • Access data/Mnist and run: "python3 generate_niid_20users.py"
  • We can change the number of user and number of labels for each user using 2 variable NUM_USERS = 20 and NUM_LABELS = 2

• To generate idd MNIST Data (we do not use iid data in the paper):

  • Access data/Mnist and run: "python3 generate_iid_20users.py"

• To generate niid Synthetic:

  • Access data/Synthetic and run: "python3 generate_synthetic_05_05.py". Similar to MNIST data, the Synthetic data is configurable with the number of users and the numbers of labels for each user.

• The datasets also are available to download at: https://drive.google.com/drive/folders/1-Z3FCZYoisqnIoLLxOljMPmP70t2TGwB?usp=sharing

Produce experiments and figures

• There is a main file "main.py" which allows running all experiments.

Using same parameters

• To produce the comparison experiments for pFedMe using MNIST dataset:

  • Strongly Convex Case, run below commands:

    
    python3 main.py --dataset Mnist --model mclr --batch_size 20 --learning_rate 0.005 --personal_learning_rate 0.1 --beta 1 --lamda 15 --num_global_iters 800 --local_epochs 20 --algorithm pFedMe --numusers 5 --times 10
    python3 main.py --dataset Mnist --model mclr --batch_size 20 --learning_rate 0.005 --num_global_iters 800 --local_epochs 20 --algorithm FedAvg --numusers 5  --times 10
    python3 main.py --dataset Mnist --model mclr --batch_size 20 --learning_rate 0.005 --beta 0.001  --num_global_iters 800 --local_epochs 20 --algorithm PerAvg --numusers 5  --times 10
    

• It is noted that each algorithm should be run at least 10 times and then the results are averaged.

• All the train loss, testing accuracy, and training accuracy will be stored as h5py file in the folder "results". It is noted that we store the data for persionalized model and global of pFedMe in 2 separate files following format: DATASET_pFedMe_p_x_x_xu_xb_x_avg.h5 and DATASET_pFedMe_x_x_xu_xb_x_avg.h5 respectively (pFedMe for global model, pFedMe_p for personalized model of pFedMe, PerAvg_p is for personalized model of PerAvg).

• In order to plot the figure for convex case, set parameters in file main_plot.py similar to parameters run from previous experiments. It is noted that each experiment with different parameters will have different results, the configuration in the plot function should be modified for each specific case. For example. To plot the comparision in convex case for the above experiments, in the main_plot.py set:

  
  numusers = 5
  num_glob_iters = 800
  dataset = "Mnist"
  local_ep = [20,20,20,20]
  lamda = [15,15,15,15]
  learning_rate = [0.005, 0.005, 0.005, 0.005]
  beta =  [1.0, 1.0, 0.001, 1.0]
  batch_size = [20,20,20,20]
  K = [5,5,5,5]
  personal_learning_rate = [0.1,0.1,0.1,0.1]
  algorithms = [ "pFedMe_p","pFedMe","PerAvg_p","FedAvg"]
  plot_summary_one_figure_mnist_Compare(num_users=numusers, loc_ep1=local_ep, Numb_Glob_Iters=num_glob_iters, lamb=lamda,
                             learning_rate=learning_rate, beta = beta, algorithms_list=algorithms, batch_size=batch_size, dataset=dataset, k = K, personal_learning_rate = personal_learning_rate)
  

  • NonConvex case:

    
    python3 main.py --dataset Mnist --model dnn --batch_size 20 --learning_rate 0.005 --personal_learning_rate 0.09 --beta 1 --lamda 15 --num_global_iters 800 --local_epochs 20 --algorithm pFedMe --numusers 5 --times 10
    python3 main.py --dataset Mnist --model dnn --batch_size 20 --learning_rate 0.005 --num_global_iters 800 --local_epochs 20 --algorithm FedAvg --numusers 5 --times 10
    python3 main.py --dataset Mnist --model dnn --batch_size 20 --learning_rate 0.005 --beta 0.001  --num_global_iters 800 --local_epochs 20 --algorithm PerAvg --numusers 5 --times 10
    

    To plot the figure for non-convex case, we do similar to convex case, also need to change the parameters in main_plot.py.

• To produce the comparision experiment for pFedMe using Synthetic dataset:

  • Strongly Convex Case:

    
    python3 main.py --dataset Synthetic --model mclr --batch_size 20 --learning_rate 0.005 --personal_learning_rate 0.01 --beta 1 --lamda 20 --num_global_iters 600 --local_epochs 20 --algorithm pFedMe --numusers 10 --times 10
    python3 main.py --dataset Synthetic --model mclr --batch_size 20 --learning_rate 0.005 --num_global_iters 600 --local_epochs 20 --algorithm FedAvg --numusers 10 --times 10
    python3 main.py --dataset Synthetic --model mclr --batch_size 20 --learning_rate 0.005 --beta 0.001  --num_global_iters 600 --local_epochs 20 --algorithm PerAvg --numusers 10 --times 10
    

  • NonConvex case:

    
    python3 main.py --dataset Synthetic --model dnn --batch_size 20 --learning_rate 0.005 --personal_learning_rate 0.01 --beta 1 --lamda 20 --num_global_iters 600 --local_epochs 20 --algorithm pFedMe --numusers 10 --times 10
    python3 main.py --dataset Synthetic --model dnn --batch_size 20 --learning_rate 0.005 --num_global_iters 600 --local_epochs 20 --algorithm FedAvg --numusers 10 --times 10
    python3 main.py --dataset Synthetic --model dnn --batch_size 20 --learning_rate 0.005 --beta 0.001  --num_global_iters 600 --local_epochs 20 --algorithm PerAvg --numusers 10 --times 10
    

Fine-tuned Parameters:

To produce results in the table of fine-tune parameter:

• MNIST:

  • Strongly Convex Case:

    
    python3 main.py --dataset Mnist --model mclr --batch_size 20 --learning_rate 0.01 --personal_learning_rate 0.1 --beta 2 --lamda 15 --num_global_iters 800 --local_epochs 20 --algorithm pFedMe --numusers 5 --times 10
    python3 main.py --dataset Mnist --model mclr --batch_size 20 --learning_rate 0.02 --num_global_iters 800 --local_epochs 20 --algorithm FedAvg --numusers 5 --times 10
    python3 main.py --dataset Mnist --model mclr --batch_size 20 --learning_rate 0.03 --beta 0.003  --num_global_iters 800 --local_epochs 20 --algorithm PerAvg --numusers 5 --times 10
    

  • NonConvex Case:

    
    python3 main.py --dataset Mnist --model dnn --batch_size 20 --learning_rate 0.01 --personal_learning_rate 0.05 --beta 2 --lamda 30 --num_global_iters 800 --local_epochs 20 --algorithm pFedMe --numusers 5 --times 10
    python3 main.py --dataset Mnist --model dnn --batch_size 20 --learning_rate 0.02 --num_global_iters 800 --local_epochs 20 --algorithm FedAvg --numusers 5 --times 10
    python3 main.py --dataset Mnist --model dnn --batch_size 20 --learning_rate 0.02 --beta 0.001  --num_global_iters 800 --local_epochs 20 --algorithm PerAvg --numusers 5 --times 10
    

• Sythetic:

  • Strongly Convex Case:

    
    python3 main.py --dataset Synthetic --model mclr --batch_size 20 --learning_rate 0.01 --personal_learning_rate 0.01 --beta 2 --lamda 20 --num_global_iters 600 --local_epochs 20 --algorithm pFedMe --numusers 10 --times 10
    python3 main.py --dataset Synthetic --model mclr --batch_size 20 --learning_rate 0.02 --num_global_iters 600 --local_epochs 20 --algorithm FedAvg --numusers 10 --times 10
    python3 main.py --dataset Synthetic --model mclr --batch_size 20 --learning_rate 0.02 --beta 0.002  --num_global_iters 600 --local_epochs 20 --algorithm PerAvg --numusers 10 --times 10
    

  • NonConvex Case:

    
    python3 main.py --dataset Synthetic --model dnn --batch_size 20 --learning_rate 0.01 --personal_learning_rate 0.01 --beta 2 --lamda 30 --num_global_iters 600 --local_epochs 20 --algorithm pFedMe --numusers 10 --times 10
    python3 main.py --dataset Synthetic --model dnn --batch_size 20 --learning_rate 0.03 --num_global_iters 600 --local_epochs 20 --algorithm FedAvg --numusers 10 --times 10
    python3 main.py --dataset Synthetic --model dnn --batch_size 20 --learning_rate 0.01 --beta 0.001  --num_global_iters 600 --local_epochs 20 --algorithm PerAvg --numusers 10 --times 10
    

Effect of hyper-parameters:

For all the figures for effect of hyper-parameters, we use Mnist dataset and fix the learning_rate == 0.005 and personal_learning_rate == 0.09 for all experiments. Other parameters are changed according to the experiments. Only in the experiments for the effects of $\beta$, in case $\beta = 4$, we use learning_rate == 0.003 to stable the algorithm.

CIFAR-10 dataset:

The implementation of Cifar10 has been finished. However, we haven't fine-tuned the parameters for all algorithms on Cifar10. Below is the comment to run cifar10 on pFedMe.

python3 main.py --dataset Cifar10 --model cnn --batch_size 20 --learning_rate 0.01 --personal_learning_rate 0.01 --beta 1 --lamda 15 --num_global_iters 800 --local_epochs 20 --algorithm pFedMe --numusers 5