Federated Optimization in Heterogeneous Networks

This repository contains the code and experiments for the paper:

Federated Optimization in Heterogeneous Networks

MLSys 2020

Federated Learning is a distributed learning paradigm with two key challenges that differentiate it from traditional distributed optimization: (1) significant variability in terms of the systems characteristics on each device in the network (systems heterogeneity), and (2) non-identically distributed data across the network (statistical heterogeneity). In this work, we introduce a framework, FedProx, to tackle heterogeneity in federated networks, both theoretically and empirically.

This repository contains a set of detailed empirical evaluation across a suite of federated datasets. We show that FedProx allows for more robust convergence than FedAvg. In particular, in highly heterogeneous settings, FedProx demonstrates significantly more stable and accurate convergence behavior relative to FedAvg—improving absolute test accuracy by 22% on average.

General Guidelines

Note that if you would like to use FedProx as a baseline and run our code:

• If you are using different datasets, then at least the learning rates and the mu parameter need to be tuned based on your metric. You might want to tune mu from {0.001, 0.01, 0.1, 0.5, 1}. There are no default mu values that would work for all settings.

• If you are using the same datasets as those used here, then need to use the same learning rates and mu reported in our paper.

Preparation

Dataset generation

We already provide four synthetic datasets that are used in the paper under corresponding folders. For all datasets, see the README files in separate data/$dataset folders for instructions on preprocessing and/or sampling data.

The statistics of real federated datasets are summarized as follows.

Downloading dependencies

pip3 install -r requirements.txt  

Run on synthetic federated data

(1) You don't need a GPU to run the synthetic data experiments:

export CUDA_VISIBLE_DEVICES=

(2) Run the instructions as follows, and the log files will be automatically stored for drawing figures later.

bash run_fedavg.sh synthetic_iid 0 | tee log_synthetic/synthetic_iid_client10_epoch20_mu0
bash run_fedprox.sh synthetic_iid 0 1 | tee log_synthetic/synthetic_iid_client10_epoch20_mu1
bash run_fedavg.sh synthetic_0_0 0 | tee log_synthetic/synthetic_0_0_client10_epoch20_mu0
bash run_fedprox.sh synthetic_0_0 0 1 | tee log_synthetic/synthetic_0_0_client10_epoch20_mu1
bash run_fedavg.sh synthetic_0.5_0.5 0 | tee log_synthetic/synthetic_0.5_0.5_client10_epoch20_mu0
bash run_fedprox.sh synthetic_0.5_0.5 0 1 | tee log_synthetic/synthetic_0.5_0.5_client10_epoch20_mu1
bash run_fedavg.sh synthetic_1_1 0 | tee log_synthetic/synthetic_1_1_client10_epoch20_mu0
bash run_fedprox.sh synthetic_1_1 0 1 | tee log_synthetic/synthetic_1_1_client10_epoch20_mu1

(3) Draw figures to reproduce results on synthetic data

python plot_fig2.py loss     # training loss
python plot_fig2.py accuracy # testing accuracy
python plot_fig2.py dissim   # dissimilarity metric

The training loss, testing accuracy, and dissimilarity metric figures are saved as loss.pdf, accuracy.pdf and dissim.pdf respectively, under the current folder where you call plot_fig2.py. You can check that these figures reproduce the results in Figure 2 in the paper. Make sure to use the default hyper-parameters in run_fedavg.sh/run_fedprox.sh for synthetic data.

For example, the training loss for synthetic datasets would look like this:

Run on real federated datasets

(1) Specify a GPU id if needed:

export CUDA_VISIBLE_DEVICES=available_gpu_id

Otherwise just run to CPUs [might be slow if testing on Neural Network models]:

export CUDA_VISIBLE_DEVICES=

(2) Run on one dataset. First, modify the run_fedavg.sh and run_fedprox.sh scripts, specify the corresponding model of that dataset (choose from flearn/models/$DATASET/$MODEL.py and use $MODEL as the model name), specify a log file name, and configure all other parameters such as learning rate (see all hyper-parameters values in the appendix of the paper).

For example, for all the synthetic data:

fedavg.sh:

python3  -u main.py --dataset=$1 --optimizer='fedavg'  \
            --learning_rate=0.01 --num_rounds=200 --clients_per_round=10 \
            --eval_every=1 --batch_size=10 \
            --num_epochs=20 \
            --drop_percent=$2 \
            --model='mclr' 

fedprox.sh:

python3  -u main.py --dataset=$1 --optimizer='fedprox'  \
            --learning_rate=0.01 --num_rounds=200 --clients_per_round=10 \
            --eval_every=1 --batch_size=10 \
            --num_epochs=20 \
            --drop_percent=$2 \
            --model='mclr' \
            --mu=$3

Then run:

mkdir synthetic_1_1
bash run_fedavg.sh synthetic_1_1 0 | tee synthetic_1_1/fedavg_drop0
bash run_fedprox.sh synthetic_1_1 0 0 | tee synthetic_1_1/fedprox_drop0_mu0
bash run_fedprox.sh synthetic_1_1 0 1 | tee synthetic_1_1/fedprox_drop0_mu1

bash run_fedavg.sh synthetic_1_1 0.5 | tee synthetic_1_1/fedavg_drop0.5
bash run_fedprox.sh synthetic_1_1 0.5 0 | tee synthetic_1_1/fedprox_drop0.5_mu0
bash run_fedprox.sh synthetic_1_1 0.5 1 | tee synthetic_1_1/fedprox_drop0.5_mu1

bash run_fedavg.sh synthetic_1_1 0.9 | tee synthetic_1_1/fedavg_drop0.9
bash run_fedprox.sh synthetic_1_1 0.9 0 | tee synthetic_1_1/fedprox_drop0.9_mu0
bash run_fedprox.sh synthetic_1_1 0.9 1 | tee synthetic_1_1/fedprox_drop0.9_mu1

And the test accuracy, training loss, and dissimilarity numbers will be saved in the log files.

(3) After you collect logs for all the 5 datasets in Figure 1 (synthetic, mnist, femnist, shakespeare, sent140) (the log directories should be [synthetic_1_1, mnist, femnist, shakespeare, sent140]), run:

python plot_final_e20.py loss

to reproduce results in Figure 1 (the generated figure is called loss_full.pdf).

Note: If you only want to quickly verify the results on the first synthetic dataset, you can modify the plot_final_e20.py script by changing range(5) in Line 54 to range(1), and run python plot_final_e20.py loss.

Note: It might take a much longer time to run on real datasets than synthetic data because real federated datasets are larger and some of the models are deep neural networks.

References

See our FedProx paper for more details as well as all references.