DONE: Distributed Newton Method for Federated Edge Learning (IEEE Transactions on Parallel and Distributed Systems 2022).

This repository is for the Experiment Section of the paper: "DONE: Distributed Newton Method for Federated Edge Learning"

Link: https://arxiv.org/pdf/2012.05625.pdf, https://ieeexplore.ieee.org/abstract/document/9695269

Software requirements:

• numpy, scipy, pytorch, Pillow, matplotlib.

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

• The code can be run on any pc.

Dataset: We use 2 datasets: MNIST, and Synthetic

• To generate non-idd MNIST Data:

  • Access data/Mnist and run: "python3 generate_niid_32users.py"
  • We can change the number of user and number of labels for each user using 2 variable NUM_USERS = 32 and NUM_LABELS = 3

• To generate non-iid Synthetic:

  • Access data/Linear_synthetic and run: "python3 generate_niid_linear_32users_updated.py". Synthetic data is configurable with the number of users, the numbers of labels for each user, and the value of $\kappa$.

• The datasets also are available to download at: https://drive.google.com/drive/folders/1LkBjkP0PzfRNiAY9ImN85r9vBIuW4U6-?usp=sharing

Produce experiments and figures

• There is a main file "main.py" which allows running all experiments, and 2 files: "plot_mnist.py", "plot_synthetic.py" to plot all results after runing all experiment. Only run "plot_mnist.py" and "plot_synthetic.py" after getting the results from training process.

Performance comparison with different distributed algorithms (table 2 in our paper)

• For MNIST: <pre><code> python3 main.py --dataset Mnist --model mclr --algorithm DONE --batch_size 0 --alpha 0.03 --num_global_iters 100 --local_epochs 40 --numedges 32 python3 main.py --dataset Mnist --model mclr --algorithm Newton --batch_size 0 --alpha 0.03 --num_global_iters 100 --local_epochs 40 --numedges 32 python3 main.py --dataset Mnist --model mclr --algorithm Newton --batch_size 0 --alpha 0.05 --num_global_iters 100 --local_epochs 40 --numedges 32 python3 main.py --dataset Mnist --model mclr --algorithm DANE --batch_size 0 --eta 1 --learning_rate 0.04 --num_global_iters 100 --local_epochs 40 --numedges 32 python3 main.py --dataset Mnist --model mclr --algorithm FEDL --batch_size 0 --eta 1 --learning_rate 0.04 --num_global_iters 100 --local_epochs 40 --numedges 32 python3 main.py --dataset Mnist --model mclr --algorithm GD --batch_size 0 --learning_rate 0.2 --num_global_iters 100 --numedges 32 python3 main.py --dataset Mnist --model mclr --algorithm GT --batch_size 0 --alpha 0.03 --L 1 --num_global_iters 100 --local_epochs 40 --numedges 32 </code></pre>

• For FEMNIST: <pre><code> python3 main.py --dataset Nist --model mclr --algorithm DONE --batch_size 0 --alpha 0.01 --num_global_iters 100 --local_epochs 40 --numedges 32 python3 main.py --dataset Nist --model mclr --algorithm Newton --batch_size 0 --alpha 0.01 --num_global_iters 100 --local_epochs 40 --numedges 32 python3 main.py --dataset Nist --model mclr --algorithm Newton --batch_size 0 --alpha 0.01 --num_global_iters 100 --local_epochs 40 --numedges 32 python3 main.py --dataset Nist --model mclr --algorithm DANE --batch_size 0 --eta 1 --learning_rate 0.02 --num_global_iters 100 --local_epochs 40 --numedges 32 python3 main.py --dataset Nist --model mclr --algorithm FEDL --batch_size 0 --eta 1 --learning_rate 0.02 --num_global_iters 100 --local_epochs 40 --numedges 32 python3 main.py --dataset Nist --model mclr --algorithm GD --batch_size 0 --learning_rate 0.02 --num_global_iters 100 --numedges 32 python3 main.py --dataset Nist --model mclr --algorithm GT --batch_size 0 --alpha 0.01 --L 1 --num_global_iters 100 --local_epochs 40 --numedges 32 </code></pre>

• For Human Activities: <pre><code> python3 main.py --dataset human_activity --model mclr --algorithm DONE --batch_size 0 --alpha 0.01 --num_global_iters 100 --local_epochs 40 --numedges 30 python3 main.py --dataset human_activity --model mclr --algorithm Newton --batch_size 0 --alpha 0.01 --num_global_iters 100 --local_epochs 40 --numedges 30 python3 main.py --dataset human_activity --model mclr --algorithm Newton --batch_size 0 --alpha 0.03 --num_global_iters 100 --local_epochs 40 --numedges 30 python3 main.py --dataset human_activity --model mclr --algorithm DANE --batch_size 0 --eta 1 --learning_rate 0.05 --num_global_iters 100 --local_epochs 40 --numedges 30 python3 main.py --dataset human_activity --model mclr --algorithm FEDL --batch_size 0 --eta 1 --learning_rate 0.05 --num_global_iters 100 --local_epochs 40 --numedges 30 python3 main.py --dataset human_activity --model mclr --algorithm GD --batch_size 0 --learning_rate 0.1 --num_global_iters 100 --numedges 30 python3 main.py --dataset human_activity --model mclr --algorithm GT --batch_size 0 --alpha 0.01 --L 1 --num_global_iters 100 --local_epochs 40 --numedges 30 </code></pre>

Effect of hyper-parameters: $\alpha$, $R $, and $\kappa$

• To produce the Fig.1 : Effects of various values of $\alpha$ and $R$ on synthetic with diffent $\kappa = 10^2$ <pre><code> python3 main.py --dataset Linear_synthetic --model linear_regression --algorithm DONE --batch_size 0 --alpha 0.06 --num_global_iters 100 --local_epochs 20 --numedges 32 python3 main.py --dataset Linear_synthetic --model linear_regression --algorithm DONE --batch_size 0 --alpha 0.08 --num_global_iters 100 --local_epochs 20 --numedges 32 python3 main.py --dataset Linear_synthetic --model linear_regression --algorithm DONE --batch_size 0 --alpha 0.1 --num_global_iters 100 --local_epochs 20 --numedges 32 python3 main.py --dataset Linear_synthetic --model linear_regression --algorithm DONE --batch_size 0 --alpha 0.2 --num_global_iters 100 --local_epochs 20 --numedges 32 python3 main.py --dataset Linear_synthetic --model linear_regression --algorithm DONE --batch_size 0 --alpha 0.1 --num_global_iters 100 --local_epochs 5 --numedges 32 python3 main.py --dataset Linear_synthetic --model linear_regression --algorithm DONE --batch_size 0 --alpha 0.1 --num_global_iters 100 --local_epochs 10 --numedges 32 python3 main.py --dataset Linear_synthetic --model linear_regression --algorithm DONE --batch_size 0 --alpha 0.1 --num_global_iters 100 --local_epochs 20 --numedges 32 python3 main.py --dataset Linear_synthetic --model linear_regression --algorithm DONE --batch_size 0 --alpha 0.1 --num_global_iters 100 --local_epochs 30 --numedges 32 </code></pre>

Effects of various values of $\alpha$ and $R$

• MNIST

  <pre><code> python3 main.py --dataset Mnist --model mclr --algorithm DONE --batch_size 0 --alpha 0.005 --num_global_iters 100 --local_epochs 40 --numedges 32 python3 main.py --dataset Mnist --model mclr --algorithm DONE --batch_size 0 --alpha 0.01 --num_global_iters 100 --local_epochs 40 --numedges 32 python3 main.py --dataset Mnist --model mclr --algorithm DONE --batch_size 0 --alpha 0.02 --num_global_iters 100 --local_epochs 40 --numedges 32 python3 main.py --dataset Mnist --model mclr --algorithm DONE --batch_size 0 --alpha 0.03 --num_global_iters 100 --local_epochs 10 --numedges 32 python3 main.py --dataset Mnist --model mclr --algorithm DONE --batch_size 0 --alpha 0.03 --num_global_iters 100 --local_epochs 20 --numedges 32 python3 main.py --dataset Mnist --model mclr --algorithm DONE --batch_size 0 --alpha 0.03 --num_global_iters 100 --local_epochs 30 --numedges 32 </code></pre>

• FEMNIST

  <pre><code> python3 main.py --dataset Nist --model mclr --algorithm DONE --batch_size 0 --alpha 0.004 --num_global_iters 100 --local_epochs 40 --numedges 32 python3 main.py --dataset Nist --model mclr --algorithm DONE --batch_size 0 --alpha 0.006 --num_global_iters 100 --local_epochs 40 --numedges 32 python3 main.py --dataset Nist --model mclr --algorithm DONE --batch_size 0 --alpha 0.008 --num_global_iters 100 --local_epochs 40 --numedges 32 python3 main.py --dataset Nist --model mclr --algorithm DONE --batch_size 0 --alpha 0.01 --num_global_iters 100 --local_epochs 10 --numedges 32 python3 main.py --dataset Nist --model mclr --algorithm DONE --batch_size 0 --alpha 0.01 --num_global_iters 100 --local_epochs 20 --numedges 32 python3 main.py --dataset Nist --model mclr --algorithm DONE --batch_size 0 --alpha 0.01 --num_global_iters 100 --local_epochs 30 --numedges 32 </code></pre>

• Human activies

  <pre><code> python3 main.py --dataset human_activity --model mclr --algorithm DONE --batch_size 0 --alpha 0.005 --num_global_iters 100 --local_epochs 40 --numedges 30 python3 main.py --dataset human_activity --model mclr --algorithm DONE --batch_size 0 --alpha 0.01 --num_global_iters 100 --local_epochs 40 --numedges 30 python3 main.py --dataset human_activity --model mclr --algorithm DONE --batch_size 0 --alpha 0.015 --num_global_iters 100 --local_epochs 40 --numedges 30 python3 main.py --dataset human_activity --model mclr --algorithm DONE --batch_size 0 --alpha 0.02 --num_global_iters 100 --local_epochs 10 --numedges 30 python3 main.py --dataset human_activity --model mclr --algorithm DONE --batch_size 0 --alpha 0.02 --num_global_iters 100 --local_epochs 20 --numedges 30 python3 main.py --dataset human_activity --model mclr --algorithm DONE --batch_size 0 --alpha 0.02 --num_global_iters 100 --local_epochs 30 --numedges 30 </code></pre>

Effects of mini batch

• MNIST <pre><code> python3 main.py --dataset Mnist --model mclr --algorithm DONE --batch_size 256 --alpha 0.01 --num_global_iters 100 --local_epochs 120 --numedges 32 python3 main.py --dataset Mnist --model mclr --algorithm DONE --batch_size 128 --alpha 0.01 --num_global_iters 100 --local_epochs 120 --numedges 32