On the Convergence of Federated Averaging with Cyclic Client Participation

Authors: Yae Jee Cho, Pranay Sharma, Gauri Joshi, Zheng Xu, Satyen Kale, Tong Zhang

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental 5. Experimental Results Setup. We train ML models on standard datasets using Fed Avg with Cy CP for different client local updated procedures to see how cyclicity affects the performance of FL. We experiment with image classification using an MLP for the FMNIST (Xiao et al., 2017) dataset and EMNIST dataset (Cohen et al., 2017)...
Researcher Affiliation Collaboration 1Carnegie Mellon University, USA 2Google Research, USA 3Hong Kong University of Science and Technology, Hong Kong.
Pseudocode Yes Algorithm 1 Cy CP Framework in FL
Open Source Code Yes The code used for all experiments is included in the supplementary material.
Open Datasets Yes We experiment with image classification using an MLP for the FMNIST (Xiao et al., 2017) dataset and EMNIST dataset (Cohen et al., 2017) with 62 labels where we have 100 and 500 clients in total and select 5 and 10 clients per communication round respectively.
Dataset Splits Yes For all experiments, the data is partitioned to 80%/10%/10% for training/validation/test data, where the training data then is again partitioned across the clients heterogeneously.
Hardware Specification Yes All experiments are conducted on clusters equipped with one NVIDIA Titan X GPU.
Software Dependencies Yes The algorithms are implemented in Py Torch 1. 11. 0.
Experiment Setup Yes Specifically, we do a grid search over the learning rate: η {0.05, 0.01, 0.005, 0.001}, batch size: b {32, 64, 128}, and local iterations: τ {5, 10, 30, 50} to find the hyper-parameters with the highest test accuracy for each benchmark.