FetchSGD: Communication-Efficient Federated Learning with Sketching

Authors: Daniel Rothchild, Ashwinee Panda, Enayat Ullah, Nikita Ivkin, Ion Stoica, Vladimir Braverman, Joseph Gonzalez, Raman Arora

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We prove that Fetch SGD has favorable convergence guarantees, and we demonstrate its empirical effectiveness by training two residual networks and a transformer model.
Researcher Affiliation Collaboration 1University of California, Berkeley, California, USA 2Johns Hopkins University, Baltimore, Maryland 3Amazon.
Pseudocode Yes Fetch SGD is presented in full in Algorithm 1.
Open Source Code Yes Code available at https://github.com/ kiddyboots216/Comm Efficient. Git commit at the time of camera-ready: 833ca44.
Open Datasets Yes CIFAR10 and CIFAR100 (Krizhevsky et al., 2009)... Federated EMNIST... (Caldas et al., 2018)... We finetune a pretrained GPT2 on the Persona Chat dataset... (Zhang et al., 2018).
Dataset Splits Yes CIFAR10 (CIFAR100) we use 10,000 (50,000) clients... We report accuracy on the test datasets... We report final accuracies on the validation dataset.
Hardware Specification No The paper does not specify the exact hardware (e.g., CPU, GPU models, cloud instance types) used for running the experiments.
Software Dependencies No We implement and compare Fetch SGD, gradient sparsification (local top-k), and Fed Avg using Py Torch (Paszke et al., 2019).
Experiment Setup Yes Details on which hyperparameters were chosen for each task can be found in Appendix A.1.