Revisiting Ensembling in One-Shot Federated Learning

Authors: Youssef Allouah, Akash Dhasade, Rachid Guerraoui, Nirupam Gupta, Anne-marie Kermarrec, Rafael Pinot, Rafael Pires, Rishi Sharma

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We showcase the effectiveness of FENS through exhaustive experiments spanning several datasets and heterogeneity levels. Through extensive evaluations on several benchmark datasets (CIFAR100, CIFAR-10, SVHN, and AG-News) across different heterogeneity levels, we demonstrate the efficacy of FENS in achieving FL-like accuracy at OFL-like communication cost.
Researcher Affiliation Academia 1EPFL 2University of Copenhagen 3Sorbonne Université and Université Paris Cité, CNRS, LPSM
Pseudocode Yes Algorithms 1 and 2 (Appendix C) provide the pseudo for FENS.
Open Source Code Yes Source code available at: https://github.com/sacs-epfl/fens.
Open Datasets Yes We consider three standard vision datasets with varying level of difficulty, including SVHN [30], CIFAR-10 [30] and CIFAR-100 [30], commonly used in several OFL works [7, 10, 46] as well as one language dataset AG-News [47]. For our experiments involving the realistic healthcare FLamby benchmark, we experiment with 3 datasets: Fed-Camelyon16, Fed-Heart-Disease, and Fed-ISIC2019.
Dataset Splits Yes The testing set of each dataset is split (50-50%) for validation and testing. We use the validation set to tune hyperparameters and always report the accuracy on the testing split. In FENS, each client performs local training using 90% of their local training data while reserving 10% for the iterative aggregator training.
Hardware Specification Yes We use a cluster comprising a mix of 2x Intel Xeon Gold 6240 @ 2.5 GHz of 36 hyper-threaded cores and 2x AMD EPYC 7302 @ 3 GHz of 64 hyper-threaded cores, equipped with 4x NVIDIA Tesla V100 32G and 8x NVIDIA Tesla A100 40G GPU respectively.
Software Dependencies No We use the standard Py Torch quantization library to implement quantization in FENS. The paper mentions "Py Torch quantization library" but no specific version number for it or other software dependencies.
Experiment Setup Yes For the CIFAR-10 and CIFAR-100 datasets, each client conducts local training for 500 epochs utilizing SGD as the local optimizer with an initial learning rate of 0.0025. For the SVHN and AG-News datasets, local training extends to 50 and 20 epochs respectively with a learning rate of 0.01. The learning rate is decayed using Cosine Annealing across all datasets. We tune learning rates for each algorithm. (Table 6)