FedRolex: Model-Heterogeneous Federated Learning with Rolling Sub-Model Extraction

Authors: Samiul Alam, Luyang Liu, Ming Yan, Mi Zhang

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the performance of Fed Rolex under two regimes: i) small-model small-dataset regime, and ii) large-model large-dataset regime
Researcher Affiliation Collaboration 1Michigan State University, 2The Ohio State University, 3Google Research, 4The Chinese University of Hong Kong, Shenzhen
Pseudocode Yes The pseudocode of Fed Rolex is in Algorithm 1.
Open Source Code Yes Our code is available at: https://github.com/AIo TMLSys-Lab/Fed Rolex.
Open Datasets Yes We train pre-activated Res Net18 (Pre Res Net18) models [26] on CIFAR-10 and CIFAR-100 [27]. ... For Stack Overflow, we followed [2] to use a 10% dropout rate to prevent overfitting, and 200 clients are randomly selected from a pool of 342, 477 clients in each communication round. The statistics of the datasets are listed in Table 2.
Dataset Splits Yes Table 2: Dataset statistics. ... Stack Overflow: Validation Clients 38,758, Validation Examples 16,491,230
Hardware Specification Yes We implemented Fed Rolex and PT-based baselines using Py Torch [31] and Ray [32], and conducted our experiments on 8 NVIDIA A6000 GPUs.
Software Dependencies No The paper mentions using 'Py Torch [31] and Ray [32]' but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes The details of the hyper-parameters for model training are included in Appendix A.7. ... To ensure a fair comparison, all the PT-based baselines are trained using the same learning rate, number of communication rounds, and multi-step learning rate decay schedule. The details of the schedule for each dataset and experiment are described in Appendix A.7.