Accelerating Heterogeneous Federated Learning with Closed-form Classifiers
Authors: Eros Fanı̀, Raffaello Camoriano, Barbara Caputo, Marco Ciccone
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we empirically evaluate the performances of our proposed methods in terms of accuracy, convergence speed, communication, and computational costs. |
| Researcher Affiliation | Academia | Eros Fanì 1 Raffaello Camoriano 1 2 Barbara Caputo 1 3 Marco Ciccone 1 1Department of Computing and Control Engineering, Polytechnic University of Turin, Italy 2Istituto Italiano di Tecnologia, Genoa, Italy 3CINI Consortium, Rome, Italy. |
| Pseudocode | Yes | Algorithm 1 FED3R and FED3R-RF |
| Open Source Code | Yes | Official website: https://fed-3r.github.io/. |
| Open Datasets | Yes | For the evaluation we choose two large-scale image classification datasets, Landmarks (Weyand et al., 2020) and i Naturalist (Van Horn et al., 2018), both parti- tioned as proposed in (Hsu et al., 2020) 2. |
| Dataset Splits | Yes | For the evaluation we choose two large-scale image classification datasets, Landmarks (Weyand et al., 2020) and i Naturalist (Van Horn et al., 2018), both parti- tioned as proposed in (Hsu et al., 2020) 2. |
| Hardware Specification | Yes | We run all the experiments using an NVIDIA A100-SXM4-40GB using the FL clients partitions provided by (Hsu et al., 2020) for Landmarks (Weyand et al., 2020) and i Naturalist (Van Horn et al., 2018). |
| Software Dependencies | No | The paper mentions optimizers (SGD) and network architectures (Mobile Net V2) but does not provide specific version numbers for software libraries or dependencies (e.g., Python, PyTorch, CUDA versions). |
| Experiment Setup | Yes | We conducted the Landmarks experiments for 3000 rounds, the i Naturalist experiments for 5000 rounds, and the Cifar100 experiments for 1500 rounds. We sampled 10 clients per round in all three cases unless stated otherwise. We utilized SGD as the client optimizer with a learning rate (lr) of 0.1 and a weight decay (wd) of 4 10 5, a batch size of 50, and 5 local epochs for both Landmarks and i Naturalist, and 1 local epoch for Cifar100. Additionally, we employed SGD as the server optimizer (Reddi et al., 2021) with a learning rate (slr) set to 1.0 and no momentum (smom). |