Towards Personalized Federated Learning via Heterogeneous Model Reassembly
Authors: Jiaqi Wang, Xingyi Yang, Suhan Cui, Liwei Che, Lingjuan Lyu, Dongkuan (DK) Xu, Fenglong Ma
NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results demonstrate that p Fed HR outperforms baselines on three datasets under both IID and Non-IID settings. Additionally, p Fed HR effectively reduces the adverse impact of using different public data and dynamically generates diverse personalized models in an automated manner. |
| Researcher Affiliation | Collaboration | Jiaqi Wang1 Xingyi Yang2 Suhan Cui1 Liwei Che1 Lingjuan Lyu3 Dongkuan Xu4 Fenglong Ma1 1The Pennsylvania State University 2National University of Singapore 3Sony AI 4North Carolina State University |
| Pseudocode | Yes | Algorithm 1: Reassembly Candidate Search [...] Algorithm 1: Algorithm Flow of p Fed HR. |
| Open Source Code | Yes | Source code can be found in the link https://github.com/Jackqq Wang/pfed HR |
| Open Datasets | Yes | Datasets. We conduct experiments for the image classification task on MNIST, SVHN, and CIFAR10 datasets under both IID and non-IID data distribution settings, respectively. We split the datasets into 80% for training and 20% for testing. |
| Dataset Splits | No | The paper mentions 'We split the datasets into 80% for training and 20% for testing' but does not specify a separate validation split or how validation was performed if it was implicitly part of training. |
| Hardware Specification | Yes | The proposed p Fed HR is implemented in Pytorch 2.0.1 and runs on NVIDIA A100 with CUDA version 12.0 on a Ubuntu 20.04.6 LTS server. |
| Software Dependencies | Yes | The proposed p Fed HR is implemented in Pytorch 2.0.1 and runs on NVIDIA A100 with CUDA version 12.0 on a Ubuntu 20.04.6 LTS server. |
| Experiment Setup | Yes | The hyperparameter λ in Eq. (6) is 0.2. We use Adam as the optimizer. The learning rate of the local client learning and the server fine-tuning learning rate are both equal to 0.001. |