FedLoGe: Joint Local and Generic Federated Learning under Long-tailed Data
Authors: Zikai Xiao, Zihan Chen, Liyinglan Liu, YANG FENG, Joey Tianyi Zhou, Jian Wu, Wanlu Liu, Howard Hao Yang, Zuozhu Liu
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experimental results on CIFAR-10/100-LT, Image Net-LT, and i Naturalist demonstrate the advantage of our method over state-of-the-art p FL and Fed-LT approaches. |
| Researcher Affiliation | Collaboration | 1Zhejiang University, 2Singapore University of Technology and Design, 3University of Electronic Science and Technology of China, 4Angelalign Technology Inc, 5IHPC, Agency for Science, Technology and Research, Singapore, 6CFAR, Agency for Science, Technology and Research, Singapore |
| Pseudocode | Yes | Algorithm 1 An overview of Fed Lo Ge framework |
| Open Source Code | Yes | Our codes are available at https://github.com/Zack Zikai Xiao/Fed Lo Ge. |
| Open Datasets | Yes | We consider image classification tasks for performance evaluation on benchmark long-tailed datasets: CIFAR-10/100-LT, Image Net-LT, and i Naturalist-User-160k (Van Horn et al., 2018). The CIFAR-10/100-LT datasets are sampled into a long-tailed distribution employing an exponential distribution governed by the Imbalance Factor (IF) in Cao et al. (2019). |
| Dataset Splits | No | The paper describes its training and testing procedures and mentions 'global balanced test set' and 'local test set', but does not explicitly specify a separate 'validation' dataset split for hyperparameter tuning or early stopping. |
| Hardware Specification | Yes | we conducted the experiments three times on the Py Torch platform utilizing the NVIDIA Ge Force RTX 3090; the average cost is 11 minutes and 52 seconds. |
| Software Dependencies | No | The paper mentions 'Py Torch platform' but does not specify its version or any other software dependencies with version numbers. |
| Experiment Setup | Yes | We use α = 1, 0.5 and IF = 50, 100 in CIFAR-10/100-LT. We use α = 0.1, 0.5 in Image Net-LT and i Naturalist, respectively. Res Net-18 is trained over K = 40 clients on CIFAR-10-LT, while Res Net-34 and Res Net-50 are implemented on CIFAR-100-LT and Image Net-LT, respectively, with K = 20 clients. We train the generic model with T = 500 rounds via applying SGD optimizer for local training in all experiments unless otherwise stated. For CIFAR10/100-LT datasets, we engaged 40 clients with full participation in each training round, setting the number of local epochs to 5. For Image Net/i Naturalist, the experiments involved 20 clients, with a 40% participation rate in each training round, and the number of local epochs was set to 3. |