CLIP-Guided Federated Learning on Heterogeneity and Long-Tailed Data
Authors: Jiangming Shi, Shanshan Zheng, Xiangbo Yin, Yang Lu, Yuan Xie, Yanyun Qu
AAAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experimental results on several benchmarks demonstrate that CLIP2FL achieves impressive performance and effectively deals with data heterogeneity and long-tail distribution. The code is available at https://github.com/shijiangming1/CLIP2FL. |
| Researcher Affiliation | Academia | Jiangming Shi1, Shanshan Zheng2, Xiangbo Yin2, Yang Lu2, Yuan Xie3, 4 , Yanyun Qu1, 2* 1 Institute of Artificial Intelligence, Xiamen University 2 School of Informatics, Xiamen University 3 East China Normal University 4 Chongqing Institute of East China Normal University |
| Pseudocode | Yes | Algorithm 1: Training Process for Round t |
| Open Source Code | Yes | The code is available at https://github.com/shijiangming1/CLIP2FL. |
| Open Datasets | Yes | Datasets. We implement CLIP2FL on three frequently used datasets with the long-tailed data: CIFAR-10/100LT (Krizhevsky, Hinton et al. 2009) and Image Net-LT (Russakovsky et al. 2015). |
| Dataset Splits | No | No explicit mention of specific train/validation/test dataset splits (e.g., percentages, sample counts) for the overall datasets was found. The paper describes data partitioning among clients and long-tailed distribution generation. |
| Hardware Specification | Yes | Experiments were conducted using Py Torch on four NVIDIA Ge Force RTX 3090 GPUs. |
| Software Dependencies | No | The paper mentions 'Py Torch' but does not specify a version number or other software dependencies with their versions. |
| Experiment Setup | Yes | The number of clients is set to 20, and 40% of them are randomly selected as online clients to participate in training. The batch size of client-side training is set to 32 for all datasets and we set the number of federated features to 100 for each class. ...We employed the standard cross-entropy loss by default and executed 200 communication rounds. ... Three important hyperparameters in our CLIP2FL are β, η and m. We found that CLIP2FL achieved the best performance when β = 3.0, η {0.001, 0.0001, 1e 5} and m = 100. |