Better Generative Replay for Continual Federated Learning
Authors: Daiqing Qi, Handong Zhao, Sheng Li
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results on multiple benchmark datasets demonstrate that our method significantly outperforms baselines. Code is available at: https://github.com/daiqing98/Fed CIL. |
| Researcher Affiliation | Collaboration | Daiqing Qi1, Handong Zhao2, Sheng Li1 1University of Virginia, 2Adobe Research {daiqing.qi, shengli}@virginia.edu, hazhao@adobe.com |
| Pseudocode | No | The paper does not contain any pseudocode or clearly labeled algorithm blocks. |
| Open Source Code | Yes | Code is available at: https://github.com/daiqing98/Fed CIL. |
| Open Datasets | Yes | We adopt commonly used datasets for federated learning (FL) (Zhu et al., 2021b), suggested by the LEAF benchmark (Caldas et al., 2018). Proper modifications are made according to the definition of our CI-FL setting. For vision tasks, MNIST, EMINST and Celeb A are suggested. We do not use Celeb A because it is built as a binary classification task in FL and not suitable for continual learning setting. Besides, we add use CIFAR-10, which is a simple but still challenging dataset in data-free class-incremental learning (CIL). They are simple datasets for classification, but difficult and far from being solved in non-IID FL and CIL (Prabhu et al., 2020). For all datasets: MNIST, EMNIST-Letters, EMNIST-Balanced and CIFAR-10, we build 5 tasks for each client with 2 classes every task. Details of datasets and data processing can be found in Appendix A. |
| Dataset Splits | No | The paper mentions training and testing sets (e.g., "MNIST: It is a digit image classification dataset of 10 classes with a training set of 60,000 instances and a test set of 10,000 instances.") but does not explicitly state a validation split or the use of a separate validation set for hyperparameter tuning or early stopping. |
| Hardware Specification | No | The paper does not provide specific details about the hardware used to run the experiments, such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The paper mentions the use of "Adam optimizer" and "learning rate 1e-4" but does not specify any software dependencies with version numbers (e.g., Python, PyTorch, TensorFlow versions, or specific library versions). |
| Experiment Setup | Yes | Unless otherwise specified, we set local training iteration T = 400 and global communication round R = 200 for all models. For each local iteration, we adopt mini-batch size B = 32 for MNIST, EMNIST-L, EMNIST-B and B = 100 for CIFAR-10. The Adam optimizer is used with the learning rate 1e-4. |