Low Precision Local Training is Enough for Federated Learning

Authors: Zhiwei Li, Yiqiu LI, Binbin Lin, Zhongming Jin, Weizhong Zhang

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments across extensive benchmarks are conducted to showcase the effectiveness of our proposed method.
Researcher Affiliation Collaboration 1Fudan University 2Zhejiang University 3Alibaba Cloud Computing 4Fullong Inc.
Pseudocode Yes The details are given in Algorithm 1. Our pseudocode in Algorithm 2 depicts the process of low precision local training on the client device and high precision aggregation on the server.
Open Source Code Yes Code is released at https://github.com/digbangbang/LPT-FL.
Open Datasets Yes We conduct experiments over four commonly used datasets: Fashion Mnist [34], CIFAR10 [19], CIFAR100 [19] and CINIC10 [8].
Dataset Splits Yes In our experiment, we set the test dataset as the validation dataset.
Hardware Specification Yes All of our models are trained on s Ge Force RTX 4090.
Software Dependencies No The paper does not provide specific version numbers for the software dependencies (e.g., libraries, frameworks) used in the experiments.
Experiment Setup Yes For Fashion MNIST, CIFAR10, CINIC10 and CIFAR100, we run 200 communication rounds with local epoch set to 1. There are 80 clients in total, and the participation ratio in each round is set to 40%. We use Dirichlet distribution to simulate non-iid data distribution and set α to 0.01, 0.04, and 0.16. The local learning rate is set to 10-3 with Adam optimizer [17]. We report the last 5 round global model s average performance evaluated using the test split of the datasets. For quantization method, we adopt the Block FLoating Point Quantization with the number of bits used set to 6, 8 and 32 (without quantization). Some of the other hyperparameter settings are included in the Appendix C.