FedScale

FedScale是一个包含20个真实世界数据集的联邦学习基准套件，旨在评估和优化联邦学习在不同任务和规模下的性能。这些数据集覆盖了从图像分类、目标检测到语言建模和语音识别等多个关键任务。每个数据集都配备了统一的评估协议，使用真实世界的数据分割和评估指标。FedScale还提供了一个可扩展的运行时环境，支持高效实现联邦学习算法，并在多样化的硬件和软件后端上进行大规模部署和评估。该数据集的应用领域广泛，旨在解决联邦学习中的性能瓶颈和优化问题。

FedScale is a federated learning benchmark suite containing 20 real-world datasets, developed to evaluate and optimize the performance of federated learning across diverse tasks and scales. These datasets cover multiple critical tasks ranging from image classification, object detection to language modeling and speech recognition. Each dataset is equipped with a unified evaluation protocol that utilizes real-world data splits and standard evaluation metrics. FedScale also provides a scalable runtime environment that supports efficient implementation of federated learning algorithms, as well as large-scale deployment and evaluation across heterogeneous hardware and software backends. This suite has wide application scenarios and aims to address performance bottlenecks and optimization problems in federated learning.