flwrlabs/fed-urbansound8k
收藏资源简介:
该数据集是UrbanSound8K的一个联邦化、非独立同分布(non-IID)的重新分区版本,用于客户端分区的音频分类实验。源数据集包含来自10个城市声音类别的标记音频片段,每个片段最长4秒。派生数据集保留了原始音频示例和元数据,但将其重新组织为50个模拟联邦客户端。每个示例包含一个`clientID`列,标识其所属的模拟客户端/分区。数据集构建过程包括加载源数据集、按`fsID`分组、创建模拟客户端、分配客户端标签偏好和大小目标,并最终将客户端本地训练和测试子集合并为全局`train`和`test`分割。该数据集适用于联邦学习、非IID音频分类、客户端分布偏移等研究。
This dataset is a federated, non-independent and identically distributed (non-IID) re-partitioned variant of the UrbanSound8K dataset, designed for audio classification experiments with client partitions. The source dataset contains labeled audio clips from 10 urban sound categories, with each clip lasting up to 4 seconds. The derived dataset retains the original audio samples and metadata, but reorganizes the data into 50 simulated federated clients. Each sample includes a `clientID` column that identifies the simulated client/partition it belongs to. The dataset construction process includes loading the source dataset, grouping samples by `fsID`, creating simulated clients, assigning client label preferences and size targets, and finally merging the local training and test subsets of all clients into the global `train` and `test` splits. This dataset is suitable for research on federated learning, non-IID audio classification, client distribution shift, and other related topics.




