Stream

该数据集被称为Stream数据集，它由71个公开可用的图像分类数据集拼接而成，形成了一系列任务流。这样的设计旨在评估在面对显著的领域变化和不同任务难度时，持续学习方法的表现。每个任务包含不同数量的类别，并且该数据集使用了从预训练的CLIP模型中提取的特征向量进行评估。在规模上，该数据集包含了6,770,722张训练图片和743,977张验证图片，涵盖了2866个类别，其任务类型为图像分类。This dataset is termed the Stream Dataset. It is constructed by concatenating 71 publicly available image classification datasets to form a series of task streams. This design aims to evaluate the performance of continual learning methods when confronted with substantial domain shifts and varying task difficulties. Each task comprises a distinct number of classes, and evaluation of this dataset is conducted using feature vectors extracted from a pre-trained CLIP model. In terms of scale, the dataset includes 6,770,722 training images and 743,977 validation images, covering 2,866 classes, with the task type being image classification.