OoD-ViT-NAS
收藏资源简介:
OoD-ViT-NAS是由新加坡科技设计大学创建的一个基准测试数据集,旨在评估Vision Transformer(ViT)在分布外(OoD)泛化能力上的表现。该数据集包含3000个ViT架构,涵盖了广泛的计算预算,并在8个常见的OoD数据集上进行了评估,包括ImageNet-C、ImageNet-A等。数据集通过One-Shot NAS方法构建,利用预训练的超网络权重来高效生成子网络架构。该数据集的应用领域主要集中在计算机视觉中的OoD泛化问题,旨在解决ViT架构在真实世界场景中的泛化能力不足的问题。
OoD-ViT-NAS is a benchmark dataset developed by the Singapore University of Technology and Design, aimed at evaluating the out-of-distribution (OoD) generalization performance of Vision Transformers (ViTs). This dataset includes 3000 ViT architectures spanning a wide range of computational budgets, and has been evaluated on 8 common OoD datasets such as ImageNet-C, ImageNet-A, and others. Constructed via the One-Shot NAS method, the dataset leverages pre-trained supernetwork weights to efficiently generate subnet architectures. Its application mainly focuses on OoD generalization problems in computer vision, with the goal of addressing the insufficient generalization capability of ViT architectures in real-world scenarios.




