Blox

Blox数据集是由三星人工智能中心剑桥创建的一个大型宏观神经架构搜索（NAS）基准。该数据集包含91,125个在CIFAR-100数据集上训练的独特模型，旨在为NAS算法提供一个系统的性能研究平台。Blox数据集不仅包括模型的训练结果，还包括在多种硬件平台上的运行时测量，以支持对NAS算法的全面评估。此数据集特别关注于块的多样性，允许模型中的每个块具有不同的架构，从而推动了宏观搜索空间的研究。通过Blox，研究人员可以更深入地理解不同NAS算法的性能，并为设计更高效的宏观搜索算法提供实证分析。

The Blox dataset is a large-scale macro neural architecture search (NAS) benchmark developed by Samsung AI Center, Cambridge. This dataset comprises 91,125 unique models trained on the CIFAR-100 dataset, and is designed to serve as a systematic platform for performance research on NAS algorithms. The Blox dataset encompasses not only the training outcomes of these models, but also runtime measurements across diverse hardware platforms, enabling comprehensive evaluations of NAS algorithms. This dataset specifically focuses on block diversity, allowing each block within a model to feature distinct architectures, thereby advancing research on macro search spaces. Leveraging the Blox dataset, researchers can gain deeper insights into the performance of various NAS algorithms, and derive empirical analyses to facilitate the design of more efficient macro search algorithms.