TPUGRAPHS

TPUGRAPHS数据集是由Google研究团队创建的，专注于大型张量计算图的性能预测。该数据集包含3100多万个图和配置对，每个数据样本包括计算图、编译配置和执行时间。这些图来自开源机器学习程序，涵盖了多种流行的模型架构，如ResNet、EfficientNet等。数据集的创建旨在解决机器学习编译器在优化决策中的挑战，特别是在处理大规模图时的可扩展性、训练效率和模型质量问题。TPUGRAPHS数据集的应用领域包括加速机器学习编译器的优化过程，提高模型在实际生产环境中的性能。

TPUGRAPHS dataset is created by the Google Research team, focusing on performance prediction of large-scale tensor computation graphs. This dataset contains over 31 million graph-and-configuration pairs, with each data sample consisting of a computation graph, a compilation configuration, and execution time. These graphs are sourced from open-source machine learning programs, covering a variety of popular model architectures such as ResNet, EfficientNet, etc. The dataset is developed to address the challenges in optimization decisions of machine learning compilers, particularly issues related to scalability, training efficiency, and model quality when handling large-scale graphs. Application scenarios of the TPUGRAPHS dataset include accelerating the optimization process of machine learning compilers and improving the performance of models in actual production environments.