NAS-HPO-Bench-II

NAS-HPO-Bench-II是由横滨国立大学创建的一个用于联合优化卷积神经网络架构和训练超参数的基准数据集。该数据集包含4K基于单元的卷积神经网络架构在CIFAR-10数据集上训练的性能数据，共计192K配置。数据集创建过程中，模型进行了12个周期的训练，并构建了一个代理模型来预测200个周期训练后的准确性。NAS-HPO-Bench-II旨在解决深度学习中手动调整机器学习系统的困难和耗时问题，通过自动化机器学习（AutoML）优化机器学习系统的组件，特别是在神经架构搜索（NAS）领域。

NAS-HPO-Bench-II is a benchmark dataset developed by Yokohama National University for joint optimization of convolutional neural network (CNN) architectures and training hyperparameters. This dataset contains performance data of 4K cell-based convolutional neural network architectures trained on the CIFAR-10 dataset, with a total of 192K configurations. During the dataset construction process, models were trained for 12 epochs, and a surrogate model was built to predict the accuracy obtained after 200 epochs of training. NAS-HPO-Bench-II aims to address the difficulties and time-consuming issues associated with manually tuning machine learning systems in deep learning, by optimizing the components of machine learning systems via Automated Machine Learning (AutoML), especially in the field of Neural Architecture Search (NAS).

搜集汇总

背景与挑战

背景概述

NAS-HPO-Bench-II是由横滨国立大学创建的基准数据集，用于联合优化卷积神经网络架构和训练超参数，包含4K个架构在CIFAR-10上的192K个性能配置。该数据集通过自动化机器学习（AutoML）支持神经架构搜索（NAS），旨在解决手动调整机器学习系统耗时且困难的问题，提高优化效率。