TransNAS-Bench-101

TransNAS-Bench-101 是一个神经架构搜索 (NAS) 基准数据集，包含七个任务的网络性能，涵盖分类、回归、像素级预测和自我监督任务。这种多样性提供了在任务之间转移 NAS 方法的机会，并允许发展更复杂的转移方案。我们探索了两种根本不同类型的搜索空间：单元级搜索空间和宏观级搜索空间。通过对 7 个任务的 7,352 个主干进行评估，提供了 51,464 个具有详细训练信息的训练模型。通过 TransNAS-Bench-101，我们希望鼓励出现卓越的 NAS 算法，将跨任务搜索效率和通用性提高到一个新的水平。

TransNAS-Bench-101 is a neural architecture search (NAS) benchmark dataset that contains network performance metrics across seven tasks, covering classification, regression, pixel-level prediction, and self-supervised learning tasks. This diversity provides opportunities for transferring NAS methods across tasks and enables the development of more sophisticated transfer strategies. We investigate two fundamentally distinct types of search spaces: cell-level search spaces and macro-level search spaces. By evaluating 7,352 backbones across the seven tasks, this benchmark provides 51,464 trained models with detailed training information. With TransNAS-Bench-101, we intend to encourage the emergence of exceptional NAS algorithms that elevate cross-task search efficiency and generality to unprecedented levels.

搜集汇总

背景与挑战

背景概述

TransNAS-Bench-101是一个神经架构搜索基准数据集，涵盖七个任务的网络性能评估，包括分类、回归和像素级预测等。它提供了两种搜索空间和基于大量主干评估的模型训练信息，旨在推动跨任务NAS算法的通用性和效率提升。