===============================================================================
In computer vision research, the process of automating ar-chitecture engineering, Neural Architecture Search (NAS), has gainedsubstantial interest. Due to the high computational costs, most recentapproaches to NAS as well as the few available benchmarks only pro-vide limited search spaces. In this paper we propose a surrogate modelfor neural architecture performance prediction built upon Graph Neu-ral Networks (GNN). We demonstrate the effectiveness of this surrogatemodel on neural architecture performance prediction for structurally un-known architectures (i.e. zero shot prediction) by evaluating the GNNon several experiments on the NAS-Bench-101 dataset.
The NAS-Bench-101 dataset is split into 70%/20%/10% training-,test- and validation set and can be downloaded NB-101
Unzip and store all in the folder "data/".
For more information, see the paper
J.Lukasik, D.Friede, H.Stuckenschmidt, M.Keuper. Neural Architecture Performance PredictionUsing Graph Neural Networks. arXiv preprint arXiv: 2010.10024, 2020