GCN
我们是同时将SGAS用到GCN的网络结构搜索中的。GCN网络结构由普通单元(normal cell) 组成。其搜索空间由10个运算组成:conv-1×1, MRConv, EdgeConv, GAT, SemiGCN, GIN, SAGE, RelSAGE,skip-connect, and zero operation。
SGAS在ModelNet10的训练集与测试集搜索结构,并在ModelNet40训练集和测试集上进行训练与测试,结果如表3所示:
我们也将SGAS应用到生物信息图的结点预测上。我们在PPI (protein protein intersection) 数据集的训练集与验证集搜索结构,并在PPI的训练集和测试集上进行训练与测试,结果如表4所示:
我们SGAS在GCN上的实验,超越了之前最好的模型。我们在ModelNet40以及PPI数据集上成为了新的state-of-the-art.
参考资料[1] Barret Zoph and Quoc V Le. Neuralarchitecture search with reinforcement learning. arXiv preprintarXiv:1611.01578, 2016.[2] Esteban Real, Alok Aggarwal, YanpingHuang, and Quoc V Le. Regularized evolution for image classifier architecturesearch. In Proceedings of the AAAI Conference on Artificial Intelligence,volume 33, pages 4780–4789, 2019.[3] Hanxiao Liu, Karen Simonyan, and YimingYang. Darts: Differentiable architecture search. arXiv preprintarXiv:1806.09055, 2018.[4] Maurice G Kendall. A new measure of rank correlation. Biometrika,30(1/2):81–93, 1938.