Group Differentiable Architecture Search

Differentiable architecture search (DARTS) has received great attention due to its simplicity and efficiency. However, there are two annoying problems. One is that searched architecture of normal cell tends to be shallow. The other is skip-connect aggregation caused by the unfair competition between operations. We find that fewer operations per edge is helpful to search for deeper architectures, so we divide the operations into groups and train these groups in turn. To explore competitiveness among all the operations in the search space, candidate operations will be regrouped in each epoch. In addition, the random grouping prevents the overfitting of the super network, and consequently avoids the skip-connect aggregation. We named this method GroupDARTS and evaluated these searched architectures, achieving a state-of-the-art result of 97.68% on CIFAR10 and a top-1 accuray of 75.5% on ImageNet.