- Original Pdf: pdf
- Abstract: In this paper, we present a deep convolutional neural network (CNN) which performs arbitrary resize operation on intermediate feature map resolution at stage-level. Motivated by weight sharing mechanism in neural architecture search, where a super-network is trained and sub-networks inherit the weights from the super-network, we present a novel CNN approach. We construct a spatial super-network which consists of multiple sub-networks, where each sub-network is a single scale network that obtain a unique spatial configuration, the convolutional layers are shared across all sub-networks. Such network, named as Resizable Neural Networks, are equivalent to training infinite single scale networks, but has no extra computational cost. Moreover, we present a training algorithm such that all sub-networks achieve better performance than individually trained counterparts. On large-scale ImageNet classification, we demonstrate its effectiveness on various modern network architectures such as MobileNet, ShuffleNet, and ResNet. To go even further, we present three variants of resizable networks: 1) Resizable as Architecture Search (Resizable-NAS). On ImageNet, Resizable-NAS ResNet-50 attain 0.4% higher on accuracy and 44% smaller than the baseline model. 2) Resizable as Data Augmentation (Resizable-Aug). While we use resizable networks as a data augmentation technique, it obtains superior performance on ImageNet classification, outperform AutoAugment by 1.2% with ResNet-50. 3) Adaptive Resizable Network (Resizable-Adapt). We introduce the adaptive resizable networks as dynamic networks, which further improve the performance with less computational cost via data-dependent inference.