CNASV: A Convolutional Neural Architecture Search-Train Prototype for Computer Vision Task

Abstract

Neural Architecture Search (NAS) has become more and more prevalent in the field of deep learning in the past two years. Existing works often focus on image classification, and few works recently extend NAS to another computer vision task, such as semantic image segmentation. The semantic image segmentation is essentially a dense prediction for each pixel on whole image. Therefore, we choose the same basic primitive operations to build the search space for the two computer vision task respectively. Searching good neural network architectures and then training them from scratch is a regular procedure for NAS. In this paper, we design a prototype system that deploy search module and train module to collaborate with each other. Follow the former research, we initialize over-parameterized cells architecture and then transform to the continuous relaxation of the architecture to derive the good subnetwork by gradient descent. Our system can support any differential search algorithm, such as one-shot, DARTS or ProxylessNAS. We illustrate the effectiveness of our chosen primitive operations in the image classification and ability to transfer these operations to build search space for semantic image segmentation.