A Zero-Cost Darts Base on Multi-Step Optimization

Abstract

DARTS has achieved great result in Image classification field, the accuracy predictor and computation costs are the key of DNAS algorithm. Searching for a high-performance architecture always costs Large amount of computation. With a gradient-based bi-level optimization, DARTS using one-step optimization which makes the process available within a few GPU day, because of the one-step optimization , there exists a great gap between the architectures in search and evaluation. In this paper, we propose a zero-cost DARTS method which using multi-step optimization to address the above issues. To further reduce the computational requirements, we use the zen-score to estimate architectures in evaluation stage. Experiments on CIFAR-10 and our private data sets show that our algorithm play a certain role in solving the above problems.