Progressive Differentiable Architecture Search: Bridging the Depth Gap between Search and Evaluation

TL;DR

This paper introduces a progressive differentiable architecture search method that gradually increases architecture depth during training, reducing search time and improving transferability and accuracy of neural architectures.

Contribution

It proposes a novel progressive search algorithm that addresses depth gap issues, enabling efficient and stable architecture search with state-of-the-art results.

Findings

Achieves state-of-the-art performance on CIFAR and ImageNet datasets.

Reduces search time to approximately 7 hours on a single GPU.

Improves transferability of searched architectures across datasets.

Abstract

Recently, differentiable search methods have made major progress in reducing the computational costs of neural architecture search. However, these approaches often report lower accuracy in evaluating the searched architecture or transferring it to another dataset. This is arguably due to the large gap between the architecture depths in search and evaluation scenarios. In this paper, we present an efficient algorithm which allows the depth of searched architectures to grow gradually during the training procedure. This brings two issues, namely, heavier computational overheads and weaker search stability, which we solve using search space approximation and regularization, respectively. With a significantly reduced search time (~7 hours on a single GPU), our approach achieves state-of-the-art performance on both the proxy dataset (CIFAR10 or CIFAR100) and the target dataset (ImageNet). Code is available at https://github.com/chenxin061/pdarts.