Work

DSNAS: Direct Neural Architecture Search Without Parameter Retraining

Year: 2020
Type: conference-paper
Abstract: If NAS methods are solutions, what is the problem? Most existing NAS methods require two-stage parameter optimization. However, performance of the same architecture in the two stages correlates poorly. In this work, we propose a new problem definition for NAS, task-specific end-to-end, based on this observation. We argue that given a computer vision task for which a NAS method is expected, this definition can reduce the vaguely-defined NAS evaluation to i) accuracy of this task and ii) the total computation consumed to finally obtain a model with satisfying accuracy. Seeing that most existing methods do not solve this problem directly, we propose DSNAS, an efficient differentiable NAS framework that simultaneously optimizes architecture and parameters with a low-biased Monte Carlo estimate. Child networks derived from DSNAS can be deployed directly without parameter retraining. Comparing with two-stage methods, DSNAS successfully discovers networks with comparable accuracy (74.4\%) on ImageNet in 420 GPU hours, reducing the total time by more than 34\%.
Language: English
Open Access status: green
OpenAlex ID: W3034309359
FWCI: 7.702
Cites: 34
Cited by: 172
Related to: 10