Introduction

The paper introduces the NAS-Bench-201, which is extended from NAS-Bench-101.

Motivation

Previously NAS-Bench-101 and NAS-HPO-Bench are proposed. However,

1. some NAS algorithms can not be applied directly on NASBench-101.

   101 has constraints on the nodes/edges of each cell. Some NAS algorithms based on weight-sharing cannot be applied into it.

2. NAS-HPO-Bench only has 144 candidate architectures, which may be insufficient to evaluate NAS algorithms.

Contributions

In summary, the paper makes the following contributions.

1. The search space: Cell-based, each cell has 4 nodes and 5 operations, which results in 15625 cell/architectures candidates in total.
2. Each architecture is trained on three datasets (CIFAR10, CIFAR100, ImageNet-16-120 ), and the paper records their loss, accuracy, number of parameters, and FLOPs.
3. The paper benchmarked 10 NAS algorithms on the above search space.

NAS-Bench-201

Discussion

---

• ← Previous Post
• Next Post →