Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Single Path One-Shot Neural Architecture Search with Uniform Sampling

About

We revisit the one-shot Neural Architecture Search (NAS) paradigm and analyze its advantages over existing NAS approaches. Existing one-shot method, however, is hard to train and not yet effective on large scale datasets like ImageNet. This work propose a Single Path One-Shot model to address the challenge in the training. Our central idea is to construct a simplified supernet, where all architectures are single paths so that weight co-adaption problem is alleviated. Training is performed by uniform path sampling. All architectures (and their weights) are trained fully and equally. Comprehensive experiments verify that our approach is flexible and effective. It is easy to train and fast to search. It effortlessly supports complex search spaces (e.g., building blocks, channel, mixed-precision quantization) and different search constraints (e.g., FLOPs, latency). It is thus convenient to use for various needs. It achieves start-of-the-art performance on the large dataset ImageNet.

Zichao Guo, Xiangyu Zhang, Haoyuan Mu, Wen Heng, Zechun Liu, Yichen Wei, Jian Sun• 2019

Related benchmarks

TaskDatasetResultRank
Object DetectionCOCO 2017 (val)
AP30.7
2643
Image ClassificationImageNet-1K
Top-1 Acc74.4
1239
Image ClassificationImageNet (val)
Top-1 Acc76.6
1206
Image ClassificationImageNet 1k (test)
Top-1 Accuracy74.7
848
Image ClassificationImageNet-1k (val)
Top-1 Accuracy74.8
844
Image ClassificationImageNet-1k (val)
Top-1 Acc74.7
706
Semantic segmentationCityscapes
mIoU71.6
658
Image ClassificationImageNet
Top-1 Accuracy75
431
Image ClassificationImageNet
Top-1 Accuracy75
366
Image ClassificationImageNet (val)
Top-1 Accuracy74.3
354
Showing 10 of 29 rows

Other info

Follow for update