spos

SPOS

Single Path One-Shot Neural Architecture Search with Uniform Sampling

Abstract

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.

Introduction

Supernet pre-training on ImageNet

python ./tools/mmcls/train_mmcls.py \
  configs/nas/spos/spos_supernet_shufflenetv2_8xb128_in1k.py \
  --work-dir $WORK_DIR

Search for subnet on the trained supernet

python ./tools/mmcls/search_mmcls.py \
  configs/nas/spos/spos_evolution_search_shufflenetv2_8xb2048_in1k.py \
  $STEP1_CKPT \
  --work-dir $WORK_DIR

Subnet retraining on ImageNet

python ./tools/mmcls/train_mmcls.py \
  configs/nas/spos/spos_subnet_shufflenetv2_8xb128_in1k.py \
  --work-dir $WORK_DIR \
  --cfg-options algorithm.mutable_cfg=$STEP2_SUBNET_YAML  # or modify the config directly

Results and models

Dataset	Supernet	Subnet	Params(M)	Flops(G)	Top-1 (%)	Top-5 (%)	Config	Download	Remarks
ImageNet	ShuffleNetV2	mutable	3.35	0.33	73.87	91.6	config	model | log	MMRazor searched
ImageNet	MobileNet-ProxylessGPU	mutable	5.94	0.49*	75.98	92.77	config		AngleNAS searched

Note:

1. There might be(not all the case) some small differences in our experiment in order to be consistent with other repos in OpenMMLab. For example, normalize images in data preprocessing; resize by cv2 rather than PIL in training; dropout is not used in network. Please refer to corresponding config for details.
2. For ShuffleNetV2, we retrain the subnet reported in paper with their official code, Top-1 is 73.6 and Top-5 is 91.6.
3. For AngleNAS searched MobileNet-ProxylessGPU, we obtain params and FLOPs using this script, which may be different from AngleNAS.

Citation

@inproceedings{guo2020single,
  title={Single path one-shot neural architecture search with uniform sampling},
  author={Guo, Zichao and Zhang, Xiangyu and Mu, Haoyuan and Heng, Wen and Liu, Zechun and Wei, Yichen and Sun, Jian},
  booktitle={European Conference on Computer Vision},
  pages={544--560},
  year={2020},
  organization={Springer}
}