Neural Architecture Search with Random Labels(RLNAS)

Introduction

This project provides an implementation for Neural Architecture Search with Random Labels (CVPR 2021 poster) on Pytorch. Experiments are evaluated on multiple datasets (NAS-Bench-201 and ImageNet) and multiple search spaces (DARTS-like and MobileNet-like). RLNAS achieves comparable or even better results compared with state-of-the-art NAS methods such as PC-DARTS, Single Path One-Shot, even though the counterparts utilize full ground truth labels for searching. We hope our finding could inspire new understandings on the essential of NAS.

Requirements

• Pytorch 1.4
• Python3.5+

Search results

1.Results in NAS-Benchmark-201 search space

2.Results in DARTS searh space

Architeture visualization

1) Architecture searched on CIFAR-10

• RLDARTS = Genotype(
  normal=[
  ('sep_conv_5x5', 0), ('sep_conv_3x3', 1),
  ('dil_conv_3x3', 0), ('sep_conv_5x5', 2),
  ('sep_conv_3x3', 0), ('dil_conv_5x5', 3),
  ('dil_conv_5x5', 1), ('dil_conv_3x3', 2)],
  normal_concat=[2, 3, 4, 5],
  reduce=[
  ('sep_conv_5x5', 0), ('dil_conv_3x3', 1),
  ('sep_conv_3x3', 0), ('sep_conv_5x5', 2),
  ('dil_conv_3x3', 1), ('sep_conv_3x3', 3),
  ('max_pool_3x3', 1), ('sep_conv_5x5', 2,)],
  reduce_concat=[2, 3, 4, 5])

• Normal cell:

• Reduction cell:

2) Architecture searched on ImageNet-1k without FLOPs constrain

• RLDARTS = Genotype( normal=[
  ('sep_conv_3x3', 0), ('sep_conv_3x3', 1),
  ('sep_conv_3x3', 1), ('sep_conv_3x3', 2),
  ('sep_conv_3x3', 0), ('sep_conv_5x5', 1),
  ('sep_conv_3x3', 0), ('sep_conv_3x3', 1)],
  normal_concat=[2, 3, 4, 5],
  reduce=[
  ('sep_conv_3x3', 0), ('sep_conv_3x3', 1),
  ('sep_conv_5x5', 0), ('sep_conv_3x3', 2),
  ('sep_conv_5x5', 0), ('sep_conv_5x5', 2),
  ('sep_conv_3x3', 2), ('sep_conv_3x3', 4)],
  reduce_concat=[2, 3, 4, 5])

• Normal cell:

• Reduction cell:

3) Architecture searched on ImageNet-1k with 600M FLOPs constrain

• RLDARTS = Genotype(
  normal=[
  ('sep_conv_3x3', 0), ('sep_conv_3x3', 1),
  ('skip_connect', 1), ('sep_conv_3x3', 2),
  ('sep_conv_3x3', 1), ('sep_conv_3x3', 2),
  ('skip_connect', 0), ('sep_conv_3x3', 4)],
  normal_concat=[2, 3, 4, 5],
  reduce=[ ('sep_conv_3x3', 0), ('max_pool_3x3', 1),
  ('sep_conv_3x3', 0), ('skip_connect', 1),
  ('sep_conv_3x3', 0), ('dil_conv_3x3', 1),
  ('skip_connect', 0), ('sep_conv_3x3', 1)],
  reduce_concat=[2, 3, 4, 5])

• Normal cell:

• Reduction cell:

3.Results in MobileNet search space

The MobileNet-like search space proposed in ProxylessNAS is adopted in this paper. The SuperNet contains 21 choice blocks and each block has 7 alternatives:6 MobileNet blocks (combination of kernel size {3,5,7} and expand ratio {3,6}) and ’skip-connect’.

Architeture visualization

Usage

• RLNAS in NAS-Benchmark-201

1)enter the work directory

cd nas_bench_201

2)train supernet with random labels

bash ./scripts-search/algos/train_supernet.sh cifar10 0 1

3)evolution search with angle

bash ./scripts-search/algos/evolution_search_with_angle.sh cifar10 0 1

4)calculate correlation

bash ./scripts-search/algos/cal_correlation.sh cifar10 0 1

• RLNAS in DARTS search space

1)enter the work directory

cd darts_search_space

search architecture on CIFAR-10

cd cifar10/rlnas/

or search architecture on ImageNet

cd imagenet/rlnas/

2)train supernet with random labels

cd train_supernet
bash run_train.sh

3)evolution search with angle

cd evolution_search
cp ../train_supernet/models/checkpoint_epoch_50.pth.tar ./model_and_data/
cp ../train_supernet/models/checkpoint_epoch_0.pth.tar ./model_and_data/
bash run_server.sh
bash run_test.sh

4)architeture evaluation

cd retrain_architetcure

add searched architecture to genotypes.py

bash run_retrain.sh

• RLNAS in MobileNet search space

The conduct commands are almost the same steps like RLNAS in DARTS search space, excepth that you need run 'bash run_generate_flops_lookup_table.sh' before evolution search.

Note: setup a server for the distributed search

tmux new -s mq_server
sudo apt update
sudo apt install rabbitmq-server
sudo service rabbitmq-server start
sudo rabbitmqctl add_user test test
sudo rabbitmqctl set_permissions -p / test '.*' '.*' '.*'

Before search, please modify host and username in the config file evolution_search/config.py.

Citation

If you find that this project helps your research, please consider citing some of the following papers:

@article{zhang2021neural,
  title={Neural Architecture Search with Random Labels},
  author={Zhang, Xuanyang and Hou, Pengfei and Zhang, Xiangyu and Sun, Jian},
  booktitle={Proceedings of the IEEE conference on computer vision and pattern recognition},
  year={2021}
}

@inproceedings{hu2020angle,
  title={Angle-based search space shrinking for neural architecture search},
  author={Hu, Yiming and Liang, Yuding and Guo, Zichao and Wan, Ruosi and Zhang, Xiangyu and Wei, Yichen and Gu, Qingyi and Sun, Jian},
  booktitle={European Conference on Computer Vision},
  pages={119--134},
  year={2020},
  organization={Springer}
}

@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}
}