MAC-AutoML / rethinking_performance_estimation_in_NAS

Rethinking Performance Estimation in Neural Architecture Search

There is the code of the paper Rethinking Performance Estimation in Neural Architecture Search for searching. We provide the implementations of Reinforcement Learning(RL), Evolution Algorithm(EA), Random Search(RS) and Differentiable Architecture Search(DARTS) coped with the proposed BPE method.

Two hyperparameter settings for searching, named BPE1 and BPE2 respectively, are defined in param_setting.py. BPE1 takes only 0.33 GPU hours to train a full network while BPE2 takes 0.5 GPU hours.

Reinforcement Learning

1. Train

git clone https://github.com/zhengxiawu/rethinking_performance_estimation_in_NAS.git
cd rethinking_performance_estimation_in_NAS

python run_rl.py --run_id=0 --output_path=experiment/RL --n_iters=100 --lr=1e-1 --param=BPE1/BPE2

The parameter --n_iters indicates the number of iterations, 100 for default setting and the --lr is the learning rate for agent optimization.

2. Parse the best architecture from Json file

python parse_json.py --method=RL --param=BPE1/BPE2 --run_id=0

Evolution Algorithm

1. Train

python run_evolution.py --run_id=0 --output_path=experiment/EA --n_iters=100 --pop_size=50 --param=BPE1/BPE2

The parameter --n_iters indicates the total number of iterations, while the --pop_size is the number iterations to generate populations.

2. Sampling the best architecture from supernet

• sampling 10 nets from the best supernet

python parse_json.py --method=EA --param=BPE1/BPE2 --run_id=0

• augment the sampled nets as the same way with Random Search, and find the best cell architecture

Random Search

1. Randomly generate 100 cell architectures

python random_darts_generator.py --num=100

2. Train these random architectures from scratch

• For BPE1:

python augment.py --name=RS_BPE1 --file=random_darts_architecture.txt --data_path=data/ --save_path=experiment/ --batch_size=128 --lr=0.03 --layers=6 --init_channels=8 --epochs=10 --cutout_length=0 --image_size=16

• For BPE2:

python augment.py --name=RS_BPE2 --file=random_darts_architecture.txt --data_path=data/ --save_path=experiment/ --batch_size=128 --lr=0.03 --layers=16 --init_channels=16 --epochs=30 --cutout_length=0 --image_size=16

Differentiable Architecture Search

• For BPE1:

python search.py --name=DARTS_BPE1 --batch_size=128 --w_lr=0.03 --layers=6 --init_channels=8 --epochs=10 --cutout_length=0 --image_size=16

• For BPE2:

python search.py --name=DARTS_BPE2 --batch_size=128 --w_lr=0.03 --layers=16 --init_channels=16 --epochs=30 --cutout_length=0 --image_size=16

Search Trails

We also release the examples we sampled in our paper in Search_Hyperparameters_Results.xlsx

Reference Project

https://github.com/khanrc/pt.darts

https://github.com/automl/fanova

https://github.com/automl/nas_benchmarks

Cite this work

@INPROCEEDINGS {zheng_rethinking,
  author = {X. Zheng and R. Ji and Q. Wang and Q. Ye and Z. Li and Y. Tian and Q. Tian},
  booktitle = {2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  title = {Rethinking Performance Estimation in Neural Architecture Search},
  year = {2020},
  month = {jun}
}