Noisy Differentiable Architecture Search [BMVC 2021]

This repository includes the implementation of NoisyDARTS.

Requirements

To install requirements:

pip install -r requirements.txt

We use CIFAR-10 and ImageNet datasets, which can be downloaded by torchvision automatically. We follow the standard preprocessing for both datasets.

Searching

To perform a standard NoisyDARTS search, execute for example:

nohup python -u train_search.py --exp_name noisy_darts_a --factor_skip 0.2 --add_noise_skip --gpu 0 --seed 10 > noisy_darts_a.log 2>&1 &

Training

To train the model(s) in the paper, run for example:

nohup python -u train.py --auxiliary --cutout --save noisy_darts_a --arch noisy_darts_a --gpu 0 > noisy_darts_a.log 2>&1 &

Evaluation

To evaluate CIFAR models, run for example:

python verify.py --auxiliary --arch noisy_darts_a --model-path pretrained/noisy_darts_a.pt

More evalutaion commands can be found in scripts/run_verify.sh.

Pre-trained Models

You can download pretrained models here:

• All NoisyDARTS models can be downloaded from this link.

Results

NoisyDARTS Models Searched on CIFAR-10

Models	Search Strategy	Multiply-adds (M)	Parameters (M)	Top-1 (%)
NoisyDARTS-a	Gaussian, lambda=0.2	534	3.25	97.61
NoisyDARTS-b	Gaussian, lambda=0.1	511	3.09	97.53
NoisyDARTS-c	Uniform, lambda=0.2	539	3.33	97.40
NoisyDARTS-d	Uniform, lambda=0.1	501	3.06	97.42
NoisyDARTS-e	Gaussian, Multiplicative, std=0.1	539	3.24	97.55
NoisyDARTS-f	Gaussian, Multiplicative, std=0.2	443	2.68	97.18
NoisyDARTS-g	Gaussian, lambda=0.2, mean=0.5	549	3.32	97.49
NoisyDARTS-h	Gaussian, lambda=0.2, mean=1.0	511	3.01	97.35
NoisyDARTS-i	Gaussian, lambda=0.1, mean=0.5	495	3.07	97.28
NoisyDARTS-j	Gaussian, lambda=0.1, mean=1.0	476	2.94	97.21

Our model achieves the following performance on :

Image Classification on ImageNet

Model name	Top 1 Accuracy	Top 5 Accuracy
NoisyDARTS-A	77.9%	94.0%

Citation

@inproceedings{chu2021noisy,
  title={Noisy Differentiable Architecture Search},
  author={Chu, Xiangxiang and Zhang, Bo},
  booktitle={BMVC},
  year={2021}
}