Fair DARTS: Eliminating Unfair Advantages in Differentiable Architecture Search

Differential Architecture Search (DARTS) is now a widely disseminated weight-sharing neural architecture search method. However, there are two fundamental weaknesses remain untackled. First, we observe that the well-known aggregation of skip connections during optimization is caused by an unfair advantage in an exclusive competition. Second, there is a non-negligible incongruence when discretizing continuous architectural weights to a one-hot representation. Because of these two reasons, DARTS delivers a biased solution that might not even be suboptimal. In this paper, we present a novel approach to curing both frailties. Specifically, as unfair advantages in a pure exclusive competition easily induce a monopoly, we relax the choice of operations to be collaborative, where we let each operation have an equal opportunity to develop its strength. We thus call our method Fair DARTS. Moreover, we propose a zero-one loss to directly reduce the discretization gap. Experiments are performed on two mainstream search spaces, in which we achieve new state-of-the-art networks on ImageNet.

We Are Still Hiring (both interns and full-time)!

Hi folks! We are AutoML Team from Xiaomi AI Lab, based in Beijing, China. There are few open positions (internships included), welcome global applications from students, new graduates and professionals skilled in NAS and Deep Learning (Vision, Speech, NLP etc.)!

Please send your resume to zhangbo11@xiaomi.com
人工智能算法/软件工程师（含实习生）职位，简历请发送至 zhangbo11@xiaomi.com
QQ群交流：小米 AutoML 交流反馈, 群号：702473319 (加群请填写“神经网络架构搜索”的英文简称)

User Guide

Prerequisites

Python 3

pip install -r requirements.txt

The fairdarts folder includes our search, train and evaluation code. The darts folder consists of random and noise experiments on the original DARTS.

Run Search

python train_search.py --aux_loss_weight 10 --learning_rate 0.005 --batch_size 128 --parse_method threshold_sparse --save 'EXP-lr_0005_alw_10'

Default batch-size is 128

Single Model Training

python train.py --auxiliary --cutout --arch FairDARTS_a --parse_method threshold --batch_size 128 --epoch 600

Single Model Evaluation

python evaluate_model.py  --arch FairDARTS_b --model_path ../best_model/FairDARTS-b.tar --parse_method threshold

Searched Architectures by FairDARTS

Note that we select architecture by barring with threshold σ, and |edge| <= 2 per node.

FairDARTS_a:

FairDARTS_b

FairDARTS_c

FairDARTS_d

FairDARTS_e

FairDARTS_f

FairDARTS_g

The isolated nodes (in gray) are ignored after parsing the genotypes.

Evaluation Results on CIFAR-10

Performance Stability

We run FairDARTS 7 times, all searched architectures have close performance.

Model	Flops	Params	Performance
FairDARTS_a	373M	2.83M	97.46
FairDARTS_b	536M	3.88M	97.49
FairDARTS_c	400M	2.59M	97.50
FairDARTS_d	532M	3.84M	97.51
FairDARTS_e	414M	3.12M	97.47
FairDARTS_f	497M	3.62M	97.35
FairDARTS_g	453M	3.38M	97.46
mean,var	~457.85M	~3.32M	97.46±0.049

Note: We remove batch normalization for FLOPs' calculation in thop package. This is to follow status quo treamtment.

Comparison with Other State-of-the-art Results (CIFAR-10)

Model	FLOPs	Params	Batch size	lr	DP	Optimizer	Performance
FairDARTS-a	373M	2.83	96	0.025	0.2	SGD+CosineAnnealingLR	97.46
FairDARTS-b	536M	3.88	96	0.025	0.2	SGD+CosineAnnealingLR	97.49
DARTS_V2	522M	3.36	96	0.025	0.2	SGD+CosineAnnealingLR	96.94*
PC-DARTS	558M	3.63	96	0.025	0.2	SGD+CosineAnnealingLR	97.31*
PDARTS	532M	3.43	96	0.025	0.2	SGD+CosineAnnealingLR	97.53*

*: Results obtained by training their published code.

Citation

@article{chu2019fairdarts,
    title={{Fair DARTS: Eliminating Unfair Advantages in Differentiable Architecture Search}},
    author={Chu, Xiangxiang and Zhou, Tianbao and Zhang, Bo and Li, Jixiang},
    journal={arXiv preprint arXiv:1911.12126},
    url={https://arxiv.org/abs/1911.12126.pdf},
    year={2019}
}

hyqyoung/FairDARTS