SCARLET-NAS: Bridging the Gap between Stability and Scalability in Weight-sharing Neural Architecture Search

16 Aug 2019  ·  Xiangxiang Chu, Bo Zhang, Qingyuan Li, Ruijun Xu, Xudong Li ·

To discover powerful yet compact models is an important goal of neural architecture search. Previous two-stage one-shot approaches are limited by search space with a fixed depth. It seems handy to include an additional skip connection in the search space to make depths variable. However, it creates a large range of perturbation during supernet training and it has difficulty giving a confident ranking for subnetworks. In this paper, we discover that skip connections bring about significant feature inconsistency compared with other operations, which potentially degrades the supernet performance. Based on this observation, we tackle the problem by imposing an equivariant learnable stabilizer to homogenize such disparities. Experiments show that our proposed stabilizer helps to improve the supernet's convergence as well as ranking performance. With an evolutionary search backend that incorporates the stabilized supernet as an evaluator, we derive a family of state-of-the-art architectures, the SCARLET series of several depths, especially SCARLET-A obtains 76.9% top-1 accuracy on ImageNet. Code is available at https://github.com/xiaomi-automl/ScarletNAS.

PDF Abstract
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Neural Architecture Search ImageNet SCARLET-A Top-1 Error Rate 23.1 # 76
Accuracy 76.9 # 61
Params 6.7M # 15
MACs 365M # 107
Image Classification ImageNet SCARLET-A4 Top 1 Accuracy 82.3% # 541
Number of params 27.8M # 638
Hardware Burden 12G # 1
Operations per network pass 0.42G # 1
GFLOPs 8.4 # 289
Image Classification ImageNet SCARLET-C Top 1 Accuracy 75.6% # 920
Number of params 6M # 446
GFLOPs 0.560 # 58
Image Classification ImageNet SCARLET-B Top 1 Accuracy 76.3% # 895
Number of params 6.5M # 453
GFLOPs 0.658 # 78
Image Classification ImageNet SCARLET-A Top 1 Accuracy 76.9% # 874
Number of params 6.7M # 458
GFLOPs 0.730 # 90
Neural Architecture Search ImageNet SCARLET-C Top-1 Error Rate 24.4 # 103
Accuracy 75.6 # 82
Params 6.0M # 23
MACs 280M # 87
Neural Architecture Search ImageNet SCARLET-B Top-1 Error Rate 23.7 # 86
Accuracy 76.3 # 70
Params 6.5M # 16
MACs 329M # 98

Methods