Skip to content

Latest commit

 

History

History
67 lines (50 loc) · 7.61 KB

README.md

File metadata and controls

67 lines (50 loc) · 7.61 KB
Raw

Cyclic Differentiable Architecture Search

This is an official implementation of CDARTS

In this work, we propose new joint optimization objectives and a novel Cyclic Differentiable ARchiTecture Search framework, dubbed CDARTS. Considering the structure difference, CDARTS builds a cyclic feedback mechanism between the search and evaluation networks with introspective distillation. First, the search network generates an initial architecture for evaluation, and the weights of the evaluation network are optimized. Second, the architecture weights in the search network are further optimized by the label supervision in classification, as well as the regularization from the evaluation network through feature distillation. Repeating the above cycle results in a joint optimization of the search and evaluation networks and thus enables the evolution of the architecture to fit the final evaluation network.

CDARTS overview

Model Zoo

For evaluation, we provide the checkpoints and configs of our models in Google Drive.

After downloading the models, you can do the evaluation following the description in SETUP.md.

Model download links:

DARTS Search Space

CIFAR10

Top-1 Acc. % 97.60 97.45 97.52 97.53 97.54 97.77
Cell Download link Cell-1 Cell-2 Cell-3 Cell-4 Cell-5 Cell-6

ImageNet

Top-1 Acc. % 75.90 75.93 76.40 76.60 76.44
Cell Download link Cell-1 Cell-2 Cell-3 Cell-4 Cell-5

NATS-Bench

Model CIFAR10 Validation CIFAR10 Test CIFAR100 Validation CIFAR100 Test ImageNet-16-120 Validation ImageNet-16-120 Test Download link
Cell1 91.50% 94.37% 73.31% 73.09% 45.59% 46.33% Cell, Log
Cell2 91.37% 94.09% 72.64% 72.57% 45.46% 45.63% Cell, Log
Cell3 90.51% 93.62% 70.43 70.10% 44.23% 44.57% Cell, Log

Chain-structured Search Space

Model Params. Flops Top-1 Acc. % Download link
CDARTS-a 7.0M 294M 77.4 Model, Config, Log
CDARTS-b 6.4M 394M 78.2 Model, Config, Log

Object Detection

Backbone Input Size Params. Flops AP AP_50 AP_75 AP_S AP_M AP_L Download link
CDARTS-a 1280x800 6.0G 7.0M 35.2 55.5 37.5 19.8 38.7 47.5 Model, Config, Log
CDARTS-b 1280x800 8.1G 6.4M 36.2 56.7 38.3 20.9 39.8 48.5 Model, Config, Log

Semantic Segmentation

Dataset Encoder Input Size Params. Flops mIoU % Download link
Cityscapes CDARTS-b 1024x2048 5.9M 20.7G 78.1 Model,Config,Log
ADE20K CDARTS-b 640x640 2.7M 5.9G 40.4 Model,Config,Log

Bibtex

If this repo is helpful for you, please consider to cite it. Thank you! :)

@article{CDARTS,
  title={Cyclic Differentiable Architecture Search},
  author={Yu, Hongyuan and Peng, Houwen and Huang, Yan and Fu, Jianlong and Du, Hao and Wang, Liang and Ling, Haibin},
  booktitle={IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI)},
  year={2022}
}