Accurate and Robust Neural Architecture Search via a Flexible Supernet

This is the code implementation for the paper "Accurate and Robust Neural Architecture Search via a Flexible Supernet"

Requirements

• Python 3.7.4
• torch 1.8.0 + cu111
• torchvision 0.9.0 + cu111
• torchattacks 3.5.1

Search, Training, and Testing

Search for adversarially robust neural architectures in NAS-Bench-101/NAS-Bench-201/DARTS search spaces:

• NAS-Bench-101 search space:

python ./nb101/train_search.py

• NAS-Bench-201 search space:

python ./nb201/train_search.py

• DARTS search space:

python ./search/train_search.py

Train the derived architectures for evaluation in NAS-Bench-101/NAS-Bench-201/DARTS search spaces:

• NAS-Bench-101 search space:

python ./nb101/adv_train.py
python ./nb101/adv_test.py

• NAS-Bench-201 search space:

python ./nb201/adv_train.py
python ./nb201/adv_test.py

• DARTS search space:

Training on CIFAR-10/CIFAR-100/SVHN:

python ./train_eval/adv_train.py

Training on Tiny-ImageNet-200:

python ./train_eval/adv_train_tinyimagenet.py

Training on ImageNet-1K:

python ./train_eval/adv_train_imagenet.py

Test on CIFAR-10/CIFAR-100/SVHN/Tiny-ImageNet-200:

python ./train_eval/adv_test.py

Test on ImageNet-1K:

python ./train_eval/adv_test_imagenet.py

Test trained architectures under black-box settings:

python ./train_eval/adv_test_blackbox.py

Visualizations

Get the visualization about the derived architecture parameters:

python ./train_eval/vis_arch_parameter.py

Get the visualization regarding the filter number ratio:

python ./train_eval/vis_delta_m.py

Get the visualization in terms of the loss landscape:

python ./train_eval/vis_loss_landscape.py