HA-ENAS

Code accompanying the paper.

Requirements

Python >= 3.7.13 , PyTorch >= 1.11.0 , torchvision >= 0.12.0 , torchattacks >= 3.5.1 , timm >= 0.5.4  , tqdm 

Usage

search

Collect supernet and place it to checkpoints, Then run

python search/cifar10_search.py

The log will be saved in the Job folder, where you can view the search process and results.

Best structure searched

The structure of ha-net-a1, with 94.79% accuracy on cifar10, has excellent robustness under multiple attacks.

Architecture validation

To validate the results by training from scratch, run

python validation/hanet_train.py

To validate the robustness under multiple attacks，run

python validation/transfer_attack_test.py --batch_size 32 --test_model HA_Net

• You may need to adjust the batch_size depending on your GPU memory.

• You can modify the test_model to select the networks you want to test, the optional parameters are HA_Net,resnet50,vit and mixer. you need to put the weights of these networks in the checkpoints folder before you test them.

You can download checkpoints here，and attacked dataset here.

Custom encoded HA-Net

You can add a custom encoded HA-Net after line 125 of validation/hanet/hanet.py, written as follows. And reference the network for training and testing in validation/hanet_train.py and validation/hanet_test.py

def HA_Net_custom(pretrained=False, **kwargs):
    # change your code here
    encodings = [[2, 1, 2], [0, 1, 2, 2], [0, 0, 0, 0, 0, 0], [0, 0, 0]]
    model = HaNet(encoding=encodings)
    if pretrained:
        checkpoint_path = 'checkpoints/hanet_custom.pth'
        checkpoint = torch.load(checkpoint_path)['state_dict']
        model.load_state_dict(checkpoint)
    print(model)
    return model

Citations

@article{
  title={Hybrid Architecture-based Evolutionary Robust Neural Architecture Search},
  author={Shangshang Yang, Xiangkun Sun, Ke Xu, Yuanchao Liu, Ye Tian, Member, IEEE and Xingyi Zhang, Senior Member, IEEE},
  journal={IEEE Transactions on Emerging Topics in Computational Intelligence},
  year={2024}
}