[NeurIPS-23] Pytorch Implementation of Revisiting Visual Model Robustness: A Frequency Long-Tailed Distribution View

• Paper link

• Supplementary material

• If you have any questions about the paper or code, please contact us

Environments

• pytorch
• numpy
• autoattack

Visualization of analytical experiments

• The Visualize.ipynb file contains the model's under-fitting Behavior visualization and sensitivity analysis experimental code for high-frequency components.
• We provide the data files required for visualization in the link, including input gradients under standard training and sampling results of the model loss space.

How to use Balance Spectrum Sampling（BaSS）strategy

We provide two methods of using BaSS in the data preprocessing stage:

• for CIFAR10 or CIFAR100 dataset

train_set = torchvision.datasets.CIFAR10(root='/data/cifar', train=True, download=True)
test_set = torchvision.datasets.CIFAR10(root='/data/cifar', train=False, download=True)

print("CIFAR-B dataset :)")

train_set.data = get_BaSS_dataset(train_set.data.copy())
test_set.data = get_BaSS_dataset(test_set.data.copy())

We provide the algorithm details of the get_BaSS_dataset() function in the BaSS_cifar.py file

• for ImageNet dataset

TRAIN_TRANSFORMS_IMAGENET = transforms.Compose([
        transforms.Resize((128,128)),

        ## applying BaSS 
        FreqLog(),
        ##

        transforms.RandomHorizontalFlip(),
        transforms.ColorJitter(
            brightness=0.1,
            contrast=0.1,
            saturation=0.1
        ),
        transforms.ToTensor(),
        Lighting(0.05, IMAGENET_PCA['eigval'], 
                      IMAGENET_PCA['eigvec'])
    ])

TEST_TRANSFORMS_IMAGENET = transforms.Compose([
        transforms.Resize((128,128)),

        ## applying BaSS
        FreqLog(),
        ##

        transforms.ToTensor(),
    ])

We provide the algorithm details of the FreqLog() function in the BaSS_imagegnet.py file

Examples of BaSS Working in Conjunction with Adversarial Training

Our adversarial training framework is based on Bag of Tricks for Adversarial Training (ICLR 2021, code) and MART(ICLR 2020, code).

Training

We only modified the data processing part based on the original code framework. For detailed code description, please refer to the readme file in the original repo. In order to train quickly, we edited the easy_train.py file to include the training script.

cd Bag-of-Tricks-for-AT
python easy_train.py

cd MART
python easy_train.py

We set the --f-train switch in the training script whether to use BaSS, for example

## standard training
script1 = 'CUDA_VISIBLE_DEVICES=0 python train.py --model resnet --dataset cifar10'

## training with BaSS
script2 = 'CUDA_VISIBLE_DEVICES=1 python train.py --model resnet --dataset cifar10 --f-train'

Evaluation

Our evaluation results are based on Clean acc, PGD(CE loss,CW loss),Autoattack.

cd Bag-of-Tricks-for-AT
python eval_cifar.py --out-dir 'path_to_the_model' --ATmethods 'TRADES' --f-test

cd MART
python test.py --dataset cifar10 --model resnet --f-test