• Paper link

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

• Starting with standard traning by easy_train.py

# an example of training on cifar10 with resnet18
python easy_train.py --dataset cifar10 --arch resnet18 --gpu 0

• Obtaining evolution of gradient spectrum and visualization by grad.py

python grad.py --dataset cifar10 --arch resnet18 --gpu 0

Convolutional Density Enhancement Strategy (CDES)

Hypothesize: Spectral density can serve as an explanation for frequency bias in image classification tasks. We propose a framework called Convolutional Density Enhancement Strategy (CDES) to modify the spectral density of natural images.

• Train toy kernels

Since simply adding noise is difficult to modify the spectral density as we expect, we first propose to perform convolution operations on original images with a set of trainable convolution filters.

cd SpectralDensity
python train_toy_models.py --number 16 --kernel_size 7 --cuda_index 0 --weight_decay 1e-4 --momentum 0.9 --epochs 20 --lr 0.01 --step_lr 5 --step_lr_gamma 0.1 --exp_name SCR --batch_size 100 --workers 4

--number (int): number of convolutional kernels, default 16.

--kernel_size (int): size of convolutional kernels, default 7.

--exp_name (str): The type of generate dataset. default [SCR, WCR].

• Generate toy datasets

cd SpectralDensity
python generate_toy_dataset.py --exp_name SCR --cuda_index 0

--cuda_index (int): your gpu ids, default 0.

--exp_name (str): The type of generate dataset, default [SCR, WCR].

• Visualization of SCR-dataset(take class 0 as an example)

• Visualization of WCR-dataset

Process of building HARS-dataset

Hypothesize: We hypothesize that the model only pick up the frequency components with the property of class consistency. For example, if we break up the class consistency on LFC, model will pay more attention to employing HFC. In order to verify this hypothesis, we construct another variant dataset of CIFAR-10 called HARS-dataset

• Creating the new dataset called HARS-dataset

cd tools
python generate_freq_dataset.py

We use r=12 in the code to control the mixing radius of low frequency and high frequency of different types of images.

• Training model with HARS-dataset

python easy_train.py --dataset cifar10 --arch resnet18 --gpu 0 --r 12

• Evaluation of model

python test.py --dataset cifar10 --arch resnet18 --freq low --r 4

--freq (str): The frequency component of image, low or high frequencies.

--r (int): The frequency radius of spectrum.

The code refers to https://github.com/MadryLab/robustness and https://github.com/cc-hpc-itwm/UpConv

If you find this repo useful for your research, please consider citing the paper

@inproceedings{ijcai2022p101,
  title     = {Investigating and Explaining the Frequency Bias in Image Classification},
  author    = {Lin, Zhiyu and Gao, Yifei and Sang, Jitao},
  booktitle = {Proceedings of the Thirty-First International Joint Conference on
               Artificial Intelligence, {IJCAI-22}},
  publisher = {International Joint Conferences on Artificial Intelligence Organization},
  editor    = {Lud De Raedt},
  pages     = {717--723},
  year      = {2022},
  month     = {7},
  note      = {Main Track},
  doi       = {10.24963/ijcai.2022/101},
  url       = {https://doi.org/10.24963/ijcai.2022/101},
}