Adaout

Adaout is a practical and flexible regularization method with high generalization and interpretability.

Requirements

• python 3.6 (Anaconda version >=5.2.0 is recommended)
• torch (torch version >=1.1.0 is recommended)
• torchvision (torchvision version >=0.3.0 is recommended)
• pandas
• numpy
• NVIDIA GPU + CUDA CuDNN

Datasets

• CIFAR-10, CIFAR-100, SVHN, ImageNet and others

Getting started

• Download datasets and extract it inside data
• Train: python train.py, python train100.py or python train_svhn.py
• Evaluate:
  • Pretrained models for CIFAR-10 and CIFAR-100 are available at this link. Download and extract them in the save_model/resnet56_10 or save_model/resnet56_100 directory.
  • You should achieve the top-1 classification accuracy of 94.63% on CIFAR-10, and 74.18% on CIFAR-100 datasets.

Citation

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

@article{guo2023generalize,
  title={Generalize Deep Neural Networks With Adaptive Regularization for Classifying},
  author={Guo, Kehua and Tao, Ze and Zhang, Lingyan and Hu, Bin and Kui, Xiaoyan},
  journal={IEEE Transactions on Computational Social Systems},
  year={2023},
  publisher={IEEE}
}

For any questions, please contact Ze Tao ([email protected]).

Acknowledgement

• Thanks for Disout. We build this library based on the Disout's codebase.