theot1 / dign

Diverse Gaussian Noise Consistency Regularization for Robustness and Uncertainty Calibration

Diverse Gaussian Noise Consistency Regularization for Robustness and Uncertainty Calibration

GitHub repo for IJCNN 2023 paper

[pdf] [arxiv]

Directories

• models: contains saved model checkpoints
• results: contains .pkl files generated from testing
• CIFAR-10-C: .npy files for each corruption
• CIFAR-100-C: .npy files for each corruption
• Tiny-ImageNet-C: folders for each corruption
  • 'brightness'
  • 1: severity 1
  • class1

Training

To train CIFAR-10 ResNet-18 model using Standard training, batch size 512:

python train.py cifar10 resnet18 512 standard

To train CIFAR-10 ResNet-18 model using DiGN training, batch size 512:

python train.py cifar10 resnet18 512 dign_gn

Other baselines are available in train.py, including:

• AT
• TRADES
• RSE
• DeepAugment
• AugMix
• AugMax
• DiGN w.o. CR

Robustness Evaluation

To evaluate robustness under corruptions for a CIFAR-10 ResNet-18 model:

python test-rob-c.py cifar10 resnet18 512 [Boolean_test_time_ensemble] [model_name] [Boolean_noise_only_eval]

To evaluate a model only on digital noise corruptions (mCA-N), assuming the model in models directory saved as 'RN18_cifar10_DIGN_model_X', use:

python test-rob-c.py cifar10 resnet18 512 False DIGN_model_X True

and for evaluating on all common corruptions (mCA), use instead:

python test-rob-c.py cifar10 resnet18 512 False DIGN_model_X False

Uncertainty Calibration Evaluation

To evaluate uncertainty calibration under corruptions for a CIFAR-10 ResNet-18 model:

python test-cal-c.py cifar10 resnet18 512 [Boolean_test_time_ensemble] [model_name] [Boolean_noise_only_eval]

To evaluate a model only on digital noise corruptions (RMSE-N), assuming the model in models directory saved as 'RN18_cifar10_DIGN_model_X', use:

python test-cal-c.py cifar10 resnet18 512 False DIGN_model_X True

and for evaluating on all common corruptions (RMSE), use instead:

python test-cal-c.py cifar10 resnet18 512 False DIGN_model_X False

Citation

For citing this paper or code, please use the following:

@inproceedings{DiGN,
  title={Diverse Gaussian Noise Consistency Regularization for Robustness and Uncertainty Calibration},
  author={Theodoros Tsiligkaridis and Athanasios Tsiligkaridis},
  booktitle={Proceedings of the International Joint Conference on Neural Networks, IJCNN},
  month={June},
  year={2023}
}