Skip to content

valeman/calibration-study

 
 

Repository files navigation

Calibration Dataset

This dataset is based on NATS-Bench and is used for calibration. There are 6 files in this folder, which are:

  • cifar10_sss.csv: the calibration dataset for CIFAR-10 with 32,768 architectures in SSS search space.
  • cifar10_tss.csv: the calibration dataset for CIFAR-10 with 15,624 architectures in TSS search space.
  • cifar100_sss.csv: the calibration dataset for CIFAR-100 with 32,768 architectures in SSS search space.
  • cifar100_tss.csv: the calibration dataset for CIFAR-100 with 15,624 architectures in TSS search space.
  • ImageNet_sss.csv: the calibration dataset for ImageNet16-120 with 32,768 architectures in SSS search space.
  • ImageNet_tss.csv: the calibration dataset for ImageNet16-120 with 15,624 architectures in TSS search space.

The meaning of each column is as follows, note that the robustness data is from Neural Architecture Design and Robustness: A Dataset:

  • config: the architecture index.
  • acc: the accuracy of the architecture at epoch 200.
  • correct_confidence: the mean confidence of the correct prediction.
  • incorrect_confidence: the mean confidence of the incorrect prediction.
  • arch: the architecture string.
  • pre_ECE_5: the pre-temperature-scaling ECE with bin size = 5.
  • post_ECE_5: the post-temperature-scaling ECE with bin size = 5.
  • pre_ECE_10: the pre-temperature-scaling ECE with bin size = 10.
  • post_ECE_10: the post-temperature-scaling ECE with bin size = 10.
  • pre_ECE_15: the pre-temperature-scaling ECE with bin size = 15.
  • post_ECE_15: the post-temperature-scaling ECE with bin size = 15.
  • pre_ECE_20: the pre-temperature-scaling ECE with bin size = 20.
  • post_ECE_20: the post-temperature-scaling ECE with bin size = 20.
  • pre_ECE_25: the pre-temperature-scaling ECE with bin size = 25.
  • post_ECE_25: the post-temperature-scaling ECE with bin size = 25.
  • pre_ECE_50: the pre-temperature-scaling ECE with bin size = 50.
  • post_ECE_50: the post-temperature-scaling ECE with bin size = 50.
  • pre_ECE_100: the pre-temperature-scaling ECE with bin size = 100.
  • post_ECE_100: the post-temperature-scaling ECE with bin size = 100.
  • pre_ECE_200: the pre-temperature-scaling ECE with bin size = 200.
  • post_ECE_200: the post-temperature-scaling ECE with bin size = 200.
  • pre_ECE_500: the pre-temperature-scaling ECE with bin size = 500.
  • post_ECE_500: the post-temperature-scaling ECE with bin size = 500.
  • pre_ECE_em_5: the pre-temperature-scaling ECE with equal mass bin size = 5.
  • post_ECE_em_5: the post-temperature-scaling ECE with equal mass bin size = 5.
  • pre_ECE_em_10: the pre-temperature-scaling ECE with equal mass bin size = 10.
  • post_ECE_em_10: the post-temperature-scaling ECE with equal mass bin size = 10.
  • pre_ECE_em_15: the pre-temperature-scaling ECE with equal mass bin size = 15.
  • post_ECE_em_15: the post-temperature-scaling ECE with equal mass bin size = 15.
  • pre_ECE_em_20: the pre-temperature-scaling ECE with equal mass bin size = 20.
  • post_ECE_em_20: the post-temperature-scaling ECE with equal mass bin size = 20.
  • pre_ECE_em_25: the pre-temperature-scaling ECE with equal mass bin size = 25.
  • post_ECE_em_25: the post-temperature-scaling ECE with equal mass bin size = 25.
  • pre_ECE_em_50: the pre-temperature-scaling ECE with equal mass bin size = 50.
  • post_ECE_em_50: the post-temperature-scaling ECE with equal mass bin size = 50.
  • pre_ECE_em_100: the pre-temperature-scaling ECE with equal mass bin size = 100.
  • post_ECE_em_100: the post-temperature-scaling ECE with equal mass bin size = 100.
  • pre_ECE_em_200: the pre-temperature-scaling ECE with equal mass bin size = 200.
  • post_ECE_em_200: the post-temperature-scaling ECE with equal mass bin size = 200.
  • pre_ECE_em_500: the pre-temperature-scaling ECE with equal mass bin size = 500.
  • post_ECE_em_500: the post-temperature-scaling ECE with equal mass bin size = 500.
  • pre_cwCE_5: the pre-temperature-scaling cwCE with bin size = 5.
  • post_cwCE_5: the post-temperature-scaling cwCE with bin size = 5.
  • pre_cwCE_10: the pre-temperature-scaling cwCE with bin size = 10.
  • post_cwCE_10: the post-temperature-scaling cwCE with bin size = 10.
  • pre_cwCE_15: the pre-temperature-scaling cwCE with bin size = 15.
  • post_cwCE_15: the post-temperature-scaling cwCE with bin size = 15.
  • pre_cwCE_20: the pre-temperature-scaling cwCE with bin size = 20.
  • post_cwCE_20: the post-temperature-scaling cwCE with bin size = 20.
  • pre_cwCE_25: the pre-temperature-scaling cwCE with bin size = 25.
  • post_cwCE_25: the post-temperature-scaling cwCE with bin size = 25.
  • pre_cwCE_50: the pre-temperature-scaling cwCE with bin size = 50.
  • post_cwCE_50: the post-temperature-scaling cwCE with bin size = 50.
  • pre_cwCE_100: the pre-temperature-scaling cwCE with bin size = 100.
  • post_cwCE_100: the post-temperature-scaling cwCE with bin size = 100.
  • pre_cwCE_200: the pre-temperature-scaling cwCE with bin size = 200.
  • post_cwCE_200: the post-temperature-scaling cwCE with bin size = 200.
  • pre_cwCE_500: the pre-temperature-scaling cwCE with bin size = 500.
  • post_cwCE_500: the post-temperature-scaling cwCE with bin size = 500.
  • pre_cwCE_em_5: the pre-temperature-scaling cwCE with equal mass bin size = 5.
  • post_cwCE_em_5: the post-temperature-scaling cwCE with equal mass bin size = 5.
  • pre_cwCE_em_10: the pre-temperature-scaling cwCE with equal mass bin size = 10.
  • post_cwCE_em_10: the post-temperature-scaling cwCE with equal mass bin size = 10.
  • pre_cwCE_em_15: the pre-temperature-scaling cwCE with equal mass bin size = 15.
  • post_cwCE_em_15: the post-temperature-scaling cwCE with equal mass bin size = 15.
  • pre_cwCE_em_20: the pre-temperature-scaling cwCE with equal mass bin size = 20.
  • post_cwCE_em_20: the post-temperature-scaling cwCE with equal mass bin size = 20.
  • pre_cwCE_em_25: the pre-temperature-scaling cwCE with equal mass bin size = 25.
  • post_cwCE_em_25: the post-temperature-scaling cwCE with equal mass bin size = 25.
  • pre_cwCE_em_50: the pre-temperature-scaling cwCE with equal mass bin size = 50.
  • post_cwCE_em_50: the post-temperature-scaling cwCE with equal mass bin size = 50.
  • pre_cwCE_em_100: the pre-temperature-scaling cwCE with equal mass bin size = 100.
  • post_cwCE_em_100: the post-temperature-scaling cwCE with equal mass bin size = 100.
  • pre_cwCE_em_200: the pre-temperature-scaling cwCE with equal mass bin size = 200.
  • post_cwCE_em_200: the post-temperature-scaling cwCE with equal mass bin size = 200.
  • pre_cwCE_em_500: the pre-temperature-scaling cwCE with equal mass bin size = 500.
  • post_cwCE_em_500: the post-temperature-scaling cwCE with equal mass bin size = 500.
  • pre_MCE_5: the pre-temperature-scaling MCE with bin size = 5.
  • post_MCE_5: the post-temperature-scaling MCE with bin size = 5.
  • pre_MCE_10: the pre-temperature-scaling MCE with bin size = 10.
  • post_MCE_10: the post-temperature-scaling MCE with bin size = 10.
  • pre_MCE_15: the pre-temperature-scaling MCE with bin size = 15.
  • post_MCE_15: the post-temperature-scaling MCE with bin size = 15.
  • pre_MCE_20: the pre-temperature-scaling MCE with bin size = 20.
  • post_MCE_20: the post-temperature-scaling MCE with bin size = 20.
  • pre_MCE_25: the pre-temperature-scaling MCE with bin size = 25.
  • post_MCE_25: the post-temperature-scaling MCE with bin size = 25.
  • pre_MCE_50: the pre-temperature-scaling MCE with bin size = 50.
  • post_MCE_50: the post-temperature-scaling MCE with bin size = 50.
  • pre_MCE_100: the pre-temperature-scaling MCE with bin size = 100.
  • post_MCE_100: the post-temperature-scaling MCE with bin size = 100.
  • pre_MCE_200: the pre-temperature-scaling MCE with bin size = 200.
  • post_MCE_200: the post-temperature-scaling MCE with bin size = 200.
  • pre_MCE_500: the pre-temperature-scaling MCE with bin size = 500.
  • post_MCE_500: the post-temperature-scaling MCE with bin size = 500.
  • pre_KSCE: the pre-temperature-scaling KSCE.
  • post_KSCE: the post-temperature-scaling KSCE.
  • pre_KDECE: the pre-temperature-scaling KDECE.
  • post_KDECE: the post-temperature-scaling KDECE.
  • pre_MMCE: the pre-temperature-scaling MMCE.
  • post_MMCE: the post-temperature-scaling MMCE.
  • pre_NLL: the pre-temperature-scaling NLL.
  • post_NLL: the post-temperature-scaling NLL.
  • pre_Brier: the pre-temperature-scaling Brier.
  • post_Brier: the post-temperature-scaling Brier.
  • auc_cifar10c: the AUC on CIFAR-10-C (corrupted with guassian noise).
  • auc_svhn: the AUC on SVHN.
  • fgsm_Linf_acc0_1: fgsm attack with Linf norm and epsilon = 0.1.
  • fgsm_Linf_acc0_5: fgsm attack with Linf norm and epsilon = 0.5.
  • fgsm_Linf_acc1_0: fgsm attack with Linf norm and epsilon = 1.0.
  • fgsm_Linf_acc2_0: fgsm attack with Linf norm and epsilon = 2.0.
  • fgsm_Linf_acc3_0: fgsm attack with Linf norm and epsilon = 3.0.
  • fgsm_Linf_acc4_0: fgsm attack with Linf norm and epsilon = 4.0.
  • fgsm_Linf_acc5_0: fgsm attack with Linf norm and epsilon = 5.0.
  • fgsm_Linf_acc6_0: fgsm attack with Linf norm and epsilon = 6.0.
  • fgsm_Linf_acc7_0: fgsm attack with Linf norm and epsilon = 7.0.
  • fgsm_Linf_acc8_0: fgsm attack with Linf norm and epsilon = 8.0.
  • fgsm_Linf_acc255_0: fgsm attack with Linf norm and epsilon = 255.0.
  • pgd_Linf_acc0_1: pgd attack with Linf norm and epsilon = 0.1.
  • pgd_Linf_acc0_5: pgd attack with Linf norm and epsilon = 0.5.
  • pgd_Linf_acc1_0: pgd attack with Linf norm and epsilon = 1.0.
  • pgd_Linf_acc2_0: pgd attack with Linf norm and epsilon = 2.0.
  • pgd_Linf_acc3_0: pgd attack with Linf norm and epsilon = 3.0.
  • pgd_Linf_acc4_0: pgd attack with Linf norm and epsilon = 4.0.
  • pgd_Linf_acc8_0: pgd attack with Linf norm and epsilon = 8.0.
  • aa_apgd_ce_Linf_acc0_1: aa_apgd attack with Linf norm and epsilon = 0.1.
  • aa_apgd_ce_Linf_acc0_5: aa_apgd attack with Linf norm and epsilon = 0.5.
  • aa_apgd_ce_Linf_acc1_0: aa_apgd attack with Linf norm and epsilon = 1.0.
  • aa_apgd_ce_Linf_acc2_0: aa_apgd attack with Linf norm and epsilon = 2.0.
  • aa_apgd_ce_Linf_acc3_0: aa_apgd attack with Linf norm and epsilon = 3.0.
  • aa_apgd_ce_Linf_acc4_0: aa_apgd attack with Linf norm and epsilon = 4.0.
  • aa_apgd_ce_Linf_acc8_0: aa_apgd attack with Linf norm and epsilon = 8.0.
  • aa_square_Linf_acc0_1: aa_square attack with Linf norm and epsilon = 0.1.
  • aa_square_Linf_acc0_5: aa_square attack with Linf norm and epsilon = 0.5.
  • aa_square_Linf_acc1_0: aa_square attack with Linf norm and epsilon = 1.0.
  • aa_square_Linf_acc2_0: aa_square attack with Linf norm and epsilon = 2.0.
  • aa_square_Linf_acc3_0: aa_square attack with Linf norm and epsilon = 3.0.
  • aa_square_Linf_acc4_0: aa_square attack with Linf norm and epsilon = 4.0.
  • aa_square_Linf_acc8_0: aa_square attack with Linf norm and epsilon = 8.0.
  • brightness: accuracy on dataset corrupted with brightness.
  • contrast: accuracy on dataset corrupted with contrast.
  • defocus_blur: accuracy on dataset corrupted with defocus_blur.
  • elastic_transform: accuracy on dataset corrupted with elastic_transform.
  • fog: accuracy on dataset corrupted with fog.
  • frost: accuracy on dataset corrupted with frost.
  • gaussian_noise: accuracy on dataset corrupted with gaussian_noise.
  • glass_blur: accuracy on dataset corrupted with glass_blur.
  • impulse_noise: accuracy on dataset corrupted with impulse_noise.
  • jpeg_compression: accuracy on dataset corrupted with jpeg_compression.
  • motion_blur: accuracy on dataset corrupted with motion_blur.
  • pixelate: accuracy on dataset corrupted with pixelate.
  • shot_noise: accuracy on dataset corrupted with shot_noise.
  • snow: accuracy on dataset corrupted with snow.
  • zoom_blur: accuracy on dataset corrupted with zoom_blur.

About

Dataset for "A Benchmark Study on Calibration" ICLR 2024

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages

  • Jupyter Notebook 99.5%
  • Other 0.5%