Datasets and splits used for experiments for the paper 'Towards Scalable and Unified Example-based Explanation and Outlier Detection'

link to paper https://arxiv.org/abs/2011.05577

For LSUN experiments

To prepare LSUN dataset (lsun_split folder) :

Refer to https://github.com/fyu/lsun to download the whole latest dataset. After downloading the dataset, run the script file command_to_run_data.sh (specify paths in the script) to export 10,000 train images per class and all the given validation images as 128 x 128 PNG files.

The train, val, and test splits used in our experiments are given in the folder:

• lsun_split/new_allcls_lsun_split/
  • new10k_train.txt
  • new10k_val.txt
  • new_test.txt
  • new_test_subset100.txt     (subset used in LRP perturbation)

Each line is in the above textfiles is in the form of:

• imagename   class_label

To create LSUN strokes and altered color outlier samples, cd to lsun_split/codes_genoutlier_data folder. Run the following commands:

python gen_synthetic_strokes.py --dataset lsun --thickness 5 --rootPath /some/path/to/LSUN_images_PL/ python gen_altered_colors.py --dataset lsun --rootPath /some/path/to/LSUN_images_PL/

For PCam experiments

To prepare PCam dataset (pcam_split folder) :

We used the train, val, and test split from https://github.com/basveeling/pcam.

The tar file pcam_split/new_allcls_patchcam_split.tar.xz contains the following test subsets used in LRP perturbation and outlier evaluation:

• camelyonpatch_level_2_split_test_subset500_x.h5     (subset used in LRP perturbation)
• camelyonpatch_level_2_split_test_subset500_y.h5     (subset used in LRP perturbation)
• camelyonpatch_level_2_split_test_subset1500_x.h5     (subset used in outlier evaluation)
• camelyonpatch_level_2_split_test_subset1500_y.h5     (subset used in outlier evaluation)

To create PCam strokes and altered color outlier samples, cd to pcam_split/codes_genoutlier_data folder. Run the following commands:

python gen_synthetic_strokes.py --dataset patchcam --thickness 5 --rootPath /rootpath/to/new_allcls_patchcam_split/ python gen_altered_colors.py --dataset patchcam --rootPath /rootpath/to/new_allcls_patchcam_split/

For Stanford Cars experiments

To prepare Stanford dataset (lsun_split folder) :

Download the dataset from http://ai.stanford.edu/~jkrause/cars/car_dataset.html and put them in the folder stanford_cars. Then perform data augmentation on the images in the folder stanford_cars/cars_train/ using the same data augmentation as the authors from the ProtoPNet work https://github.com/cfchen-duke/ProtoPNet/blob/master/img_aug.py

We used only samples from the first 50 car models. The following textfiles contain the unaugmented train, val, and test splits:

• stanfordcar_split/new_allcls_stanfordcar_split/
  • new_train90_50classes_withoutdataug.txt     (unaugmented training images)
  • new_val10_50classes.txt
  • test_50classes.txt

For training, we used the augmented images of the original image in new_train90_50classes_withoutdataug.txt and thus we have 1835 x 30 = 55,050 training samples in total. For validation and test, we used the unaugmented images as listed in new_val10_50classes.txt and test_50classes.txt, respectively.

To create Stanford Cars strokes and altered color outlier samples, cd to stanfordcar_split/codes_genoutlier_data folder. Run the following commands:

python gen_synthetic_strokes.py --dataset stanfordcar --thickness 5 --rootPath /path/to/stanford_cars/ python gen_altered_colors.py --dataset stanfordcar --rootPath /path/to/stanford_cars/