Data for "Exploring specialization and sensitivity of convolutional neural networks in the context of simultaneous image augmentations"

This repository contains data collected under the following study:P.Kharyuk, S.Matveev, I.Oseledets. Exploring specialization and sensitivity of convolutional neural networks in the context of simultaneous image augmentations, arXiv:2503.03283.

Corresponding source code repository: https://github.com/kharyuk/activation_sa

1_sensitivity_values.7z: sensitivity values (Sobol indices, Shapley values) computed for the first experimental series. In addition, logs and npz-files containing the sampled parameters were packed. To be used within the jupyter notebooks, all .hdf5 files should be placed into the 'results' directory of the source code repository.

2_guided_masking_predictions.7z: these files include the top-5 class predictions and corresponding classifying layer's outputs computed for the second experimental series. Every .hdf5 file should be placed into the 'results' directory of the source code repository. Logs and pkl files with sampling parameters are also included in the archive.

3_single_channelled_segments.7z: sensitivity values (Sobol indices, Shapley values) computed for the third experimental series (for the standalone segments). As for the previous experimental series, logs and npz-files tracing the sampled parameters are provided within the archive. Every .hdf5 file should be placed into the 'results/single_units' directory of the source code repository.

4_test_references.7z: examples of testing scripts output (core computations with a limited number of samples and checkpoints). Not related to the article course.

The unpacking procedure for archives 1-3 is automated by the bash script located in tools/unpack_results.sh of the GitHub repository.