From Disentangled Representation to Concept Ranking: Interpreting Deep Representations in Image Classification tasks - Paper Code

General Aspects

The codes presented here were developed for the XKDD-2022 workshop paper. To execute the methods presented here, the NetDissect code must be executed first. The project link is https://github.com/CSAILVision/NetDissect-Lite. I ran it using the "settings.py" file presented here.

In this paper we used the Action40 dataset - http://vision.stanford.edu/Datasets/40actions.html - to extract the features.

The codes "feature_extraction_global.py" and "feature_extraction_local.py" present the general idea about how I extracted global and local concepts from the model, based on the NetDissection result.

The "linear_classification.py" shows how the ranking based on the classification task was developed.

The jupyter notebook "XKDD_metrics.ipynb" presents the metric calculation for the paper.

Reproducibility

The structure for the reproducibility is:

• Run NetDissect (https://github.com/CSAILVision/NetDissect-Lite) using the "settings.py"
• Download the Action40 dataset - http://vision.stanford.edu/Datasets/40actions.html
• Run "feature_extraction_global.py" to extract the global concepts.
• Run "linear_classification.py" to generate the ranked global concepts.
• Run "feature_extraction_local.py" to extract the local concepts and generate the ranked local concepts.
• Run "XKDD_metrics.ipynb" to analyse the result.

Reference

  @inproceedings{ferreira2023disentangled,
    title={From Disentangled Representation to Concept Ranking: Interpreting Deep Representations in Image Classification Tasks},
    author={Ferreira dos Santos, Eric and Mileo, Alessandra},
    booktitle={Machine Learning and Principles and Practice of Knowledge Discovery in Databases: International Workshops of ECML PKDD 2022, Grenoble, France, September 19--23, 2022, Proceedings, Part I},
    pages={322--335},
    year={2023},
    organization={Springer}
  }