Skip to content

Repository with files and scripts for the human study of the CRP (Concept-Relevance-Propagation) paper

Notifications You must be signed in to change notification settings

maxdreyer/crp-human-study

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

crp-human-experiment

This is the repository for the human study of the Concept-Relevance-Propagation (CRP) paper: Achtibat, Reduan, et al. "From" Where" to" What": Towards Human-Understandable Explanations through Concept Relevance Propagation." (2022).

The CRP explanations used in this study are based on the official CRP implementation for PyTorch.

We have three models trained on ImageNet:

  1. vgg16 (standard torchvision pretrained),
  2. vgg16_corrupted (trained on border artifact for 30 classes) and
  3. vgg16_uncorrupted (trained on border artifact for all classes).

Model checkpoints can be found here.

In directory ./scripts/ there are several .sh files for running experiments.

  • run_prepare_crp.sh to run analysis run for crp
  • run_all.sh to run all methods for the samples given

Code is tested with Python 3.8.10. Required packages are listed in requirements.txt.

Please note the HTML template files for amazon mechanical turk as well as instructory images in directory ./html_templates/. Study results are in .csv format in ./results/csv/ where _T or _F indicates the answers for the corrupted and uncorrupted model, respectively.

Regarding the CRP explanation generation in ./experiments/crp_.py, concept labels are predicted using MILAN (Natural Language Descriptions of Deep Visual Features) (package located in ./utils/milan)

About

Repository with files and scripts for the human study of the CRP (Concept-Relevance-Propagation) paper

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published