This is the code for the paper A Comparison of Instance Attribution Methods. All third party code is publicly available except for the pre-trained ExPred model, contact Avishek Anand for more information. This is part of the 2023 TuDelft Research Projec (https://github.com/TU-Delft-CSE/Research-Project).
- Pull FastIF, TracIn, and the pre-trained ExPred model, and import into designated files folders.
- From FastIF only
experiments
andinfluence_utils
are required. Use the includednn_influence_utils.py
. Minor changes to other files may be necessary. - In
src/TracIn
put the files alongside main_tracin.py - Install the required packages from the
requirements.txt
bypip install -r requirements.txt
- In
main.py
adjust constants as needed and run.
Not that depending on your hardware you may have to change the k
in main.py
.
├── dataset: the eraser FEVER datasets and custom dataset utils
├── expred: the pretrained expred model (not publicly available)
├── fastif: fastif from https://github.com/salesforce/fast-influence-functions
├── README.md: this file
├── tracin: tracin from https://github.com/frederick0329/TracIn
└── main.py: main execution file