TRAIS-Lab / dattri

`dattri` is a PyTorch library for developing, benchmarking, and deploying efficient data attribution algorithms.
https://trais-lab.github.io/dattri/
MIT License
27 stars 8 forks source link

[dattri.algorithm] Fix Arnoldi algorithm's bug #112

Closed tingwl0122 closed 1 month ago

tingwl0122 commented 2 months ago

Description

Fix Arnoldi algorithm's bug by using arnoldi_project defined in dattri/projection.py

1. Motivation and Context

Now the eigenspace calculated by Arnoldi algorithm will be fixed throughout the attribution.

2. Summary of the change

3. What tests have been added/updated for the change?

tingwl0122 commented 2 months ago

The performance is (slightly) verified through MNIST+LR example in dattri/example/mnist_lr/influence_function_noisy_label.py

Checked Data Sample      Found flipped Sample     
--------------------------------------------------
0                        0                        
100                      69                       
200                      92                       
300                      105                      
400                      106                      
500                      108                      
600                      108                      
700                      108                      
800                      108                      
900                      108