TRAIS-Lab / dattri

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

[dattri.benchmark] Add TracIn example (noisy label on MNIST and CIFAR-2) #78

Closed tingwl0122 closed 3 months ago

tingwl0122 commented 3 months ago

Description

This PR adapts the current example/mnist_lr/influence_function_noisy_label.py to TracIn on MNIST/CIFAR-2 .

1. Motivation and Context

Test the functionality of TracIn

2. Summary of the change

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

tingwl0122 commented 3 months ago

For MNIST, TracIn works well. For CIFAR-2, it is just like random checking. Not sure TracIn functionality yet...

(Update: IF is also similar to this)

tingwl0122 commented 3 months ago

For MNIST, TracIn works well. For CIFAR-2, it is just like random checking. Not sure TracIn functionality yet...

(Update: IF is also similar to this)

Hi @TheaperDeng , I think this is also ready to be merged. The only concern is the poor performance of TracIn on CIFAR-2 noisy label test. But since IF also gives similar result, I guess this is normal.

TheaperDeng commented 3 months ago

please fix the conflict before merge the PR.