[dattri.algorithm] Add TracIn Attributor - Githubissues

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

28 stars 8 forks source link

[dattri.algorithm] Add TracIn Attributor #54

Closed tingwl0122 closed 4 months ago

tingwl0122 commented 4 months ago

Description

This PR implements TracIn Attributor as another TDA algorithm family.

1. Motivation and Context

To implement TracIn-related TDA methods

2. Summary of the change

Copy-and-paste Junwei's base class (dattri.algorithm.base from #32 )
Implement TracIn attributor, which covers TracIn, CosIn, Grad-Dot, Grad-Cos
Allows the attributor to use random_projection
Allows the usage of multiple parameter checkpoints

Note

[ ] TODO: allow multiple model paths (checkpoints) for computing gradients -> this should be covered in `dattri/model_utils/ensemble.py
[x] TODO: use double-loop to avoid explicit storing grad_train and grad_test
[x] TODO: allow projectors to be different across checkpoints

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

[x] Unit test: Typically, this should be included if you implemented a new function/fixed a bug.