Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[WIP] Fix batching in DIG #114

Open
wants to merge 25 commits into
base: main
Choose a base branch
from
Open

[WIP] Fix batching in DIG #114

wants to merge 25 commits into from

Conversation

gsarti
Copy link
Member

@gsarti gsarti commented Dec 2, 2021

Description

Fixes #113

Type of Change

🔧 Bug fix (non-breaking change which fixes an issue)

@gsarti
Copy link
Member Author

gsarti commented Nov 30, 2022

Current status:

  • DIG attribution was made compatible with newer changes (decoder-only model support, etc.) and supports both source and source-target attribution.
  • The problematic behavior of inconsistent outputs when testing the same input alone vs. inside a batch persist. The issue is probably in the scale_input function call inside DiscretizedIntegratedGradient.attribute, since the rest is mostly compatible with the original IG attribution.
  • Open question: is return_convergence_delta meaningful in this setting? First thoughts: it could provide information on the distance from a good non-orthogonal path, but should not be used as in IG.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

Inconsistent batching for DiscretizedIntegratedGradients attributions
1 participant