Skip to content

Hyperparam tuning #25

Description

@NightlyTwo58

So far, we have some really weird results where PGD and Autoattack are getting really low 0-1% accuracy, inconsistent with existing published papers.
We don't think this is due to a bug in our code; our normalization and quantization handling should be solidified at this point.

We propose the following

  • removing normalization for AutoAttack (it appears it may not be necessary)
  • reducing EPS for PGD and furthermore
  • migrating from Linf to L2 PGD
  • remove normalization for AutoAttack

Metadata

Metadata

Labels

No labels
No labels

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions