brainstorming
brainstorming
You MUST use this before any creative work - creating features, building components, adding functionality, modifying behavior, or planning anything new, in software or out of it (a talk, a business, a renovation).
Explore skills
Development · SKILL.md
dlr_margin_loss
dlr_margin_lossCompute DLR-style logit margin losses for APGD-DLR adversarial optimization checks.
Independent GitHub snapshot · Automated classification · Original documentation governs
Skill documentation
GitHubname: dlr_margin_loss description: Compute DLR-style logit margin losses for APGD-DLR adversarial optimization checks.
DLR Margin Loss
Use this skill when an adversarial attack needs the AutoAttack DLR loss or when validating that a proxy attack preserves APGD-DLR semantics. Do not use it for binary classifiers or tasks without ordered class logits.
Inputs
- A logit vector with at least three classes.
- The true class index.
Outputs
- Scalar untargeted DLR loss.
- Ranking diagnostics from the helper script.
Workflow
- Sort logits while retaining the true-class position.
- Identify the strongest non-true competitor.
- Normalize the true-versus-competitor margin by the spread between top and third logits.
- Maximize this loss during untargeted APGD.
Validation
Run python tests/test_dlr.py or the generated skill validator.
Limitations
This pure-Python helper is intended for deterministic small examples and mirrors the untargeted loss needed by the recovery harness.