Skip to content

T40: HF ViT analyze: adapter design and spike #345

Description

@mateuszwalo

Track: dev · Level: stretch · Effort: ~10h · Depends on: Write access

Why this matters

Today analyze targets torchvision-style CNNs. A huge share of modern practitioners use Hugging Face transformers, and the common objection to BNNR is that it is 'just demo CNNs'. Making analyze work on a real ViT removes that objection and opens BNNR to a much larger audience, which is a direct path toward standard status.

Steps

  1. Study how existing adapters work (src/bnnr/adapter.py) and how analyze consumes a model and its target layer.
  2. Spike: load a transformers ViT and produce one saliency map through BNNR (a ViT exposes attention differently from a CNN, so identify the right target).
  3. Write a short design note describing the adapter and any limitations, for maintainer review.

Done when

A design note approved by the maintainer and a spike that produces a saliency map from a ViT.


How to take this: comment "taking this" and wait to be assigned. Branch t40-short-desc from upstream/main, and put Closes #<this issue number> in your PR. Full workflow: the Cohort Handbook (pinned in Discord).

Metadata

Metadata

Assignees

No one assigned

    Labels

    cohortBNNR 2026 cohort tasklevel:stretchStretch tasktrack:devCohort development task

    Type

    No type

    Fields

    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions