Split out from #404 per Emlyn's suggestion, to stop it blocking the rest of that PR's testing-infrastructure work.
Open question: .github/workflows/scheduled-training-check.yml (added in #404) currently re-downloads and re-trains against whatever DINOv2 weights timm resolves to on a weekly schedule - Emlyn flagged a licensing risk in this, since DINOv2 itself started under a non-commercial license before later moving to Apache 2.0, and an unpinned weekly pull wouldn't catch a similar future change.
A ready fix is prepared on branch 404-pin-dinov2-weights-version: pins timm to a specific version so the weights stay fixed until a maintainer deliberately bumps it (verified working). That approach would also remove the need for the scheduled workflow itself, per Emlyn's original comment - though there's a separate, smaller benefit the schedule currently provides (catching the notebook's own code going stale, independent of weights, as happened previously with pipeline_step_by_step.ipynb).
@emlynjdavies @nepstad - once there's agreement on direction, this is ready to implement quickly.
Split out from #404 per Emlyn's suggestion, to stop it blocking the rest of that PR's testing-infrastructure work.
Open question:
.github/workflows/scheduled-training-check.yml(added in #404) currently re-downloads and re-trains against whatever DINOv2 weightstimmresolves to on a weekly schedule - Emlyn flagged a licensing risk in this, since DINOv2 itself started under a non-commercial license before later moving to Apache 2.0, and an unpinned weekly pull wouldn't catch a similar future change.A ready fix is prepared on branch
404-pin-dinov2-weights-version: pinstimmto a specific version so the weights stay fixed until a maintainer deliberately bumps it (verified working). That approach would also remove the need for the scheduled workflow itself, per Emlyn's original comment - though there's a separate, smaller benefit the schedule currently provides (catching the notebook's own code going stale, independent of weights, as happened previously withpipeline_step_by_step.ipynb).@emlynjdavies @nepstad - once there's agreement on direction, this is ready to implement quickly.