[TAO-2511][Feature] Support frozen DINOv3 backbones in Visual ChangeNet

[nnagrajrao]

What changes are proposed in this pull request?

Add frozen DINOv3 backbone support to the existing Visual ChangeNet classification and segmentation paths with minimal changes to the AOI architecture.

• Supports ViT-S, S+, B, L, H+, and 7B through the existing ViT adapter and task heads.
• Keeps the DINOv3 ViT frozen while the AOI adapter and classification or segmentation head remain trainable.
• Gives an explicit converted TAO checkpoint precedence; when no checkpoint is configured, frozen DINOv3 loads matching timm/Hugging Face weights.
• Routes explicit euclidean checkpoints through the existing AOI parser and state-dict loader, including actionable rejection of raw SSL checkpoints.
• Adds the 7B channel map and 40-block adapter partition.
• Preserves existing C-RADIO behavior and defaults.

Example:

model:
  backbone:
    type: vit_large_dinov3
    pretrained_backbone_path: /data/pretrained_models/dinov3_vit_l_backbone.safetensors
    freeze_backbone: true

Set pretrained_backbone_path: null to use the matching timm/Hugging Face pretrained weights.

Why are the changes needed?

DEFT AOI needs to reuse DINOv3 SSL features without updating the backbone. This keeps the established Visual ChangeNet task-level fine-tuning design while allowing either a converted TAO DINOv3 checkpoint or the compatible public timm/Hugging Face weights.

Related issues

JIRA: TAO-2511.

Companion changes are in tao-core and tao-skill-bank on the same TAO-2511 branch.

Does this PR introduce any user-facing change?

Yes. Visual ChangeNet users can select all six DINOv3 variants and freeze the ViT. Existing backbone types, C-RADIO behavior, and defaults are unchanged.

How was this patch tested?

• Required TAO image, focused DINOv3 and config parity suite: 12 passed.
• Required TAO image, broader Visual ChangeNet model suite: 66 passed, 2 skipped.
• Exact PR-range CI hooks: license, pylint, pydocstyle, and flake8 passed.
• Confirmed timm 1.0.26 exposes all six configured model names.
• Independent architecture, correctness, and MR-readiness panel: 3/3 READY.
• git diff --check and no-Markdown checks passed.

Was this patch authored or co-authored using generative AI tooling?

Generated-by: OpenAI Codex

Release note

Visual ChangeNet classification and segmentation now support frozen DINOv3 ViT-S, S+, B, L, H+, and 7B backbones using converted TAO checkpoints or compatible timm/Hugging Face pretrained weights.

Checklist

• My commits are signed off per the DCO (git commit -s)
• I have read the contributing guidelines
• The code follows the project style, and lint and format checks pass locally
• I added or updated tests covering this change
• All applicable tests pass locally
• I updated inline configuration documentation and docstrings
• A version or changelog update is not required for this patch
• Optional imports remain guarded
• No secrets, credentials, proprietary data, or customer data are included
• I understand the contribution is licensed under the repository license

Notes for reviewers

No Markdown files are changed. Start with cv/backbone_v2/dino_v3.py, then the two Visual ChangeNet model constructors and the focused regression test. The change deliberately reuses existing AOI adapters and heads instead of introducing a parallel task package.

[tao-pr-bot]

For security reasons, CI does not run automatically on NVIDIA's runners — it must be triggered per commit.

• Comment /build to run CI on your latest commit, and make sure blossom-ci is green before merging.
• Re-run /build after each new push — CI is pinned to the head commit, so a stale run won't count.

Tip

Should this fix also ship in a release? Add a release/X.Y.Z label (matching the target release branch) and tao-cherry-pick-bot will backport it after merge.

Important

Before merging: every file needs a license header and every commit must be DCO signed-off — see the 📖 Contribution guide for details.

[nnagrajrao]

/build

[vpraveen-nv]

LGTM — the 7B wiring (40-block partition, 4096 channel maps) checks out against timm, the raw-SSL rejection matches the converter's remap table, and the C-RADIO/FAN paths are untouched. A few non-blocking questions inline. One sequencing note: please land the tao-core companion (NVIDIA-TAO/tao-core#31) in lockstep so schema validation doesn't reject vit_7b_dinov3 in the window between merges.

[hqtran-nv]

/build

[nnagrajrao]

/build

[hqtran-nv]

/build