---

name: data-model-setup description: "Prepare and validate DINO's COCO-style inputs, model/config shapes, and CUDA deformable-attention backend before training or evaluation." disable-model-invocation: true metadata: disco-role: operating license: Apache 2.0

DINO data and model setup

Use this route before a DINO training or evaluation run when the dataset, model configuration, or compiled backend is new, changed, or suspect. It builds no dataset and does not launch training, inference, or evaluation.

Route and gates

1. Environment gate. Run scripts/check_dino_environment.py with the strict requirements appropriate to the requested run. For the standard CUDA DINO path, use:

   python scripts/check_dino_environment.py --require-cuda \
     --require-extension --require-coco --require-runtime --pip-check --json
   

   It checks package imports, PyTorch/CUDA visibility, CUDA_HOME/nvcc observations, the MultiScaleDeformableAttention import, and optionally pip check. Add --require-panoptic only for the panoptic branch. Use --smoke-cuda and --smoke-extension only when a free visible GPU is available; import/symbol success is not the repository's numerical operator gate.

2. Data gate. Run scripts/validate_coco_layout.py against an existing COCO root. Validate only the requested split(s); add the optional panoptic root only when panoptic data is actually needed.

3. Target gate. Confirm the dataset yields the fields and coordinate conventions in data formats, including the num_classes/category-ID convention and transformed size versus orig_size.

4. Model gate. Compare the selected config with the backbone, returned feature levels, hidden dimension, attention heads, number of queries, class count, and denoising label-book rules in model architecture.

5. Extension gate. If the model will run its deformable encoder/decoder, compile and run the repository's CUDA operator test as described in environment. A package import alone is not a numerical proof.

Stop on a failed strict gate. Report the exact failed check and apply the corresponding recovery in troubleshooting rather than silently falling back to an incompatible backend.

Operating procedure

• Treat coco_path as a read-only input to this route. The standard instance layout is train2017/, val2017/, and annotations/; the expected JSON names are instances_train2017.json and instances_val2017.json. The test split uses test2017/ and image_info_test-dev2017.json when requested.
• Validate annotation references and image filenames before importing the dataset loader. Do not use the repository's optional DATA_COPY_SHILONG path here: its helper removes/recreates paths and is outside this skill's safety boundary.
• For a custom dataset, decide the maximum category ID and config class convention together. This repository's num_classes is a classifier width / maximum-object-ID-plus-one convention, not simply the count of category names. Recheck dn_labelbook_size after changing it.
• Confirm transforms are appropriate for the run. Training uses random flip, multiscale resize and a random crop branch; validation/evaluation uses a deterministic largest configured resize followed by normalization. The resulting boxes are normalized (center_x, center_y, width, height).
• Compile the extension only after confirming a CUDA-enabled PyTorch, a matching toolkit, a compatible host compiler (use GCC <= 12 when required by the toolkit), valid CUDA and CCCL include directories, and a free GPU. Keep build outputs in the normal build area; this route provides no cleanup, copy, download, or deletion helper.
• Route training launches, optimizer/scheduler settings, and checkpoints to the parent skill's training route. Route prediction, post-processing, and COCO/panoptic evaluation to its inference/evaluation route. This sub-skill supplies their setup facts only.

Expected handoff

Record the following for the next route: validated root and split(s), JSON files checked, image/annotation counts and warnings, whether masks/panoptic paths were checked, config name and salient shape settings, package/CUDA observations, extension import and numerical-test status, and unresolved backend limitations. Never record private machine prefixes, private dataset locations, or credentials.

References

• COCO data formats and targets
• DINO model architecture and compatibility
• Environment installation and extension build
• Troubleshooting and recovery

Scope and limitations

This is a preparation and validation route. It deliberately excludes dataset acquisition, checkpoint acquisition, training/evaluation launch commands, checkpoint surgery, destructive copy helpers, and claims that a CPU execution is an equivalent substitute for DINO's CUDA multi-scale deformable attention. The supplied scripts are read-only except for their --fixture mode, which creates and removes only a temporary self-test fixture.

Evidence provenance

This skill was distilled from the repository README installation/data sections; requirements.txt; datasets/coco.py, coco_panoptic.py, transforms.py, dataset.py, and data_util.py; the deformable-attention setup, test, function, and module files under models/dino/ops/; models/dino/backbone.py, deformable_transformer.py, and dino.py; util/slconfig.py; all config/DINO/*.py; tools/README.md; and the verified environment report for this repository. The report established a successful CUDA extension build and import, COCO dependency imports, and pip check; it did not establish a full training or evaluation result.