dino
Routes DINO DETR object-detection setup, COCO data validation, CUDA deformable-attention preparation, training, evaluation, inference, visualization, and bounded benchmarking tasks.
SKILL.mdskills/repositories/repo-skills/dino/SKILL.md
GitHubDINO repo skill
Use this skill for the official IDEA Research implementation of DINO: DETR with Improved DeNoising Anchor Boxes for End-to-End Object Detection. It is a router and operating guide for a DINO checkout or compatible installed runtime; it is not a replacement for model weights, COCO data, or the PyTorch/CUDA runtime.
Route by task
- Prepare or diagnose an environment, COCO layout, model shape, or compiled deformable-attention operator: read data-model-setup.
- Train, fine-tune, resume, choose scale/backbone, or construct a single/distributed/Submitit command: read training.
- Evaluate a checkpoint, run one custom image, interpret boxes, visualize, or measure GFLOPS/FPS: read inference-evaluation.
For a task spanning routes, run setup first, then training or inference. Do not start a long job until the setup route has recorded its data, config, package, and backend gates.
Minimum prerequisites
The upstream project documents Python 3.7.3, PyTorch 1.9.0, and CUDA 11.1 as
its historical reference environment. Keep torch and torchvision on a
compatible release family and compile the custom operator against the same
CUDA family. A verified modern inspection combination used Python 3.11,
PyTorch 2.5.1+cu121, torchvision 0.20.1+cu121, CUDA 12.1 nvcc, GCC 12.4,
and an A100 (compute capability 8.0). Treat that as evidence for one tested
combination, not a universal compatibility promise.
Install the documented runtime requirements in an isolated environment:
python -m pip install -r requirements.txt
python -m pip check
For standard DINO deformable attention, verify torch, torchvision,
pycocotools, and MultiScaleDeformableAttention. Add panopticapi only
for panoptic/mask workflows. Read
data-model-setup environment guidance
before compiling CUDA code.
Safe first checks
From the target project root, run the bundled read-only probes before changing anything:
python skills/disco/dino/sub-skills/data-model-setup/scripts/check_dino_environment.py \
--require-extension --require-coco --pip-check
python skills/disco/dino/sub-skills/data-model-setup/scripts/validate_coco_layout.py \
/path/to/COCODIR --split val
The first command checks imports and reports backend status; it does not install
or build. The second checks JSON references and paths without downloading,
copying, deleting, or rewriting data. Add --require-cuda and --smoke-cuda
only when a free visible GPU is intentionally selected.
Cross-cutting rules
- Pair 4-scale checkpoints with
DINO_4scale.pyand 5-scale checkpoints withDINO_5scale.py; also match backbone and class vocabulary. num_classesfollows the repository's maximum category-ID-plus-one convention. For custom data, preserve the documented denoising ruledn_labelbook_size >= num_classes + 1.- DINO's deformable encoder/decoder depends on the compiled CUDA operator. A CPU import or config parse is not evidence that the CUDA model path works.
batch_sizeis per process. Compute the effective global batch before comparing runs or changing learning-rate assumptions.--resumerestores a full training state;--pretrain_model_pathperforms partial model initialization and is the fine-tuning path.- Do not download checkpoints/data, launch Slurm jobs, mutate dataset roots, or start long training from the bundled helper scripts by accident.
For install/import/config failures, read shared troubleshooting. For provenance or refresh decisions, read repository provenance. For a reviewable evaluation command, use the bundled evaluator planner; for a reviewable operator build, use the bundled extension wrapper. Both print commands and require explicit launch flags for side effects.
Verification boundary
This skill was integrated from source commit d84a491d41898b3befd8294d1cf2614661fc0953.
The prepared environment passed package imports, pip check, config/API smoke
checks, CUDA allocation on a selected free A100, and compilation/import of the
custom operator. Full COCO training/evaluation, downloaded checkpoints,
pretrained Swin/ConvNeXt assets, the numerical operator test, and active Slurm
submission were intentionally deferred; preserve those limits in any result
report.