tao-train-grounding-dino

Imported from nvidia/skills (skills/tao-train-grounding-dino/SKILL.md) via skills.sh. Install upstream with npx skills add nvidia/skills --skill tao-train-grounding-dino. Copyright stays with the author (Apache-2.0).

Grounding DINO

Standalone install? If this session was not initialized by the TAO skill bank plugin, run the tao-setup skill first (host preflight, credentials, cross-skill discovery).

Grounding DINO for open-set object detection. Combines DINO-style detection with BERT text encoder for language-guided detection. Detects objects described by text prompts without fixed class vocabulary.

Set train.pretrained_model_path for full Grounding DINO weights or model.pretrained_backbone_path for backbone-only.

For TAO Deploy TensorRT actions (gen_trt_engine, TensorRT evaluate, and TensorRT inference), read references/tao-deploy-grounding-dino.md first. Deploy spec templates live in this skill's references/ folder with the spec_template_deploy_*.yaml prefix.

Dataclass Schemas

Generated TAO Core schemas are packaged in schemas/<action>.schema.json, with schemas/manifest.json listing available actions. Each generated schema also emits references/spec_template_<action>.yaml from the schema top-level default field. AutoML enablement is declared at the model layer in references/skill_info.yaml via automl_enabled. Runnable AutoML for an action requires schemas/<action>.schema.json and references/spec_template_<action>.yaml to exist and parse. Use the packaged selected-action schema for automl_default_parameters, automl_disabled_parameters, defaults, min/max bounds, enums, option weights, math conditions, dependencies, and popular parameters. Do not expect ~/tao-core at runtime; maintainers regenerate schemas/templates before packaging the skill bank.

Train Action Policy

This model is AutoML-enabled at the model layer. Before handling any train-stage request, read references/skill_info.yaml and resolve the run override from either an explicit automl_policy value or the user's workflow request. Use automl_policy: on by default and only expose on / off in new launch prompts. Treat phrases like "turn off AutoML", "disable AutoML", "no HPO", or "plain training" as automl_policy: off for this run only. When automl_policy: on, automl_enabled: true, and both schemas/train.schema.json and references/spec_template_train.yaml are packaged, route the train action through tao-skill-bank:tao-run-automl by default with this model's skill_dir. Preserve workflow/application overrides for datasets, specs, output directories, GPU/platform settings, parent checkpoints, and automl_policy. Use direct model training only when automl_policy: off or the packaged train schema/template is missing; in the missing-schema case, report that AutoML is enabled but not runnable for this model until schemas are generated.

Non-train actions such as evaluate, inference, export, and deploy flows stay in this model skill. The per-run automl_policy override does not change model metadata.

Training Requirements

• Dataset type: object_detection
• Formats: odvg, coco, raw
• Monitoring metric: val_mAP50

Per-Action Dataset Requirements

The runner may source image archives as images.tar.gz, but direct local Docker TAO CLI specs must point image_dir to an extracted image directory. Skill metadata marks these archive-backed image sources with runtime: extracted_folder so a fresh runner can unpack the archive before launching TAO.

Typical Spec Overrides

Data source overrides are mandatory for every action — the agent MUST construct data source paths from the Per-Action Dataset Requirements table above and include them in spec_overrides.

S3_TRAIN = "s3://bucket/data/train"
S3_EVAL = "s3://bucket/data/eval"

train (mandatory data sources):

{
    "train.num_epochs": 10,
    "train.checkpoint_interval": 10,
    "train.validation_interval": 10,
    "train.num_gpus": 1,
    "dataset.train_data_sources": [{"image_dir": f"{S3_TRAIN}/images.tar.gz", "json_file": f"{S3_TRAIN}/annotations_odvg.jsonl", "label_map": f"{S3_TRAIN}/annotations_odvg_labelmap.json"}],
    "dataset.val_data_sources": {"image_dir": f"{S3_EVAL}/images.tar.gz", "json_file": f"{S3_EVAL}/annotations.json"},
}

deploy/gen_trt_engine (use references/tao-deploy-grounding-dino.md):

{
    "gen_trt_engine.onnx_file": "<exported_onnx_uri>",
    "gen_trt_engine.trt_engine": "<output_engine_path>",
    "gen_trt_engine.tensorrt.data_type": "FP16",
}

inference (mandatory data sources):

{
    "inference.checkpoint": "<selected train/AutoML checkpoint>",
    "dataset.infer_data_sources.image_dir": [f"{S3_EVAL}/images.tar.gz"],
    "dataset.infer_data_sources.captions": [
        "fire extinguisher",
        "cone",
        "cart",
        "forklift"
    ],
}

evaluate (mandatory data sources):

{
    "evaluate.checkpoint": "<selected train/AutoML checkpoint>",
    "dataset.test_data_sources": {"image_dir": f"{S3_EVAL}/images.tar.gz", "json_file": f"{S3_EVAL}/annotations.json"},
}

quantize (mandatory data sources):

{
    "quantize.model_path": "<selected train checkpoint or exported ONNX model>",
    "dataset.train_data_sources": [{"image_dir": f"{S3_TRAIN}/images.tar.gz", "json_file": f"{S3_TRAIN}/annotations_odvg.jsonl", "label_map": f"{S3_TRAIN}/annotations_odvg_labelmap.json"}],
    "dataset.val_data_sources": {"image_dir": f"{S3_EVAL}/images.tar.gz", "json_file": f"{S3_EVAL}/annotations.json"},
    "dataset.quant_calibration_data_sources": {"image_dir": f"{S3_EVAL}/images.tar.gz", "json_file": f"{S3_EVAL}/annotations.json"},
}

Eval Dataset

Optional. Validation uses COCO-format annotations for mAP even though training can use ODVG format.

Important Parameters

• model.backbone: Default swin_tiny_224_1k. Also supports resnet_50 and other Swin variants. Swin generally performs better for grounding tasks.
• model.text_encoder_type: BERT model for text encoding. Default bert-base-uncased. max_text_len defaults to 256.
• model.max_text_len: Keep this aligned with the dataset label/token position maps. Do not shrink it for smoke tests unless the corresponding label maps are regenerated with the same length; otherwise validation can fail with a matrix shape mismatch between token probabilities and position maps.
• train.optim.lr: Learning rate. Default 2e-4. lr_backbone 2e-5. Supports bf16 precision in addition to fp16/fp32.
• dataset.max_labels: Maximum labels per image during training. Default 50. Increase for dense annotation datasets.
• model.num_queries: Object queries. Default 900 (higher than DINO's 300) due to open-vocabulary nature.
• model.num_queries / model.num_select: Keep num_queries high enough for the number of matched ODVG targets in a batch. Very small smoke values such as 20 can fail during Hungarian target indexing on dense images; use at least 100 for minimal Grounding DINO smoke runs unless the dataset is known to have fewer objects per image.
• train.optim.lr_steps: MultiStep LR schedule. Default [10].

Multi-GPU / Multi-Node

Launch method: Lightning-managed (single python process, Lightning spawns workers).

Same DDP/FSDP behavior as DINO. Multi-node requires WORLD_SIZE, NODE_RANK, MASTER_ADDR, MASTER_PORT env vars set by orchestrator.

Export / TRT Defaults

• Export input: 960x544 (larger than other OD models), opset 17. Keep Grounding-DINO export specs at the template export resolution for smoke tests; reducing export to very small image sizes such as 128x128 can trigger a PyTorch ONNX shape-inference assertion in the contrastive text head during torch.onnx.export.
• The parent PyTorch grounding_dino CLI supports train, evaluate, inference, export, and quantize. Run TensorRT engine generation, TensorRT inference, and TensorRT evaluation through references/tao-deploy-grounding-dino.md.
• TRT data types: FP32, FP16 only — INT8 is NOT supported
• TRT workspace: 8192 MB (8x larger than other OD models)
• TRT max_batch_size: 4

Hardware

Minimum 1 GPU(s), recommended 4 GPU(s). 24GB+ (A100 recommended) VRAM per GPU. Grounding DINO is heavier than standard DINO due to the text encoder (BERT). 24GB+ GPU memory recommended. Reduce batch_size for 16GB GPUs.

Error Patterns

CUDA out of memory: Reduce batch_size (4 -> 2 -> 1). The BERT text encoder adds significant memory overhead on top of the vision backbone.

Val annotation category IDs: Validation annotations should have category IDs starting from 0 for correct loss computation. Use annotation format conversion if needed.

Text encoder loading error: Ensure the container has access to download bert-base-uncased weights or provide a local path.

Quantize with a PyTorch checkpoint fails in TAO Toolkit 7.0.0-rc-226: The container's Grounding-DINO quantize script passes cap_lists=None when loading a checkpoint, which fails in post_process.py. ONNX quantization uses the exported ONNX artifact and COCO calibration data, but the default rc-226 PyTorch image also lacks the modelopt.onnx.quantization module. Treat this as an image/SDK blocker, not a checkpoint resolver issue.

mat1 and mat2 shapes cannot be multiplied in post_process.py: The text token length and label position maps are inconsistent, commonly because model.max_text_len was overridden below the default 256 while the dataset label maps still use 256-length position maps. Restore model.max_text_len or regenerate the label maps with the same length.

index is out of bounds for dimension 0 in criterion.py: model.num_queries is too small for the matched ODVG targets in the current batch. Increase model.num_queries and keep model.num_select compatible with it.

NotADirectoryError with images.tar.gz/<image>.jpg: The direct TAO CLI is trying to traverse an archive path as a directory. Extract the archive and set the relevant image_dir field to the extracted image folder; archive-backed skill data sources use runtime: extracted_folder for this reason.

Spec Param / Parent Model Inference

Model-specific inference mappings belong in this MD file, not in config.json. Generated runners should read this section and apply the mappings with SDK helpers before create_job(). This mirrors the old microservices infer_params.py flow.

Inference mappings from TAO Core grounding_dino.config.json:

For parent_model or parent_model_folder, pass the upstream train/export/AutoML child job id as parent_job_id. The SDK lists the parent result folder, filters checkpoint artifacts, and returns the selected model file or folder. Do not add these mappings back to config.json and do not patch generated runner scripts to guess checkpoint paths.

When selecting a Grounding-DINO checkpoint outside the SDK resolver, match the intended epoch/step artifact exactly, for example model_epoch_000_step_00046.pth. The gdino_model_latest.pth symlink is valid only when latest is explicitly requested. Carry structural model settings such as model.backbone, model.num_queries, model.num_select, model.num_feature_levels, model.max_text_len, and export input resolution forward into evaluate, inference, export, and deploy specs so checkpoint and engine shapes match.

Deployment

• tao-deploy-grounding-dino