Tao Train Grounding Dino

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
• Training monitoring metric: val_mAP50
• AutoML metric contract: for evaluation-backed selection, use test_mAP50 with maximize direction. The packaged evaluate action emits test_mAP50; use that evaluator KPI for AutoML trial comparison and final selection rather than the training-only val_mAP50 key.

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