Mask Grounding DINO

Mask Grounding DINO for grounded instance segmentation. Extends Grounding DINO with mask prediction head for open-set segmentation guided by text prompts.

Set train.pretrainedmodelpath for full model weights.

For TAO Deploy TensorRT actions (gentrtengine, TensorRT evaluate, and TensorRT inference), read references/tao-deploy-mask-grounding-dino.md first. Deploy spec templates live in this skill's references/ folder with the spectemplatedeploy_*.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/spectemplate<action>.yaml from the schema top-level default field. AutoML enablement is declared at the model layer in references/skillinfo.yaml via automlenabled. Runnable AutoML still requires schemas/train.schema.json and references/spectemplatetrain.yaml to exist and parse. Use the packaged train schema for automldefaultparameters, automldisabledparameters, 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/skillinfo.yaml and resolve the run override from either an explicit automlpolicy value or the user's workflow request. Use automlpolicy: 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 automlpolicy: off for this run only. When automlpolicy: on, automlenabled: true, and both schemas/train.schema.json and references/spectemplatetrain.yaml are packaged, route the train action through tao-skill-bank:tao-run-automl by default with this model's skilldir. Preserve workflow/application overrides for datasets, specs, output directories, GPU/platform settings, parent checkpoints, and automlpolicy. 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: segmentation
• Formats: odvg, coco, coco_raw
• Monitoring metric: val_loss

Per-Action Dataset Requirements

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_gpus": 1,
    "train.num_epochs": 10,
    "train.checkpoint_interval": 10,
    "train.validation_interval": 10,
    "dataset.val_data_sources.data_type": "OD",
    "model.num_region_queries": 100,
    "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"},
}

evaluate (mandatory data sources):

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

inference (mandatory data sources):

{
    "inference.checkpoint": "<selected train/AutoML checkpoint>",
    "dataset.infer_data_sources.data_type": "OD",
    "dataset.infer_data_sources": {"image_dir": f"{S3_EVAL}/images.tar.gz", "captions": ["person", "bicycle", "car"]},
}

quantize (mandatory data sources):

{
    "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_TRAIN}/images.tar.gz", "json_file": f"{S3_TRAIN}/annotations_odvg.jsonl", "label_map": f"{S3_TRAIN}/annotations_odvg_labelmap.json"},
}

Eval Dataset

Optional. Validation uses COCO-format annotations even when training uses ODVG.

Important Parameters

• model.backbone: Default swintiny224_1k. Same backbone options as Grounding DINO.
• train.optim.lr: Learning rate. Default 2e-4. lr_backbone 2e-5. Reuses GDINOTrainExpConfig — same training setup as Grounding DINO.
• model.num_queries: Object queries. Default 900.
• model.enclayers / model.declayers: Keep both at 6 for train/AutoML

runs. The mask head asserts six decoder outputs during validation, so copying Grounding DINO smoke overrides that reduce transformer layers causes an immediate failure.

• AutoML metric note: Use metric="val_loss" with

direction="minimize" for train-stage AutoML. The packaged train loop logs validation loss scalars; it does not emit [bbox] val_mAP@50 during the train job.

• model.has_mask: Enables mask prediction head. Default True. Adds mask/dice/rela loss coefficients.
• model.numregionqueries: Number of region queries for mask prediction. Default 100.
• model.losstypes: Loss components. Default [labels, boxes, masks]. Includes masklosscoef, dicelosscoef, relaloss_coef.
• evaluate.ioi_threshold: IoI threshold for mask evaluation. Default 0.5.
• evaluate.nms_threshold: NMS threshold. Default 0.2.
• evaluate.text_threshold: Text matching threshold. Default 0.3.
• dataset.hasmask: Dataset includes mask annotations. Default True. valdatasources default datatype is "VG".

Multi-GPU / Multi-Node

Launch method: Lightning-managed. Same DDP/FSDP behavior as Grounding DINO.

Hardware

Minimum 1 GPU(s), recommended 4 GPU(s). 24GB+ (A100 recommended) VRAM per GPU. Heavier than Grounding DINO due to mask prediction head. 24GB+ GPU memory recommended.

Error Patterns

CUDA out of memory: Reduce batch_size. Mask prediction adds overhead on top of Grounding DINO.

Deploy schema error for testthreshold: TAO Deploy uses evaluate.textthreshold and inference.textthreshold. Do not use testthreshold in deploy specs.

Deploy model shape mismatch: Carry transformer and mask structure fields from export into deploy evaluate/inference specs, including model.numqueries, model.numselect, model.maxtextlen, model.numregionqueries, and model.has_mask. These values must match the ONNX model used to build the TensorRT engine.

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 createjob(). This mirrors the old microservices inferparams.py flow.

Inference mappings from TAO Core maskgroundingdino.config.json:

For parentmodel or parentmodelfolder, pass the upstream train/export/AutoML child job id as parentjob_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 Mask Grounding DINO checkpoint outside the SDK resolver, match the intended epoch/step artifact exactly, for example modelepoch000step00049.pth. The maskgdinomodellatest.pth symlink is valid only when latest is explicitly requested. The parent PyTorch maskgrounding_dino CLI supports train, evaluate, inference, export, and quantize; run TensorRT engine generation, TensorRT inference, and TensorRT evaluation through references/tao-deploy-mask-grounding-dino.md.

Deployment

• [tao-deploy-mask-grounding-dino](references/tao-deploy-mask-grounding-dino.md)