Tao Train Nvpanoptix3d

NVPanoptix3D

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).

NVPanoptix3D for panoptic 3D scene reconstruction from posed RGB images. Produces 3D panoptic segmentation (semantic, instance, and panoptic masks) with occupancy completion. Built on VGGT backbone with Mask2Former-style head and 3D frustum reconstruction.

Uses 2D and 3D stage checkpoints. Set train.checkpoint_2d and train.checkpoint_3d for staged initialization.

Quick Start (docker run)

Docker-native launch — no TAO SDK and no Python on the host. Use the local Docker/platform skill instead when it gives a stricter environment-specific command (non-root UID mapping, cache redirects, remote daemons).

TAO_PYT_IMAGE_DEFAULT=nvcr.io/nvidia/tao/tao-toolkit:7.2.0-pyt  # versions-key: images.tao_toolkit.pyt
TAO_PYT_IMAGE="${TAO_PYT_IMAGE:-$TAO_PYT_IMAGE_DEFAULT}"
RUN_ROOT="${RUN_ROOT:-$PWD}"
DOCKER_COMMON=(
  --rm --gpus all --shm-size=8g
  --shm-size=8g
  --ulimit memlock=-1
  --ulimit stack=67108864
  -v "$RUN_ROOT/data:/data:ro"
  -v "$RUN_ROOT/specs:/specs:ro"
  -v "$RUN_ROOT/results:/results"
)

Train:

docker run "${DOCKER_COMMON[@]}" "$TAO_PYT_IMAGE" \
  python -m nvidia_tao_pytorch.cv.nvpanoptix3d.entrypoint.nvpanoptix3d train -e /specs/train.yaml

Evaluate:

docker run "${DOCKER_COMMON[@]}" "$TAO_PYT_IMAGE" \
  python -m nvidia_tao_pytorch.cv.nvpanoptix3d.entrypoint.nvpanoptix3d evaluate -e /specs/evaluate.yaml

Inference:

docker run "${DOCKER_COMMON[@]}" "$TAO_PYT_IMAGE" \
  python -m nvidia_tao_pytorch.cv.nvpanoptix3d.entrypoint.nvpanoptix3d inference -e /specs/inference.yaml

Export:

docker run "${DOCKER_COMMON[@]}" "$TAO_PYT_IMAGE" \
  python -m nvidia_tao_pytorch.cv.nvpanoptix3d.entrypoint.nvpanoptix3d export -e /specs/export.yaml

Every action takes its spec with -e; results_dir is set in the spec or overridden on the command line. Mount any pretrained-weights directory the spec references, and keep every in-container path consistent across actions.

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.

For AutoML, use PRQ as the optimization metric with direction=maximize. NVPanoptix3D train validation and evaluate jobs both emit PRQ, RSQ, and RRQ in status.json, so use PRQ consistently for the baseline, every recommendation, and final best-checkpoint evaluation. The model may use train.optim.monitor_name: train_loss internally for checkpointing, but minimal jobs do not reliably export a numeric train_loss to the TAO status channel; do not use it as the AutoML selection metric. Multi-fidelity promotions must obtain a fresh PRQ after the resumed epoch and must still resume from the explicit epoch/step checkpoint, produce a real checkpoint, and pass evaluate/inference. 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: nvpanoptix3d
• Formats: front3d, matterport
• Monitoring metric: PRQ
• AutoML direction: maximize
• Validation and evaluate status KPIs are PRQ, RSQ, and RRQ. Use PRQ for AutoML selection; do not use train_loss or val_loss unless a different workflow proves that exact scalar is externally emitted for every trial and obtains explicit approval to use a proxy objective.

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. For packaged S3 folders that store scene data as data/images.tar.gz, the skill metadata requests extraction into the parent data/ directory because the TAO loader expects base_dir/data/<scene_id>/....

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.enable_3d": True,
    "dataset.contiguous_id": True,
    "model.sem_seg_head.num_classes": 13,
    "dataset.frustum_mask_path": f"{S3_TRAIN}/meta/frustum_mask.npz",
    "dataset.label_map": f"{S3_TRAIN}/meta/colormap.json",
    "dataset.train.json_path": f"{S3_TRAIN}/meta/train.json",
    "dataset.train.base_dir": f"{S3_TRAIN}",
    "dataset.val.json_path": f"{S3_EVAL}/meta/val.json",
    "dataset.val.base_dir": f"{S3_EVAL}",
    "dataset.test.json_path": f"{S3_EVAL}/meta/test.json",
    "dataset.test.base_dir": f"{S3_EVAL}",
}

evaluate (mandatory data sources):

{
    "evaluate.checkpoint": "<selected train/AutoML checkpoint>",
    "dataset.enable_3d": True,
    "dataset.contiguous_id": True,
    "dataset.frustum_mask_path": f"{S3_EVAL}/meta/frustum_mask.npz",
    "dataset.label_map": f"{S3_EVAL}/meta/colormap.json",
    "dataset.val.json_path": f"{S3_EVAL}/meta/val.json",
    "dataset.val.base_dir": f"{S3_EVAL}",
    "dataset.test.json_path": f"{S3_EVAL}/meta/test.json",
    "dataset.test.base_dir": f"{S3_EVAL}",
}

inference (mandatory data sources):

{
    "inference.checkpoint": "<selected train/AutoML checkpoint>",
    "dataset.enable_3d": True,
    "dataset.frustum_mask_path": f"{S3_EVAL}/meta/frustum_mask.npz",
    "dataset.label_map": f"{S3_EVAL}/meta/colormap.json",
    "inference.images_dir": "/path/to/flat_rgb_images",
}

Eval Dataset

Optional. Val/test splits configured via dataset.val and dataset.test paths.

Important Parameters

• model.sem_seg_head.num_classes: Number of semantic classes. Default 13.
• model.mode: Prediction mode. Options: panoptic, instance, semantic. Default panoptic.
• model.backbone_type: Backbone. Default vggt (only option in schema).
• model.mask_former.num_object_queries: Object queries. Default 100.
• model.mask_former.dec_layers: Decoder layers. Default 10.
• model.frustum3d.truncation: 3D frustum truncation. Default 3.
• model.frustum3d.panoptic_weight: Panoptic loss weight. Default 25.
• model.frustum3d.completion_weights: Completion loss weights. Default [50, 25, 10].
• dataset.name: Dataset name. Options: front3d, matterport, synthetic_hospital, synthetic_warehouse.
• dataset.contiguous_id: Set True when the label-map JSON already supplies trainId values for its category IDs; leaving the default False can synthesize placeholder categories without trainId and fail during metadata construction.
• dataset.downsample_factor: Image downsample factor. Default 1 (Front3D), 2 (Matterport).
• dataset.target_size: Target image size. Default [320, 240].
• dataset.depth_min: Min depth. Default 0.4 meters.
• dataset.depth_max: Max depth. Default 6.0 meters.
• train.lr: Learning rate. Default 2e-4. backbone_multiplier=0.1.
• train.lr_scheduler: Options: MultiStep, Warmuppoly. Milestones [88, 96].
• train.precision: Only fp32 is supported by the current train code.
• train.distributed_strategy: Options: ddp, fsdp. activation_checkpoint=True by default.
• train.clip_grad_norm: Gradient clipping norm. Default 0.1.
• export.onnx_file_2d: ONNX path for 2D model component.
• export.max_voxels: Max voxels for engine input. Default 700000.
• inference.mode: Options: semantic, instance, panoptic.