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:7.1.0-pyt  # versions-key: images.tao_toolkit.pyt
TAO_PYT_IMAGE="${TAO_PYT_IMAGE:-$TAO_PYT_IMAGE_DEFAULT}"
RUN_ROOT

Train:

Evaluate:

Inference:

Export:

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 train_loss as the optimization metric with direction=minimize, and set train.optim.monitor_name: train_loss in spec_overrides. NVPanoptix3D train jobs emit PRQ, RSQ, and RRQ in status.json, and the training progress log emits train_loss; short or minimal jobs may not emit val_loss, including full-trial smoke runs. Multi-fidelity AutoML algorithms such as Hyperband, ASHA, and BOHB may promote a checkpoint to a resume job that completes without emitting a fresh train_loss line. In that case, the AutoML metric is the carried-forward metric from the source rung job that emitted train_loss; still verify the promoted job resumed from the explicit epoch/step checkpoint, produced a real checkpoint, and is usable for 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: train_loss
• AutoML direction: minimize
• For AutoML train jobs, use train_loss. For multi-fidelity resume jobs that do not emit a fresh train_loss, compare AutoML's carried metric to the source rung job that emitted it. Validation status KPIs are PRQ, RSQ, and RRQ; do not use val_loss unless a specific run is known to emit it.

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>/....

train (mandatory data sources):

evaluate (mandatory data sources):

inference (mandatory data sources):

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.

Multi-GPU / Multi-Node

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

• fsdp is NOT supported for NVPanoptix3D (code only handles ddp)
• ddp with activation checkpointing (enabled by default): find_unused_parameters=False
• ddp without: find_unused_parameters=True
• FAN backbones with 3D enabled auto-enable sync_batchnorm

Multi-node env vars (set by orchestrator): WORLD_SIZE, NODE_RANK, MASTER_ADDR, MASTER_PORT, NUM_GPU_PER_NODE.

Export / TRT Defaults

• Exports the 2D ONNX model to export.onnx_file_2d. The current export entrypoint calls export_2d_model; export.onnx_file_3d is present in the schema but not produced by this toolkit image.
• TRT data types: FP32, FP16 only
• max_voxels: 700000 (engine input tensor limit)

Hardware

Minimum 2 GPU(s), recommended 4 GPU(s). 40GB+ (A100 recommended) VRAM per GPU. 3D reconstruction is very memory intensive. Use train.precision: fp32; the current training entrypoint rejects fp16. activation_checkpoint enabled by default. FSDP for multi-node. AutoML is enabled at the model layer; preserve this GPU/VRAM guidance when routing train through AutoML.

Error Patterns

nvpanoptix3d: not found in the PyTorch image: Use the packaged module entrypoint command: python -m nvidia_tao_pytorch.cv.nvpanoptix3d.entrypoint.nvpanoptix3d <action> -e <spec>. The 7.0 PyTorch image contains the NVPanoptix3D package but does not expose a nvpanoptix3d console script.

Missing frustum mask: Ensure meta/frustum_mask.npz is present in the dataset directory.

Downsample factor mismatch: Use downsample_factor=2 for Matterport3D, 1 for Front3D / synthetic datasets.

3D occupancy OOM: Reduce frustum_dims or grid_dimensions if running out of GPU memory during 3D reconstruction.

fp16 precision rejected: The schema advertises fp16, but the current training entrypoint raises ValueError: Only fp32 precision is supported. Use train.precision: fp32 for train and resume/retrain.

Inference dataloader length is zero: inference.images_dir is scanned only for top-level .jpg and .png files. If the S3 test archive extracts to scene subdirectories, create or point to a flat folder of real RGB images before running inference.

3D ONNX missing after export: The current export entrypoint only calls the 2D ONNX exporter and writes export.onnx_file_2d. Do not require export.onnx_file_3d unless the toolkit image adds a 3D exporter.

Resume stops at the epoch boundary: A one-epoch smoke run writes an end-of-epoch checkpoint such as model_epoch_000_step_00020.pth. Resuming with train.num_epochs set only one epoch beyond the original run can restore the checkpoint and stop without producing a new epoch checkpoint. When validating actual retraining from an epoch-boundary checkpoint, set train.num_epochs at least two epochs beyond the source smoke run and raise train.optim.max_steps accordingly. For example, resuming from model_epoch_000_step_00020.pth needs train.num_epochs: 3 and enough max steps to produce a new exact epoch/step checkpoint such as model_epoch_001_step_00040.pth before handing the model to evaluate, inference, or export.

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.

Model-specific handoff mappings:

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.