PointPillars

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).

PointPillars for 3D object detection from LiDAR point clouds. Encodes point clouds into a pseudo-image via pillar-based representation, then applies 2D detection. Used in autonomous driving / robotics.

Typically trained from scratch. Provide train.resumetrainingcheckpoint_path to resume.

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

The packaged PyTorch PointPillars CLI supports datasetconvert, train, evaluate, inference, export, and prune. It does not expose a parent-model gentrtengine action; TensorRT engine generation is deploy-only. It also does not expose a separate retrain subcommand. Retraining from a pruned model uses pointpillars train -e ... with train.prunedmodel_path populated.

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 for an action requires schemas/<action>.schema.json and references/spectemplate<action>.yaml to exist and parse. Use the packaged selected-action 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: pointpillars
• Formats: default
• Monitoring metric: 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.

DATA_ROOT = "s3://bucket/data/pointpillars"
DATA_INFO = "/results/{dataset_convert_job_id}/results_dir/data_info"
CHECKPOINT = "/results/{train_job_id}/results_dir/checkpoint_epoch_1.pth"
PRUNED_MODEL = "/results/{prune_job_id}/results_dir/pruned_0.1.tlt"

The raw PointPillars data root must be an extracted folder containing matching train/lidar, train/label, val/lidar, and val/label subfolders before datasetconvert runs. If the source dataset is packaged as separate train/val archives, extract both under the same mounted data root and point dataset.datapath at that root.

train (mandatory data sources):

{
    "train.num_epochs": 30,
    "train.checkpoint_interval": 10,
    "train.validation_interval": 10,
    "train.num_gpus": 1,
    "dataset.data_path": DATA_ROOT,
    "dataset.data_info_path": DATA_INFO,
}

resume train (mandatory checkpoint):

{
    "dataset.data_path": DATA_ROOT,
    "dataset.data_info_path": DATA_INFO,
    "train.resume_training_checkpoint_path": CHECKPOINT,
}

evaluate (mandatory data sources):

{
    "dataset.data_path": DATA_ROOT,
    "dataset.data_info_path": DATA_INFO,
    "evaluate.checkpoint": CHECKPOINT,
}

export (mandatory data sources):

{
    "dataset.data_path": DATA_ROOT,
    "dataset.data_info_path": DATA_INFO,
    "export.checkpoint": CHECKPOINT,
    "export.onnx_file": "/results/{export_job_id}/results_dir/pointpillars.onnx",
}

inference (mandatory data sources):

{
    "dataset.data_path": DATA_ROOT,
    "dataset.data_info_path": DATA_INFO,
    "inference.checkpoint": CHECKPOINT,
}

prune (mandatory data sources):

{
    "dataset.data_path": DATA_ROOT,
    "dataset.data_info_path": DATA_INFO,
    "prune.model": CHECKPOINT,
}

retrain (mandatory data sources):

{
    "dataset.data_path": DATA_ROOT,
    "dataset.data_info_path": DATA_INFO,
    "train.pruned_model_path": PRUNED_MODEL,
}

For local Docker, DATAINFO must be visible inside every train/evaluate/export/prune/retrain container. Use the datasetconvert job from the same results root, or mount/copy the converted resultsdir/datainfo folder into the current run and set dataset.datainfopath to that mounted container path. If the host scratch root is mounted at /results and the conversion artifacts live under host scratch/results/<jobid>/resultsdir/datainfo, the direct-job container path is /results/results/<jobid>/resultsdir/datainfo. Do not reuse a /results/<job_id>/... path from another run root unless that folder is mounted into the current job.

For AutoML train workflows, perform this as a launch preflight before calling AutoMLRunner.run: create or materialize the datasetconvert output under the current run's RESULTSROOT, set dataset.datainfopath to that current-run container path, and verify dbinfostrain.pkl, infostrain.pkl, and infosval.pkl are present from the train container's point of view. If a runner is cloned or adapted from a prior AutoML algorithm, update the conversion artifact in the new run root; a stale CONVERTJOB_ID from another results mount is not valid.

Eval Dataset

Optional. Validation data (val.tar.gz) is separate from training. Used for mAP evaluation.

Important Parameters

• train.num_epochs: Default 80 (much higher than other TAO models). PointPillars needs more epochs for convergence on 3D detection.
• train.lr: Learning rate. Default 0.003 (adam_onecycle scheduler).
• dataset.class_names: List of 3D object classes. Default 7 classes (KITTI-style). Modify to match your dataset.
• dataset.data_path: Path to point cloud data directory.
• dataset.datainfopath: Path to data info files from dataset_convert step.
• dataset.pointcloudrange: Spatial extent of the point cloud to consider. Must match your sensor configuration.
• model.densehead.anchorgenerator_config: Anchor configurations per class. Must be tuned for your object sizes and the point cloud range.

Multi-GPU / Multi-Node

Launch method: torchrun (LIGHTNINGEXCLUDEDNETWORK). Uses PyTorch native DistributedDataParallel (NOT Lightning Trainer).

• CUDAVISIBLEDEVICES is explicitly set from TAOVISIBLEDEVICES
• Uses nn.parallel.DistributedDataParallel directly (not Lightning strategy)
• NODE_RANK is copied to RANK if RANK is unset

Multi-node env vars (set by orchestrator):

Hardware

Minimum 1 GPU(s), recommended 4 GPU(s). 16GB+ (V100 or A100) VRAM per GPU. PointPillars is relatively efficient for 3D detection. The main bottleneck is data I/O for large point cloud datasets.

Error Patterns

datasetconvert required: Training will fail if dataset.datainfopath is not populated from a prior datasetconvert job. Always run convert first, and verify the train container can see dbinfostrain.pkl and infostrain.pkl under dataset.datainfopath. A common local-Docker failure is a stale /results/<oldjobid>/... path from a different results root.

Point cloud range mismatch: If pointcloudrange does not match the actual sensor data extent, detections will be poor or empty.

Epoch numbering: PointPillars checkpoint epoch numbers may be offset by 1 from status.json reported epochs.

Checkpoint selection: PointPillars training emits checkpoints named like checkpointepoch1.pth. For evaluation, inference, export, prune, and resume, select the intended checkpoint through the model/job checkpoint resolver and pass that exact file to evaluate.checkpoint, inference.checkpoint, export.checkpoint, prune.model, or train.resumetrainingcheckpoint_path. Do not guess by taking the newest model.pth; this model does not use that filename.

Prune/retrain key: PointPillars prune writes an encrypted .tlt artifact. Keep a non-empty key in the prune and retrain specs; the packaged templates use the TAO default tltencode. If key is omitted or null, the toolkit can still exit with a container success code while logging a passphrase error and creating an empty pruned0.1.tlt. Always verify the pruned model is nonzero before using it for retrain.

Status files matter: Some PointPillars failures can be followed by Execution status: PASS in the entrypoint footer and a Docker exit code of 0. Check results_dir/status.json and the expected artifact before marking an action as passed.

Local resultsdir wiring: For direct local-Docker specs, set the top-level resultsdir as well as any action-specific *.resultsdir field. If only evaluate.resultsdir is set and the top-level field is left blank, evaluate can try to write under /opt/nvidia/eval and then still print the generic PASS footer. Treat that as a failed action unless the expected result directory and status/artifact files exist.

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 pointpillars.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.

Deployment

• [tao-deploy-pointpillars](references/tao-deploy-pointpillars.md)