CenterPose

CenterPose for keypoint / pose estimation. Detects object centers and regresses keypoint locations. Used for 6-DoF object pose estimation.

Set model.backbone.pretrainedbackbonepath.

For TAO Deploy TensorRT actions (gentrtengine, TensorRT evaluate, and TensorRT inference), use the deploy spec templates packaged 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: centerpose
• Formats: default
• Monitoring metric: val_3DIoU

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.

TRAIN_DIR = "/path/to/extracted/train"
VAL_DIR = "/path/to/extracted/val"
TEST_DIR = "/path/to/extracted/test"
INFER_DIR = VAL_DIR
CAL_IMAGE_DIRS = ["/path/to/extracted/train/<sequence_or_image_dir>"]

train (mandatory data sources):

{
    "train.num_epochs": 30,
    "train.checkpoint_interval": 10,
    "train.validation_interval": 10,
    "train.num_gpus": 1,
    "dataset.category": "bike",
    "dataset.batch_size": 4,
    "dataset.train_data": TRAIN_DIR,
    "dataset.val_data": VAL_DIR,
}

evaluate (mandatory data sources):

{
    "dataset.category": "bike",
    "dataset.test_data": TEST_DIR,
}

inference (mandatory data sources):

{
    "dataset.category": "bike",
    "dataset.inference_data": INFER_DIR,
}

gentrtengine (mandatory data sources):

{
    "gen_trt_engine.tensorrt.calibration.cal_image_dir": CAL_IMAGE_DIRS,
}

Eval Dataset

Optional. Val and test datasets are provided as separate tarballs.

Important Parameters

• dataset.num_classes: Number of object categories. Default 1.
• dataset.num_joints: Number of keypoints per object. Fixed at 8 (bbox keypoints). Valid range: exactly 8.
• dataset.input_res: Input resolution. Fixed at 512. Output resolution fixed at 128.
• dataset.category: Object category name. Default "cereal_box".
• model.backbone.modeltype: Default fansmall. Backbone options limited in schema.
• train.optim.lr: Learning rate. Default 6e-5. MultiStep scheduler with lrsteps=[90, 120], lrdecay=0.1.
• train.lossconfig: Rich loss config with toggles: mseloss, objscale, objscaleuncertainty, hpsuncertainty, regbbox, hmhp. Weights: whweight=0.1, offweight=1, hp_weight=1.
• inference.usepnp: Use PnP for 6-DoF pose. Default True. Requires camera intrinsics (focallengthx/y, principlepoint_x/y).
• export.inputwidth: Export input size. Fixed at 512x512. opsetversion=16.

Multi-GPU / Multi-Node

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

• Strategy: auto (Lightning picks the best strategy automatically)
• No explicit numnodes or distributedstrategy config — single-node only
• No sync_batchnorm

Export / TRT Defaults

• Export input: 512x512 (fixed), opset 16
• TRT data types: FP32, FP16, INT8
• TRT optbatchsize: 4, maxbatchsize: 8

Hardware

Minimum 1 GPU(s), recommended 2 GPU(s). 16GB+ VRAM per GPU. CenterPose is moderately memory-intensive depending on input resolution and number of keypoints.

Error Patterns

numjoints mismatch: Ensure dataset.numjoints matches the keypoint count in your annotations.

Extract S3 tarballs for local Docker: The starter-kit S3 data is packaged as train.tar.gz, val.tar.gz, and test.tar.gz, but the CenterPose TAO actions consume extracted folders. Extract each archive and set dataset.traindata, dataset.valdata, dataset.testdata, and dataset.inferencedata to the extracted split directories.

Checkpoint handoff: CenterPose training writes concrete checkpoints such as modelepoch000step00008.pth and a centerposemodellatest.pth symlink. Use the SDK/model checkpoint resolver or the exact epoch/step checkpoint for evaluate, inference, export, and resume. Use the symlink only when the user explicitly asks for latest.

TAO Deploy postprocessor compatibility: Use the deploy image resolved from versions.yaml or the selected platform. A successful gentrtengine run does not prove deploy evaluate or inference works; inspect those action exit codes and logs separately, especially for CenterPose postprocessor errors such as TypeError: only 0-dimensional arrays can be converted to Python scalars.

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 centerpose.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-centerpose](references/tao-deploy-centerpose.md)