Pose Classification

Pose classification using ST-GCN (Spatial Temporal Graph Convolutional Network). Classifies skeleton sequences into action categories from pose keypoint data.

Typically trained from scratch on skeleton data.

The packaged PyTorch Pose Classification CLI supports datasetconvert, train, evaluate, export, and inference. datasetconvert is conditional: run it only when the input is raw DeepStream BodyPose JSON. If the dataset is already converted to TAO-ready .npy / .pkl files, start directly with train on those files and mark dataset conversion as not run: preconverted dataset provided in validation reports. This model does not expose deploy, prune, quantize, or standalone retrain actions. Resume/retrain behavior uses poseclassification train -e ... with train.resumetrainingcheckpointpath 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 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: pose_classification
• Formats: default
• Monitoring metric: val_loss

Per-Action Dataset Requirements

Typical Spec Overrides

Data source overrides are mandatory for every action being run — the agent MUST construct data source paths from the Per-Action Dataset Requirements table above and include them in specoverrides. Do not run datasetconvert when the supplied dataset is already converted to .npy / .pkl files.

S3_TRAIN = "s3://bucket/data/purpose_built_models_pose_classification_train/nvidia"
CHECKPOINT = "/results/{train_job_id}/results_dir/model_epoch_000_step_00007.pth"

dataset_convert (optional; raw DeepStream BodyPose JSON only):

{
    "dataset_convert.data": "s3://bucket/data/<deepstream-bodypose-output>.json",
}

train (mandatory data sources):

{
    "train.num_epochs": 30,
    "train.checkpoint_interval": 10,
    "train.validation_interval": 10,
    "train.num_gpus": 1,
    "wandb.enable": False,
    "dataset.num_classes": 6,
    "dataset.label_map": {
        "class_0": 0,
        "class_1": 1,
        "class_2": 2,
        "class_3": 3,
        "class_4": 4,
        "class_5": 5,
    },
    "model.graph_layout": "nvidia",
    "dataset.train_dataset.data_path": f"{S3_TRAIN}/train_data.npy",
    "dataset.train_dataset.label_path": f"{S3_TRAIN}/train_label.pkl",
    "dataset.val_dataset.data_path": f"{S3_TRAIN}/val_data.npy",
    "dataset.val_dataset.label_path": f"{S3_TRAIN}/val_label.pkl",
}

resume train (mandatory checkpoint):

{
    "train.num_epochs": 31,
    "train.resume_training_checkpoint_path": CHECKPOINT,
    "dataset.train_dataset.data_path": f"{S3_TRAIN}/train_data.npy",
    "dataset.train_dataset.label_path": f"{S3_TRAIN}/train_label.pkl",
    "dataset.val_dataset.data_path": f"{S3_TRAIN}/val_data.npy",
    "dataset.val_dataset.label_path": f"{S3_TRAIN}/val_label.pkl",
}

evaluate (mandatory data sources):

{
    "evaluate.test_dataset.data_path": f"{S3_TRAIN}/val_data.npy",
    "evaluate.test_dataset.label_path": f"{S3_TRAIN}/val_label.pkl",
    "evaluate.checkpoint": CHECKPOINT,
}

export (mandatory checkpoint and output):

{
    "export.checkpoint": CHECKPOINT,
    "export.onnx_file": "/results/{export_job_id}/results_dir/pose_classification.onnx",
}

inference (mandatory data sources):

{
    "inference.test_dataset.data_path": f"{S3_TRAIN}/test_data.npy",
    "inference.test_dataset.label_path": f"{S3_TRAIN}/test_label.pkl",
    "inference.checkpoint": CHECKPOINT,
    "inference.output_file": "/results/pose_classification_inference.txt",
}

Dataset Convert

Dataset conversion is optional for Pose Classification. Run poseclassification datasetconvert only when the user supplies raw DeepStream BodyPose JSON. For the common S3 validation dataset, the data is already converted to traindata.npy, trainlabel.pkl, valdata.npy, vallabel.pkl, testdata.npy, and testlabel.pkl; use those files directly for train/evaluate/inference/export flows and do not synthesize fake BodyPose JSON.

Eval Dataset

Optional. Validation data is provided alongside training as valdata.npy / vallabel.pkl. TAO training emits valloss as the TensorBoard validation scalar for this model; use valloss with minimize direction for AutoML selection unless a custom evaluation hook supplies a different metric.

Important Parameters

• dataset.num_classes: Number of pose action classes. Default 6.
• model.graph_layout: Skeleton graph layout. Options: nvidia, openpose. Determines joint connectivity.
• model.graph_strategy: Graph partitioning strategy for GCN.
• train.optim.lr: Learning rate. Default 0.1 (SGD). Higher than vision models due to graph convolution properties.
• model.dropout: Dropout rate for regularization.

Multi-GPU / Multi-Node

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

• Strategy: auto (Lightning picks best strategy automatically)
• No explicit numnodes or distributedstrategy config — single-node only
• Lightweight model, single GPU typically sufficient

Hardware

Minimum 1 GPU(s), recommended 1 GPU(s). 8GB+ VRAM per GPU. Pose classification is very lightweight — skeleton data is small. Single GPU is sufficient.

Error Patterns

Graph layout mismatch: Ensure model.graph_layout matches the skeleton format in your .npy data files.

Label shape mismatch: trainlabel.pkl class indices must be in range [0, numclasses).

Missing label map: The training dataloader expects dataset.labelmap to be a dictionary. If the dataset only supplies numeric class IDs, set a synthetic contiguous map such as class0: 0 through class_5: 5 for the six-class NVIDIA sample data.

Checkpoint handoff: After AutoML/train, use the checkpoint resolver to select the intended saved .pth checkpoint under the parent result folder, such as modelepoch000step00007.pth, and pass that exact file as evaluate.checkpoint, export.checkpoint, inference.checkpoint, or train.resumetrainingcheckpointpath. pcmodellatest.pth is a latest-checkpoint symlink; use it only when the user explicitly asks for latest rather than a specific/best checkpoint. Keep the same dataset.numclasses, dataset.labelmap, and model.graphlayout overrides for downstream actions.

Dataset conversion source: datasetconvert expects the raw JSON output from the DeepStream BodyPose app. The common NVIDIA sample S3 folder is already converted to traindata.npy, trainlabel.pkl, valdata.npy, vallabel.pkl, testdata.npy, and test_label.pkl; skip conversion and start from the converted files when those are present.

Action-specific dataset paths: The evaluate and inference templates also contain the training dataset.traindataset and dataset.valdataset blocks. For evaluate, populate evaluate.testdataset.datapath and evaluate.testdataset.labelpath. For inference, populate inference.testdataset.datapath and set inference.outputfile; do not stop after replacing the first datapath or label_path in the file.

Output files: Export needs an explicit export.onnxfile path. Inference must set inference.outputfile to a writable file path; the packaged template default is an empty string, and the current PyTorch inference code opens that value directly.

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 pose_classification.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.