Re-Identification

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).

Person re-identification. Learns discriminative embeddings to match the same person across different camera views. Metric learning based.

Set model.pretrainedmodelpath for pretrained weights.

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="${RUN_ROOT:-$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" \
  re_identification train -e /specs/train.yaml

Evaluate:

docker run "${DOCKER_COMMON[@]}" "$TAO_PYT_IMAGE" \
  re_identification evaluate -e /specs/evaluate.yaml

Inference:

docker run "${DOCKER_COMMON[@]}" "$TAO_PYT_IMAGE" \
  re_identification inference -e /specs/inference.yaml

Export:

docker run "${DOCKER_COMMON[@]}" "$TAO_PYT_IMAGE" \
  re_identification 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/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.

Supported Actions

The packaged Re-Identification PyT CLI supports train, evaluate, inference, export, and defaultspecs. This model skill exposes the runnable user actions train, evaluate, inference, and export; resume/retrain is performed through train with train.resumetrainingcheckpointpath.

Do not advertise or synthesize datasetconvert, deploy, prune, quantize, gentrt_engine, or standalone retrain for this model unless the packaged model skill and real CLI add those actions.

Training Requirements

• Dataset type: re_identification
• Formats: default
• Monitoring metric: cmcrank1, maximize

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.

S3_TRAIN = "s3://bucket/data/train"
CHECKPOINT = "/results/{train_job_id}/results_dir/model_epoch_000_step_00099.pth"

train (mandatory data sources):

{
    "train.num_epochs": 30,
    "train.checkpoint_interval": 10,
    "train.validation_interval": 10,
    "train.num_gpus": 1,
    "dataset.num_classes": 100,
    "dataset.num_workers": 4,
    "dataset.batch_size": 16,
    "dataset.num_instances": 4,
    "dataset.train_dataset_dir": f"{S3_TRAIN}/sample_train.tar.gz",
    "dataset.test_dataset_dir": f"{S3_TRAIN}/sample_test.tar.gz",
    "dataset.query_dataset_dir": f"{S3_TRAIN}/sample_query.tar.gz",
}

resume train (mandatory checkpoint):

{
    "train.num_epochs": 31,
    "train.resume_training_checkpoint_path": CHECKPOINT,
    "dataset.num_classes": 100,
    "dataset.batch_size": 16,
    "dataset.num_instances": 4,
    "dataset.train_dataset_dir": f"{S3_TRAIN}/sample_train.tar.gz",
    "dataset.test_dataset_dir": f"{S3_TRAIN}/sample_test.tar.gz",
    "dataset.query_dataset_dir": f"{S3_TRAIN}/sample_query.tar.gz",
}

evaluate (mandatory data sources and checkpoint):

{
    "evaluate.test_dataset": f"{S3_TRAIN}/sample_test.tar.gz",
    "evaluate.query_dataset": f"{S3_TRAIN}/sample_query.tar.gz",
    "evaluate.checkpoint": CHECKPOINT,
    "evaluate.output_cmc_curve_plot": "/results/{evaluate_job_id}/results_dir/cmc_curve.png",
    "evaluate.output_sampled_matches_plot": "/results/{evaluate_job_id}/results_dir/sampled_matches.png",
}

export (mandatory checkpoint and output):

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

inference (mandatory data sources and checkpoint):

{
    "inference.test_dataset": f"{S3_TRAIN}/sample_test.tar.gz",
    "inference.query_dataset": f"{S3_TRAIN}/sample_query.tar.gz",
    "inference.checkpoint": CHECKPOINT,
    "inference.output_file": "/results/{inference_job_id}/results_dir/reid_inference.json",
}

For export and inference, provide explicit file paths for export.onnxfile and inference.outputfile. For evaluate, provide explicit file paths for evaluate.outputcmccurveplot and evaluate.outputsampledmatchesplot. Keep these as spec values or specparams mappings; do not declare them as file outputs in skillinfo.yaml for local Docker until the runner distinguishes files from folders during output pre-creation.

Eval Dataset

Required. Evaluation requires test and query datasets for retrieval-based metrics (CMC, mAP).

Important Parameters

• dataset.num_classes: Number of identities. Default 751. Must match the number of unique identities in training data.
• model.backbone: Default resnet_50.
• optim.base_lr: Base learning rate. Default 3.5e-4.
• dataset.batch_size: Per-GPU batch size. Default 64. Re-ID benefits from large batches for better triplet/contrastive sampling.
• dataset.num_instances: Number of instances per identity in a batch. Controls sampling strategy for metric learning.

Multi-GPU / Multi-Node

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

• Multi-GPU strategy: ddpfindunusedparameterstrue
• sync_batchnorm is always enabled
• Precision forced to FP16 (16-mixed)
• No explicit num_nodes config — single-node oriented

Hardware

Minimum 1 GPU(s), recommended 2 GPU(s). 16GB+ VRAM per GPU. Re-ID models are relatively lightweight but benefit from large batch sizes for metric learning.

Error Patterns

numclasses mismatch: Ensure dataset.numclasses equals the number of unique identity folders in the training set.

Invalid triplet batch shape: dataset.batchsize must be compatible with dataset.numinstances so each mini-batch can be reshaped for hard-example mining. For local AutoML smoke runs, keep dataset.batchsize fixed to a known valid multiple such as 16 with dataset.numinstances: 4, and tune train.optim.base_lr instead of unconstrained batch size.

Query/gallery mismatch: Query and test (gallery) datasets must share the same identity namespace.

PyTorch 2.6 checkpoint load failure on checkpoint consumers: Current Re-ID checkpoints include OmegaConf containers. For checkpoints produced by the same trusted TAO train/AutoML workflow, set TORCHFORCENOWEIGHTSONLY_LOAD=1 in downstream resume, evaluate, inference, and export job env vars so Lightning/PyTorch can load the full checkpoint. Do not use this env var for untrusted checkpoints.

AutoML metric extraction: Re-ID train status files report retrieval KPIs such as cmcrank1, cmcrank5, cmcrank10, and mAP, plus train loss. Default AutoML train launches must optimize cmcrank1 with direction: maximize; do not use val_loss as the metric for this model.

Checkpoint handoff: Use the checkpoint resolver on the best AutoML child job's resultsdir/train/ folder and select the action-appropriate modelepoch*.pth checkpoint. Re-ID also writes reidmodel_latest.pth, but that is a latest symlink and should only be used when a caller explicitly requests latest. Preserve the same dataset identity count and query/gallery archives for downstream actions.

Default spec generation: The packaged defaultspecs CLI action does not consume the normal -e <spec.yaml> experiment file for resultsdir. Invoke it with a Hydra override such as reidentification defaultspecs resultsdir=/workspace/run/results/defaultspecs. Passing only -e leaves cfg.resultsdir unset and fails with MissingMandatoryValue: resultsdir.

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 re_identification.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.