ML Recog

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).

Metric learning recognition for fine-grained visual recognition. Learns embeddings for retrieval-based matching (e.g., retail product recognition). Uses triplet/contrastive losses.

Set model.pretrainedmodelpath for pretrained backbone.

For TAO Deploy TensorRT actions (gentrtengine, TensorRT evaluate, and TensorRT inference), read references/tao-deploy-metric-learning-recognition.md first. Deploy spec templates live 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 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: ml_recog
• Formats: default
• Monitoring metric: val Precision at Rank 1

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"

train (mandatory data sources):

{
    "train.num_epochs": 30,
    "train.checkpoint_interval": 10,
    "train.validation_interval": 10,
    "train.num_gpus": 1,
    "dataset.train_dataset": f"{S3_TRAIN}/metric_learning_recognition/retail-product-checkout-dataset_classification_demo/known_classes/train.tar.gz",
    "dataset.val_dataset": {"reference": f"{S3_TRAIN}/metric_learning_recognition/retail-product-checkout-dataset_classification_demo/known_classes/reference.tar.gz", "query": f"{S3_TRAIN}/metric_learning_recognition/retail-product-checkout-dataset_classification_demo/known_classes/val.tar.gz"},
}

evaluate (mandatory data sources):

{
    "evaluate.checkpoint": "<selected train/AutoML checkpoint>",
    "dataset.val_dataset": {"reference": f"{S3_TRAIN}/metric_learning_recognition/retail-product-checkout-dataset_classification_demo/unknown_classes/reference.tar.gz", "query": f"{S3_TRAIN}/metric_learning_recognition/retail-product-checkout-dataset_classification_demo/unknown_classes/test.tar.gz"},
}

inference (mandatory data sources):

{
    "inference.checkpoint": "<selected train/AutoML checkpoint>",
    "dataset.val_dataset": {"reference": f"{S3_TRAIN}/metric_learning_recognition/retail-product-checkout-dataset_classification_demo/unknown_classes/reference.tar.gz"},
    "inference.input_path": f"{S3_TRAIN}/metric_learning_recognition/retail-product-checkout-dataset_classification_demo/unknown_classes/test.tar.gz",
}

Eval Dataset

Required. Evaluation requires reference and query datasets for retrieval metrics.

Important Parameters

• model.backbone: Default resnet50. Options: resnet50, resnet101, fansmall, fanbase, fanlarge, fantiny, nvdinov2vitlargelegacy.
• model.feat_dim: Embedding dimension. Default 256. Output feature vector size for similarity matching.
• train.batchsize: Per-GPU batch size. Default 4. valbatchsize also 4. For training and AutoML search, train.batchsize must be divisible by dataset.num_instance.
• dataset.numinstance: Instances per identity in a batch (P/K sampling). Default 4. Controls how many images of the same class appear together. If using a custom AutoML range for train.batchsize, use explicit options that are multiples of this value.
• train.optim.trunk.base_lr: Learning rate for the trunk (backbone). Default 3.5e-4 (Adam).
• train.optim.embedder.base_lr: Learning rate for the embedding head. Default 3.5e-4.
• train.optim.tripletlossmargin: Margin for triplet loss. Default 0.3. smooth_loss=True by default.
• train.optim.minerfunctionmargin: Hard mining margin. Default 0.1. Controls pair mining difficulty.
• train.optim.steps: LR decay steps. Default [40, 70] with gamma=0.1.
• dataset.train_dataset: Path to training images organized in class folders.
• dataset.val_dataset: Dict with 'reference' and 'query' keys pointing to ImageNet-format directories for retrieval evaluation.

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 oriented

Hardware

Minimum 1 GPU(s), recommended 2 GPU(s). 16GB+ VRAM per GPU. Metric learning benefits from larger batch sizes for better triplet sampling but is otherwise moderate on memory.

Error Patterns

Reference/query mismatch: Ensure reference and query datasets share compatible class namespaces for evaluation.

PyTorch 2.6 checkpoint load failure on checkpoint actions: Current TAO ML-Recog checkpoints may contain OmegaConf objects. For checkpoints produced by the same trusted TAO train/AutoML workflow, set TORCHFORCENOWEIGHTSONLY_LOAD=1 in downstream evaluate, inference, export, or resume/retrain job env vars so Lightning can load the full checkpoint. Do not use this env var for untrusted checkpoints.

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 ml_recog.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-metric-learning-recognition](references/tao-deploy-metric-learning-recognition.md)