MAL (Mask Auto-Label) for weakly-supervised segmentation. Produces segmentation masks from minimal annotations (e.g., point or box annotations). Uses ViT-MAE backbone.

Set train.pretrained_model_path for ViT-MAE pretrained weights.

Generated TAO Core schemas are packaged in , with listing available actions. Each generated schema also emits from the schema top-level field. AutoML enablement is declared at the model layer in via . Runnable AutoML still requires and to exist and parse. Use the packaged train schema for , , defaults, min/max bounds, enums, option weights, math conditions, dependencies, and popular parameters. Do not expect at runtime; maintainers regenerate schemas/templates before packaging the skill bank.

This model is AutoML-enabled at the model layer. Before handling any train-stage request, read and resolve the run override from either an explicit value or the user's workflow request. Treat phrases like "turn off AutoML", "disable AutoML", "no HPO", or "plain training" as for this run only; otherwise default to . When , , and both and are packaged, route the train action through by default with this model's . Preserve workflow/application overrides for datasets, specs, output directories, GPU/platform settings, parent checkpoints, and . Use direct model training only when 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 , , , and deploy flows stay in this model skill. The per-run override does not change model metadata.

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 .

S3_TRAIN = "s3://bucket/data/train"
S3_EVAL = "s3://bucket/data/eval"

{
    "train.num_gpus": 1,
    "train.gpu_ids": [
        0
    ],
    "train.num_epochs": 5,
    "train.checkpoint_interval": 5,
    "train.validation_interval": 5,
    "dataset.train_img_dir": f"{S3_TRAIN}/images.tar.gz",
    "dataset.train_ann_path": f"{S3_TRAIN}/annotations.json",
    "dataset.val_img_dir": f"{S3_EVAL}/images.tar.gz",
    "dataset.val_ann_path": f"{S3_EVAL}/annotations.json",
}

{
    "dataset.val_img_dir": f"{S3_EVAL}/images.tar.gz",
    "dataset.val_ann_path": f"{S3_EVAL}/annotations.json",
}

{
    "inference.img_dir": f"{S3_EVAL}/images.tar.gz",
    "inference.ann_path": f"{S3_EVAL}/annotations.json",
}

Optional. Val images and annotations configured alongside train paths.

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

train.num_gpus

train.gpu_ids

train.num_nodes

Multi-node env vars (set by orchestrator): , , , , .

Minimum 1 GPU(s), recommended 2 GPU(s). 24GB+ (A100 recommended) VRAM per GPU. ViT-MAE backbone at crop_size=512 needs 24GB+ GPU memory.

CUDA out of memory: Reduce model.crop_size (512 -> 384 -> 256) or use a smaller ViT-MAE variant (base vs large).

Model-specific inference mappings belong in this MD file, not in . Generated runners should read this section and apply the mappings with SDK helpers before . This mirrors the old microservices flow.

Inference mappings from TAO Core :

evaluate.checkpoint

parent_model

results_dir

output_dir

inference.checkpoint

parent_model

inference.label_dump_path

create_inference_result_file_mal

results_dir

output_dir

results_dir

output_dir

For or , pass the upstream train/export/AutoML child job id as . 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 and do not patch generated runner scripts to guess checkpoint paths.