MAL

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).

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.pretrainedmodelpath for ViT-MAE 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" \
  mal train -e /specs/train.yaml

Evaluate:

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

Inference:

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

Training Requirements

• Dataset type: segmentation
• Formats: default
• Monitoring metric: mIoU

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 specoverrides. MAL expects COCO-style annotation JSON plus image paths that match the JSON filename entries after the data source is prepared. Archive-only CSV/image datasets are not compatible unless they are converted to this format first.

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

train (mandatory data sources):

{
    "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",
}

evaluate (mandatory data sources):

{
    "evaluate.checkpoint": "<selected train/AutoML checkpoint>",
    "dataset.val_img_dir": f"{S3_EVAL}/images.tar.gz",
    "dataset.val_ann_path": f"{S3_EVAL}/annotations.json",
}

inference (mandatory data sources):

{
    "inference.checkpoint": "<selected train/AutoML checkpoint>",
    "inference.img_dir": f"{S3_EVAL}/images.tar.gz",
    "inference.ann_path": f"{S3_EVAL}/annotations.json",
}

For checkpoint-dependent actions, use the model resolver declared in references/skillinfo.yaml. Select the exact epoch/step checkpoint requested by the user or the best checkpoint when a best-checkpoint action is requested. The malmodel_latest.pth symlink is only appropriate when the user explicitly asks for the latest checkpoint.

Eval Dataset

Optional. Val images and annotations configured alongside train paths.

Important Parameters

• model.arch: ViT-MAE backbone variant. Default vit-mae-base/16.

Avoid vit-deit-tiny/16; the current runtime rejects tiny ViT variants.

• train.lr: Learning rate. Default 1e-6 (very low — fine-tuning ViT).
• dataset.crop_size: Training crop size. Default 512. Use this key, not

model.crop_size.

• train.warmup_epochs: Warmup epochs before full learning rate.
• model.load_mask: Whether to load pre-computed masks.

AutoML / HPO Notes

For MAL AutoML launches, keep the default smoke search space narrow and pass automl_hyperparameters=["train.lr", "train.wd"]. Use conservative Bayesian ranges around the ViT-MAE fine-tuning defaults, for example train.lr from 1e-7 to 1e-5 and train.wd from 1e-5 to 1e-2. The packaged train schema marks these two parameters as the default AutoML parameters; pass them explicitly when using a runtime that still derives MAL search metadata from its bundled config module.

Multi-GPU / Multi-Node

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

• Multi-GPU strategy: ddpfindunusedparameterstrue
• No fsdp support
• LR auto-scaling: lr = lr numdevices batchsize (learning rate is scaled automatically by device count and batch size)

Multi-node env vars (set by orchestrator): WORLDSIZE, NODERANK, MASTERADDR, MASTERPORT, NUMGPUPER_NODE.

Hardware

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.

Error Patterns

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

Key cropsize not in MALModelConfig: The crop-size override was placed under model.cropsize. Move it to dataset.crop_size.

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 mal.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.