Tao Train Mask Auto Encoder

MAE

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).

MAE (Masked Autoencoder) for self-supervised pretraining and fine-tuning. Masks random patches and reconstructs them to learn visual representations. Supports pretrain and finetune stages.

Set train.pretrained_model_path for pretrained MAE weights when fine-tuning.

For TAO Deploy TensorRT actions (gen_trt_engine), read references/tao-deploy-mask-auto-encoder.md first. Deploy spec templates live in this skill's references/ folder with the spec_template_deploy_*.yaml prefix.

The parent PyTorch mae CLI supports train, evaluate, inference, and export. Build TensorRT engines through the deploy workflow, not the model skill.

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/spec_template_<action>.yaml from the schema top-level default field. AutoML enablement is declared at the model layer in references/skill_info.yaml via automl_enabled. Runnable AutoML for an action requires schemas/<action>.schema.json and references/spec_template_<action>.yaml to exist and parse. Use the packaged selected-action schema for automl_default_parameters, automl_disabled_parameters, 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/skill_info.yaml and resolve the run override from either an explicit automl_policy value or the user's workflow request. Use automl_policy: 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 automl_policy: off for this run only. When automl_policy: on, automl_enabled: true, and both schemas/train.schema.json and references/spec_template_train.yaml are packaged, route the train action through tao-skill-bank:tao-run-automl by default with this model's skill_dir. Preserve workflow/application overrides for datasets, specs, output directories, GPU/platform settings, parent checkpoints, and automl_policy. 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: image_classification
• Formats: ssl
• Accepted dataset intents: training, evaluation, testing
• Pretraining monitoring metric: train_loss, minimized.
• AutoML metric contract: for fine-tuning, use ACC_all with maximize direction. Fine-tune validation and the packaged evaluate action emit ACC_all (plus val_loss); they do not emit train_loss. Use ACC_all for trial and final checkpoint selection when train.stage: finetune.

Per-Action Dataset Requirements

For SDK/app job inputs, the images_*.tar.gz archives are uploaded as the action inputs. For direct local Docker runs against host-mounted data, extract the archives first and point dataset.train_data_sources, dataset.val_data_sources, and dataset.test_data_sources at the extracted images_train, images_val, and images_test folders. Passing a local tar path directly to the MAE CLI can produce a zero-sample dataloader because the local dataloader does not unpack that archive path.

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

train (mandatory data sources):

{
    "dataset.train_data_sources": f"{S3_TRAIN}/images_train.tar.gz",
    "dataset.val_data_sources": f"{S3_EVAL}/images_val.tar.gz",
    "train.num_epochs": 10,
    "train.optim.lr": 2e-4,
}

evaluate (mandatory data sources):

{
    "dataset.val_data_sources": f"{S3_EVAL}/images_val.tar.gz",
    "evaluate.checkpoint": "<selected train/AutoML checkpoint>",
    "train.stage": "finetune",
}

inference (mandatory data sources):

{
    "dataset.test_data_sources": f"{S3_EVAL}/images_test.tar.gz",
    "inference.checkpoint": "<selected train/AutoML checkpoint>",
    "train.stage": "finetune",
}

Eval Dataset

Optional. Pretraining does not need eval data. Fine-tuning optionally uses val set.

Important Parameters

• train.stage: Training stage. Options: pretrain, finetune. Pretrain learns representations via masking. Finetune adds a classification head.
• model.arch: Architecture. Default convnextv2_base. For local smoke AutoML, use convnextv2_atto rather than unsupported names such as vit_tiny_patch16. Supported families include vit_base_patch16 and larger ViTs, ConvNeXtV2 atto/femto/pico/nano/tiny/base/large/huge, and Hiera tiny/small/base/large/huge.
• model.num_classes: Number of classes for fine-tuning. Default 1000 (ImageNet). Only relevant in finetune stage.
• model.mask_ratio: Fraction of patches to mask during pretraining. Typically 0.75.
• model.norm_pix_loss: Whether to normalize pixel values in reconstruction loss.
• dataset.augmentation.input_size: Keep the local smoke profile at 224 for ConvNeXtV2 MAE. Reducing to 112 can make the MAE mask grid incompatible with feature-map dimensions.
• MAE does not expose a dataset.workers spec field. Do not add it to smoke-test overrides; Hydra rejects unknown dataset keys before training.
• train.optim.lr: Learning rate. Default 2e-4.
• dataset.augmentation: Augmentation settings including mixup, cutmix for fine-tuning.