Imported from nvidia/skills (skills/tao-train-mask2former/SKILL.md) via skills.sh. Install upstream with npx skills add nvidia/skills --skill tao-train-mask2former. Copyright stays with the author (Apache-2.0).

Mask2Former

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).

Mask2Former for universal image segmentation (panoptic, instance, and semantic). Transformer-based with masked attention for high-quality segmentation results.

Set model.backbone.pretrained_weights for Swin backbone weights.

For TAO Deploy TensorRT actions (gen_trt_engine, TensorRT evaluate, and TensorRT inference), read references/tao-deploy-mask2former.md first. Deploy spec templates live in this skill's references/ folder with the spec_template_deploy_*.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/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: segmentation
• Formats: coco_panoptic, coco
• 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 spec_overrides.

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

train (mandatory data sources):

{
    "train.num_gpus": 1,
    "train.num_epochs": 10,
    "train.checkpoint_interval": 10,
    "train.validation_interval": 10,
    "model.sem_seg_head.num_classes": 133,
    "dataset.contiguous_id": True,
    "dataset.train.type": "coco_panoptic",
    "dataset.val.type": "coco_panoptic",
    "dataset.test.type": "coco_panoptic",
    "dataset.train.img_dir": f"{S3_TRAIN}/images.tar.gz",
    "dataset.label_map": f"{S3_TRAIN}/label_map_panoptic.json",
    "dataset.train.instance_json": f"{S3_TRAIN}/annotations.json",
    "dataset.train.panoptic_json": f"{S3_TRAIN}/annotations_panoptic.json",
    "dataset.train.panoptic_dir": f"{S3_TRAIN}/images_panoptic.tar.gz",
    "dataset.val.img_dir": f"{S3_EVAL}/images.tar.gz",
    "dataset.val.instance_json": f"{S3_EVAL}/annotations.json",
    "dataset.val.panoptic_json": f"{S3_EVAL}/annotations_panoptic.json",
    "dataset.val.panoptic_dir": f"{S3_EVAL}/images_panoptic.tar.gz",
    "dataset.test.img_dir": f"{S3_EVAL}/images.tar.gz",
}

evaluate (mandatory data sources):

{
    "evaluate.checkpoint": "<selected train/AutoML checkpoint>",
    "model.sem_seg_head.num_classes": 133,
    "dataset.contiguous_id": True,
    "dataset.train.type": "coco_panoptic",
    "dataset.val.type": "coco_panoptic",
    "dataset.test.type": "coco_panoptic",
    "dataset.train.img_dir": f"{S3_TRAIN}/images.tar.gz",
    "dataset.label_map": f"{S3_TRAIN}/label_map_panoptic.json",
    "dataset.train.instance_json": f"{S3_TRAIN}/annotations.json",
    "dataset.train.panoptic_json": f"{S3_TRAIN}/annotations_panoptic.json",
    "dataset.train.panoptic_dir": f"{S3_TRAIN}/images_panoptic.tar.gz",
    "dataset.val.img_dir": f"{S3_EVAL}/images.tar.gz",
    "dataset.val.instance_json": f"{S3_EVAL}/annotations.json",
    "dataset.val.panoptic_json": f"{S3_EVAL}/annotations_panoptic.json",
    "dataset.val.panoptic_dir": f"{S3_EVAL}/images_panoptic.tar.gz",
    "dataset.test.img_dir": f"{S3_EVAL}/images.tar.gz",
}

export:

{
    "export.checkpoint": "<selected train/AutoML checkpoint>",
    "export.onnx_file": "<output ONNX path>",
    "model.sem_seg_head.num_classes": "<same value used for train>",
}

inference (mandatory data sources):

{
    "inference.checkpoint": "<selected train/AutoML checkpoint>",
    "model.sem_seg_head.num_classes": "<same value used for train>",
    "dataset.contiguous_id": True,
    "dataset.train.img_dir": f"{S3_TRAIN}/images.tar.gz",
    "dataset.label_map": f"{S3_TRAIN}/label_map_panoptic.json",
    "dataset.train.instance_json": f"{S3_TRAIN}/annotations.json",
    "dataset.train.panoptic_json": f"{S3_TRAIN}/annotations_panoptic.json",
    "dataset.train.panoptic_dir": f"{S3_TRAIN}/images_panoptic.tar.gz",
    "dataset.val.img_dir": f"{S3_EVAL}/images.tar.gz",
    "dataset.val.instance_json": f"{S3_EVAL}/annotations.json",
    "dataset.val.panoptic_json": f"{S3_EVAL}/annotations_panoptic.json",
    "dataset.val.panoptic_dir": f"{S3_EVAL}/images_panoptic.tar.gz",
    "dataset.test.img_dir": f"{S3_EVAL}/images.tar.gz",
}

quantize (mandatory data sources):

{
    "quantize.model_path": "<selected train/export artifact>",
    "dataset.train.img_dir": f"{S3_TRAIN}/images.tar.gz",
    "dataset.label_map": f"{S3_TRAIN}/label_map_panoptic.json",
    "dataset.train.instance_json": f"{S3_TRAIN}/annotations.json",
    "dataset.train.panoptic_json": f"{S3_TRAIN}/annotations_panoptic.json",
    "dataset.train.panoptic_dir": f"{S3_TRAIN}/images_panoptic.tar.gz",
    "dataset.val.img_dir": f"{S3_EVAL}/images.tar.gz",
    "dataset.val.instance_json": f"{S3_EVAL}/annotations.json",
    "dataset.val.panoptic_json": f"{S3_EVAL}/annotations_panoptic.json",
    "dataset.val.panoptic_dir": f"{S3_EVAL}/images_panoptic.tar.gz",
    "dataset.test.img_dir": f"{S3_EVAL}/images.tar.gz",
    "dataset.quant_calibration_dataset.images_dir": f"{S3_TRAIN}/images.tar.gz",
}

Eval Dataset

Optional. Val data sources are part of the dataset config alongside train.

Important Parameters

• model.sem_seg_head.num_classes: Number of segmentation classes. Default 200. Must match your annotation categories.
• model.backbone.swin.type: Swin Transformer variant. Default tiny. Options include tiny, small, base, large.
• model.mode: Segmentation mode. Default panoptic. Options: panoptic, instance, semantic.
• train.optim.lr: Learning rate. Default 2e-4 (AdamW).
• dataset.train.batch_size: Per-GPU batch size. Default 1. Mask2Former is memory-intensive due to per-pixel predictions.
• dataset.contiguous_id: If true, set model.sem_seg_head.num_classes to the number of label-map categories. If false, set model.sem_seg_head.num_classes above the maximum raw category id and keep the same setting for evaluate, inference, export, deploy, and quantize. The COCO panoptic S3 sample has 133 categories with raw ids up to 200, so raw-id validation uses num_classes: 201.

Multi-GPU / Multi-Node

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

• Same DDP/FSDP behavior as DINO (activation checkpoint aware)
• FAN backbones auto-enable sync_batchnorm
• fsdp forces FP16

Multi-node env vars (set by orchestrator): WORLD_SIZE, NODE_RANK, MASTER_ADDR, MASTER_PORT, NUM_GPU_PER_NODE.

Export / TRT Defaults

• TRT data types: FP32, FP16 only — INT8 is NOT supported
• The parent PyTorch mask2former CLI supports train, evaluate, inference, export, and quantize; run TensorRT engine generation, TensorRT inference, and TensorRT evaluation through references/tao-deploy-mask2former.md. Export semantic ONNX (model.mode: semantic) when validating TensorRT evaluation because the current deploy evaluator accepts semantic engines.
• Keep export input dimensions compatible with the deploy templates. The packaged default export.input_width: 960 and export.input_height: 544 exports and builds a TensorRT engine successfully; shrinking export to tiny validation-only sizes such as 128x128 can hit a PyTorch ONNX minus_one_pos != -1 shape-inference assertion before ONNX is produced.

Hardware

Minimum 1 GPU(s), recommended 4 GPU(s). 24GB+ (A100 recommended) VRAM per GPU. Mask2Former is memory-heavy. batch_size=1 is the default for good reason. Multi-GPU recommended for reasonable training speed.

Error Patterns

CUDA out of memory: batch_size is already 1 by default. Reduce image resolution in augmentation config or use a smaller Swin variant.

Panoptic vs instance format mismatch: Ensure you provide the correct annotation format matching model.mode setting.

Deploy schema error for top-level dataset.type: TAO Deploy uses dataset.val.type and dataset.test.type. Do not put dataset.type at the top level of Mask2Former deploy specs.

Export ONNX shape assertion at very small resolution: If export fails with minus_one_pos != -1 from PyTorch ONNX shape inference, restore the template export dimensions (960x544) before retrying deploy validation. Keep training and evaluation image sizes small when needed for quick smoke tests, but do not carry those tiny dimensions into export unless the target shape has been verified.

Quantize checkpoint load error: Older PyTorch images can fail checkpoint-based mask2former quantize because the runtime quantize script passes experiment_spec to Mask2formerPlModule.load_from_checkpoint instead of the required cfg argument. Images with the quantize fix support the default torchao checkpoint flow. ONNX quantization still requires backend: modelopt.onnx, mode: static_ptq, a fixed dataset.test.target_size, and an image that includes modelopt.onnx.quantization.

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 create_job(). This mirrors the old microservices infer_params.py flow.

Inference mappings from TAO Core mask2former.config.json:

For parent_model or parent_model_folder, pass the upstream train/export/AutoML child job id as parent_job_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.

When selecting a Mask2Former checkpoint outside the SDK resolver, match the intended epoch/step artifact exactly, for example model_epoch_000_step_00100.pth. The mask2former_model_latest.pth symlink is valid only when latest is explicitly requested.

Deployment

• tao-deploy-mask2former