Tao Train Oneformer

OneFormer

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).

OneFormer for universal image segmentation. Unifies panoptic, instance, and semantic segmentation with a single architecture using task-conditioned queries.

Set train.pretrained_backbone and/or train.pretrained_model.

For TAO Deploy TensorRT actions (gen_trt_engine, TensorRT evaluate, and TensorRT inference), read references/tao-deploy-oneformer.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
• AutoML training metric: mIoU, with direction=maximize
• Standalone evaluation metric: test_mIoU, with direction=maximize

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"
S3_INFERENCE = "s3://bucket/data/inference"
S3_CALIBRATION = "s3://bucket/data/calibration"

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,
    "train.precision": "32",
    "dataset.train.images": f"{S3_TRAIN}/images.tar.gz",
    "dataset.train.annotations": f"{S3_TRAIN}/annotations.json",
    "dataset.label_map": f"{S3_TRAIN}/label_map.json",
    "dataset.train.panoptic": f"{S3_TRAIN}/images_panoptic.tar.gz",
    "dataset.val.images": f"{S3_EVAL}/images.tar.gz",
    "dataset.val.annotations": f"{S3_EVAL}/annotations.json",
    "dataset.val.panoptic": f"{S3_EVAL}/images_panoptic.tar.gz",
    "dataset.test.images": 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.images": f"{S3_TRAIN}/images.tar.gz",
    "dataset.label_map": f"{S3_TRAIN}/label_map.json",
    "dataset.train.annotations": f"{S3_TRAIN}/annotations.json",
    "dataset.train.panoptic": f"{S3_TRAIN}/images_panoptic.tar.gz",
    "dataset.val.images": f"{S3_EVAL}/images.tar.gz",
    "dataset.val.annotations": f"{S3_EVAL}/annotations.json",
    "dataset.val.panoptic": f"{S3_EVAL}/images_panoptic.tar.gz",
    "dataset.test.images": f"{S3_EVAL}/images.tar.gz",
    "dataset.test.annotations": f"{S3_EVAL}/annotations.json",
    "dataset.test.panoptic": f"{S3_EVAL}/images_panoptic.tar.gz",
}

export:

{
    "export.checkpoint": "<selected train/AutoML checkpoint>",
    "model.sem_seg_head.num_classes": 133,
    "model.export": True,
    "export.onnx_file": "/results/oneformer_export_640.onnx",
}

inference (mandatory data sources):

{
    "inference.checkpoint": "<selected train/AutoML checkpoint>",
    "dataset.train.images": f"{S3_TRAIN}/images.tar.gz",
    "dataset.label_map": f"{S3_TRAIN}/label_map.json",
    "dataset.train.annotations": f"{S3_TRAIN}/annotations.json",
    "dataset.train.panoptic": f"{S3_TRAIN}/images_panoptic.tar.gz",
    "dataset.val.images": f"{S3_EVAL}/images.tar.gz",
    "dataset.val.annotations": f"{S3_EVAL}/annotations.json",
    "dataset.val.panoptic": f"{S3_EVAL}/images_panoptic.tar.gz",
    "dataset.test.images": f"{S3_INFERENCE}/images.tar.gz",
    "inference.images_dir": f"{S3_INFERENCE}/images.tar.gz",
}

quantize (mandatory data sources):

{
    "quantize.model_path": "<selected train/AutoML checkpoint>",
    "dataset.train.images": f"{S3_TRAIN}/images.tar.gz",
    "dataset.train.annotations": f"{S3_TRAIN}/annotations.json",
    "dataset.label_map": f"{S3_TRAIN}/label_map.json",
    "dataset.train.panoptic": f"{S3_TRAIN}/images_panoptic.tar.gz",
    "dataset.val.images": f"{S3_EVAL}/images.tar.gz",
    "dataset.val.annotations": f"{S3_EVAL}/annotations.json",
    "dataset.val.panoptic": f"{S3_EVAL}/images_panoptic.tar.gz",
    "dataset.test.images": f"{S3_EVAL}/images.tar.gz",
    "dataset.quant_calibration_dataset.images_dir": f"{S3_CALIBRATION}/images.tar.gz",
}

Checkpoint Selection

OneFormer training writes epoch-step checkpoints such as model_epoch_000_step_00017.pth and may also write a oneformer_model_latest.pth symlink. For checkpoint-dependent actions, use the model-skill or SDK parent-model resolver and pass the exact selected checkpoint path into evaluate.checkpoint, inference.checkpoint, export.checkpoint, quantize.model_path, or train.resume_training_checkpoint_path. Do not pick the oneformer_model_latest.pth symlink by name unless the user explicitly asks for latest checkpoint behavior. If the resolver reports a best checkpoint, use that best checkpoint for evaluation/export/inference; if the user asks for a specific epoch or step, use the matching epoch-step checkpoint.

Eval Dataset

Optional. Val data configured alongside train in the dataset config.

Important Parameters

• model.sem_seg_head.num_classes: Number of segmentation class indices available to the head. Default 133 for COCO panoptic data when dataset.contiguous_id: True remaps raw category ids through the label map. Do not shrink this to a global workflow class count unless the label map and annotations have actually been reduced to that class set.
• model.one_former.hidden_dim: Keep at 256 for local smoke runs unless the text encoder width is changed in lock-step. Reducing hidden_dim alone causes a text feature/context dimension mismatch during training.
• model.backbone.name: Default D2SwinTransformer (Swin-based). embed_dim=192, depths=[2,2,18,2] by default.
• train.num_epochs: Default 50 — significantly higher than most TAO models. OneFormer needs more epochs for convergence.
• train.optim.lr: Learning rate. Default 1e-5. Lower than Mask2Former's 2e-4.
• model.task_toggling: Enable/disable specific tasks: semantic_on, instance_on, panoptic_on.
• export.task: Export task mode. Options: semantic, instance, panoptic. Default semantic. Export input defaults to 640x640.
• inference.mode: Inference mode. Options: semantic, instance, panoptic. Default semantic. image_size defaults to [1024, 1024].
• evaluate.iou_per_class: Report per-class IoU in evaluation. Default True.

Multi-GPU / Multi-Node

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

• Uses explicit DDPStrategy with find_unused_parameters=True, gradient_as_bucket_view=True, process_group_backend="nccl"
• sync_batchnorm is always enabled
• No fsdp support — DDP only

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

Hardware

Minimum 2 GPU(s), recommended 4 GPU(s). 24GB+ (A100 recommended) VRAM per GPU. OneFormer is memory-intensive like Mask2Former. batch_size=1 is the default. Multi-GPU needed for reasonable training speed, especially with 50 epochs.