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

train (mandatory data sources):

evaluate (mandatory data sources):

export:

inference (mandatory data sources):

quantize (mandatory data sources):

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.

Error Patterns

CUDA out of memory: batch_size is already 1. Reduce image resolution or use a smaller Swin configuration.

Extracted S3 tarball points one level too high: For local Docker runs, images.tar.gz and images_panoptic.tar.gz may extract wrapper directories such as images/ and images_panoptic/. Set dataset.*.images, dataset.*.panoptic, inference.images_dir, and quantization calibration paths to the actual folder containing image or panoptic files, not the wrapper directory. A one-level-too-high path fails with FileNotFoundError for the first annotation image even though recursive file counts look correct.

default_specs missing results_dir: The CLI default_specs subtask ignores -e experiment specs for results_dir; pass a Hydra-style override instead: oneformer default_specs results_dir=/path/to/default_specs.

Invalid Lightning precision fp32: Use train.precision: "32" in train/AutoML/evaluate/inference specs. The current Lightning stack rejects the legacy fp32 string.

PyTorch 2.6 checkpoint load failure on downstream actions: Current OneFormer checkpoints include OmegaConf objects. For checkpoints produced by the same trusted TAO train/AutoML workflow, set TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1 in downstream evaluate, inference, export, quantize, or resume job env vars so Lightning can load the full checkpoint. Do not use this env var for untrusted checkpoints.

CUDA device-side assert in matcher/class cost: If training fails in oneformer/utils/matcher.py while indexing out_prob[:, tgt_ids], compare the effective target ids with model.sem_seg_head.num_classes. The packaged COCO panoptic sample has 133 compact classes after dataset.contiguous_id: True remapping, so use model.sem_seg_head.num_classes: 133 even when a broader validation workflow passes a smaller generic num_classes value. Only use a smaller class count when the label map and annotations are reduced to that exact contiguous class set.

Inference returns PASS with no predictions: OneFormer prediction reads inference.images_dir, not dataset.test.images. Declare and populate inference.images_dir with the image folder or tarball for every inference run. dataset.test.images may still be useful for shared dataset context, but it does not drive the PyTorch predict dataloader.

Export output path pre-created as a directory: Do not declare export.onnx_file as a file output. The OneFormer exporter asserts that the ONNX path does not already exist, while the local runner pre-creates declared output paths. Set export.onnx_file explicitly in the spec to a non-existing file path under the mounted results tree. Keep the default 640x640 export shape for smoke validation; very small export shapes can trigger PyTorch ONNX shape-inference failures.

Quantize cannot find the training label map from an AutoML checkpoint: OneFormer Lightning checkpoints retain train-time absolute dataset paths in their saved hparams. When running downstream actions from an AutoML child checkpoint, keep the parent AutoML job directory accessible at its original /results/<job_id> path inside the action container in addition to passing the resolved checkpoint path. Otherwise quantize can fail while loading checkpoint hparams even when the current spec includes a valid dataset.label_map.

Slow training: 50 default epochs with batch_size=1 is slow on single GPU. Use multi-GPU distributed training.

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 oneformer.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.

Deployment

• tao-deploy-oneformer