Tao Train Segformer

SegFormer

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).

SegFormer for semantic segmentation. Lightweight transformer-based architecture with hierarchical feature extraction. Efficient for real-time segmentation tasks.

Set model.backbone.pretrained_backbone_path for backbone weights.

For TAO Deploy TensorRT actions (gen_trt_engine, TensorRT evaluate, and TensorRT inference), read references/tao-deploy-segformer.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.

Supported Actions

The packaged SegFormer PyT CLI supports train, evaluate, export, inference, quantize, and default_specs. This model skill exposes train, evaluate, export, inference, and quantize; resume/retrain is performed through train with train.resume_training_checkpoint_path.

The parent PyT CLI does not expose gen_trt_engine. Use models/segformer/deploy for TensorRT engine generation, TensorRT evaluation, and TensorRT inference.

Training Requirements

• Dataset type: segmentation
• Formats: unet
• AutoML training metric: val_miou (maximize).
• Standalone evaluation metric: test_miou. Use the train-status val_miou for AutoML ranking and the evaluator's test_miou only for checkpoint verification.

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.

SEG_TRAIN_ROOT = "/data/segformer/train"
SEG_EVAL_ROOT = "/data/segformer/eval"
SEG_INFER_ROOT = "/data/segformer/infer"
CAL_IMAGES = f"{SEG_TRAIN_ROOT}/images/train"

train (mandatory data sources):

{
    "train.num_gpus": 1,
    "train.num_epochs": 10,
    "train.checkpoint_interval": 10,
    "train.validation_interval": 10,
    "dataset.segment.batch_size": 4,
    "dataset.segment.root_dir": SEG_TRAIN_ROOT,
}

evaluate (mandatory data sources):

{
    "evaluate.batch_size": 4,
    "dataset.segment.root_dir": SEG_EVAL_ROOT,
    "evaluate.checkpoint": CHECKPOINT,
}

inference (mandatory data sources):

{
    "dataset.segment.batch_size": 1,
    "dataset.segment.root_dir": SEG_INFER_ROOT,
    "inference.checkpoint": CHECKPOINT,
}

export (mandatory data sources):

{
    "dataset.segment.root_dir": SEG_TRAIN_ROOT,
    "export.checkpoint": CHECKPOINT,
    "export.input_height": 256,
    "export.input_width": 256,
    "export.onnx_file": ONNX_FILE,
}

quantize (mandatory data sources):

{
    "dataset.segment.root_dir": SEG_TRAIN_ROOT,
    "dataset.segment.quant_calibration_dataset.images_dir": CAL_IMAGES,
    "quantize.model_path": CHECKPOINT,
}

If the source dataset is delivered as separate images/*.tar.gz and masks/*.tar.gz archives, extract them before launch so root_dir contains directories such as images/train, images/val, images/test, masks/train, and masks/val. Do not point dataset.segment.root_dir at an archive staging folder that still contains only tarballs.

Eval Dataset

Optional. Validation data is typically part of the root_dir structure.

Important Parameters

• dataset.segment.num_classes: Number of segmentation classes. Default 2 (binary). Must match the number of classes in your mask annotations.
• model.backbone.type: Default fan_small_12_p4_hybrid. Supported includes FAN variants, SegFormer MIT variants, and others.
• dataset.segment.root_dir: Root directory of the segmentation dataset.
• dataset.segment.img_size: Input image size. Default 256. Increase for finer segmentation at the cost of memory.
• train.optim.lr: Learning rate. Default 6e-5.
• model.freeze_backbone: Whether to freeze the backbone during training. Useful for fine-tuning with limited data.
• dataset.segment.batch_size: Per-GPU batch size. Default 8.
• dataset.segment.label_transform: Use the string "None" when no label transform is desired. Do not set this to JSON/YAML null; strict schema merge treats the field as a string enum.
• dataset.segment.palette: For grayscale masks, use one integer per RGB entry, for example rgb: [85]. Preserve the dataset's actual label ids and class names rather than normalizing them unless the user explicitly asks for a conversion.

Multi-GPU / Multi-Node

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

• Multi-GPU strategy: ddp_find_unused_parameters_true
• No fsdp support

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

Hardware

Minimum 1 GPU(s), recommended 2 GPU(s). 16GB+ (V100 or A100) VRAM per GPU. SegFormer is relatively lightweight. Default img_size=256 is memory-friendly. Increase img_size for higher resolution at the cost of memory and speed.