Tao Train Optical Inspection

Optical Inspection

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).

Optical inspection for defect detection using Siamese networks. Compares image pairs to detect manufacturing defects, anomalies, or quality issues.

Set train.pretrained_model_path for pretrained Siamese weights.

For TAO Deploy TensorRT actions (gen_trt_engine, TensorRT evaluate, and TensorRT inference), read references/tao-deploy-optical-inspection.md first. The parent PyT container does not expose optical_inspection gen_trt_engine; TensorRT engine generation is deploy-only. 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: optical_inspection
• Formats: default
• AutoML training metric: val_acc, with direction=maximize
• Standalone evaluation metric: test_acc, with direction=maximize

Per-Action Dataset Requirements

images.tar.gz is the transfer artifact. After staging/extraction, dataset.*_dataset.images_dir must point at the inner directory that directly contains golden/ and the board directories used by the CSV.

dataset.csv Contract

The Optical Inspection loader reads the CSV by column name and constructs both sides of every Siamese comparison as:

<images_dir>/<input_path>/<object_name>_<lighting><image_ext>
<images_dir>/<golden_path>/<object_name>_<lighting><image_ext>

Absolute input_path and golden_path values are also accepted by the loader; an absolute value replaces images_dir. Relative directories are recommended because they remain portable after staging. Do not put a filename in either path column and do not put a lighting suffix or extension in object_name.

Additional metadata columns are allowed but ignored by this loader. Empty values are invalid. One complete row, grounded in the production layout, is:

input_path,golden_path,label,object_name
690-5G190-0510-001P1/AOI_B/FXLH_690-5G190-0510-001P1_30332_P_AOI_B_20230317130332/PerComponent,golden/images/690-5G190-0510-001P1BOT/,PASS,C1018@1

With the standard four-light configuration, each row requires eight files:

<images_dir>/
├── golden/images/690-2G133-0210-000BOT/
│   ├── R821@1_LowAngleLight.jpg
│   ├── R821@1_SolderLight.jpg
│   ├── R821@1_UniformLight.jpg
│   └── R821@1_WhiteLight.jpg
└── 690-2G133-0210-000/AOI_B/<capture>/PerComponent/
    ├── R821@1_LowAngleLight.jpg
    ├── R821@1_SolderLight.jpg
    ├── R821@1_UniformLight.jpg
    └── R821@1_WhiteLight.jpg

The three selection fields must agree:

• input_map keys are the exact filename lighting suffixes. Current loader behavior iterates the YAML key insertion order; it does not sort by the integer values. Keep values contiguous and in matching order (0..N-1).
• num_input must equal the number of input_map entries. The loader opens every key, so a mismatched value does not limit the file list and produces a tensor/export shape mismatch.
• concat_type: linear stacks inputs in key order along image height. With four 128×128 inputs this produces a 512×128 tensor; grid_map is ignored.
• concat_type: grid requires an even num_input and grid_map.x * grid_map.y == num_input. Placement is row-major in key order. For the standard 2 x 2 map: LowAngle is upper-left, Solder upper-right, Uniform lower-left, and White lower-right.

If the dataset genuinely contains only *_SolderLight.jpg, use num_input: 1, input_map: {SolderLight: 0}, and concat_type: linear. Do not declare four inputs when three variants are absent.

Run the packaged preflight before train, evaluate, or inference, using the same dataset settings as the spec:

python3 skills/models/tao-train-optical-inspection/scripts/validate_dataset.py \
  --csv /data/optical-inspection/train/dataset.csv \
  --images-dir /data/optical-inspection/train/images \
  --num-input 4 \
  --concat-type grid \
  --grid-x 2 --grid-y 2

For a custom map, repeat --input-map LIGHT=INDEX in YAML key order and pass the spec's --image-ext. The validator reports missing columns, unsafe PASS case, unresolvable row directories, and every missing lighting file with its CSV row number. A two-row path-stub fixture is under tests/fixtures/dataset/valid/; it verifies the contract and is not PCB training data.

No published Optical Inspection sample dataset is discoverable from this skill bank. Obtain converted AOI data from the dataset owner for your product or organization, then stage it at the paths mounted into the container. Do not use the committed validator fixture for model training.

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.

TRAIN_ROOT = "/data/optical-inspection/train"
EVAL_ROOT = "/data/optical-inspection/eval"
INFERENCE_ROOT = "/data/optical-inspection/inference"

These are example in-container mount points after obtaining and staging data from its owner; they are not download locations.

train (mandatory data sources):

{
    "train.num_epochs": 30,
    "train.checkpoint_interval": 10,
    "train.validation_interval": 10,
    "train.num_gpus": 1,
    "dataset.batch_size": 8,
    "dataset.train_dataset.images_dir": f"{TRAIN_ROOT}/images",
    "dataset.train_dataset.csv_path": f"{TRAIN_ROOT}/dataset.csv",
    "dataset.validation_dataset.images_dir": f"{EVAL_ROOT}/images",
    "dataset.validation_dataset.csv_path": f"{EVAL_ROOT}/dataset.csv",
    "dataset.test_dataset.images_dir": f"{EVAL_ROOT}/images",
    "dataset.test_dataset.csv_path": f"{EVAL_ROOT}/dataset.csv",
}

evaluate (mandatory data sources):

{
    "evaluate.checkpoint": "<selected train/AutoML checkpoint>",
    "dataset.test_dataset.images_dir": f"{EVAL_ROOT}/images",
    "dataset.test_dataset.csv_path": f"{EVAL_ROOT}/dataset.csv",
}

Use the workflow's checkpoint resolver for downstream actions instead of guessing a filename. For Optical Inspection smoke runs, AutoML may produce model_epoch_000_step_00006.pth; resume can then produce model_epoch_001_step_00012.pth. Best-checkpoint actions should use the AutoML best child job's selected checkpoint, epoch-specific actions should pass the exact epoch/step checkpoint requested, and only explicit "latest" requests should resolve to the latest checkpoint.

export:

{
    "export.checkpoint": "<selected train/AutoML checkpoint>",
    "export.onnx_file": "/results/optical_inspection.onnx",
    "export.input_width": 128,
    "export.input_height": 512,
    "export.batch_size": 1,
}

inference (mandatory data sources):

{
    "inference.checkpoint": "<selected train/AutoML checkpoint>",
    "dataset.infer_dataset.images_dir": f"{INFERENCE_ROOT}/images",
    "dataset.infer_dataset.csv_path": f"{INFERENCE_ROOT}/dataset.csv",
}