# Tao Train Optical Inspection

> Optical Inspection for defect detection using Siamese networks. Compares image pairs to detect manufacturing defects, anomalies, or quality issues. Use when training, evaluating, exporting, or running inference for a TAO Optical Inspection model on AOI / quality-control data. Trigger phrases include "train optical inspection", "AOI defect detection", "Siamese defect classifier", "PCB / manufacturing inspection".

- Skill: `nvidia-tao/tao-train-optical-inspection` (Agent Skill, multi-file: 51 files)
- Install (CLI): `npx skillmds@latest add nvidia-tao/tao-train-optical-inspection`
- Raw SKILL.md: https://api.skillmd.com/api/skills/nvidia-tao/tao-train-optical-inspection/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- License: Apache-2.0
- Author: NVIDIA-TAO (https://skillmd.com/u/nvidia-tao)
- Updated: 2026-09-10
- Page: https://skillmd.com/skills/nvidia-tao/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

| Action | Spec Key | Source | Files | List? |
|---|---|---|---|---|
| evaluate | dataset.test_dataset.images_dir | eval_dataset | images.tar.gz | No |
| evaluate | dataset.test_dataset.csv_path | eval_dataset | dataset.csv | No |
| inference | dataset.infer_dataset.images_dir | inference_dataset | images.tar.gz | No |
| inference | dataset.infer_dataset.csv_path | inference_dataset | dataset.csv | No |
| train | dataset.train_dataset.images_dir | train_datasets | images.tar.gz | No |
| train | dataset.train_dataset.csv_path | train_datasets | dataset.csv | No |
| train | dataset.validation_dataset.images_dir | eval_dataset | images.tar.gz | No |
| train | dataset.validation_dataset.csv_path | eval_dataset | dataset.csv | No |
| train | dataset.test_dataset.images_dir | eval_dataset | images.tar.gz | No |
| train | dataset.test_dataset.csv_path | eval_dataset | dataset.csv | No |

`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:

```text
<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`.

| Column | Type | Required | Allowed values and meaning |
|---|---|---|---|
| `input_path` | string directory path | yes | Board/capture component directory, relative to `images_dir` or absolute. |
| `golden_path` | string directory path | yes | Golden/reference component directory, relative to `images_dir` or absolute. |
| `label` | string | yes | Exact case-sensitive `PASS` means non-defective (class 0). Any other non-empty defect name means defective (class 1), for example `missing`, `shift`, `excess_solder`, `lifted_lead`, `polarity`, `tombstone`, or `upside down`. `pass` is invalid because the loader would silently treat it as a defect. |
| `object_name` | string filename stem | yes | Component identifier such as `C1018@1`; no directory, `_SolderLight`, or `.jpg`. The loader forces this column to string so numeric-looking identifiers retain their text form. |

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

```csv
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:

```text
<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:

```bash
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`.

```python
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):**
```python
{
    "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):**
```python
{
    "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:**
```python
{
    "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):**
```python
{
    "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",
}
```

## Dataset Convert

Dataset conversion is optional for Optical Inspection. If the dataset is already in TAO-ready Optical Inspection format, start directly from the `images.tar.gz` plus `dataset.csv` splits and run `train`, `evaluate`, `inference`, and downstream checkpoint/export/deploy actions on that converted data.

The PyT container exposes `optical_inspection dataset_convert`, but this model skill does not package a `dataset_convert` action/template. The converter expects the raw Factory PCB layout (`root_dataset_dir`, train/val/all PCB directories, `golden_csv_dir`, `project_name`, and `bot_top`). Data owners may instead provide preconverted Optical Inspection `images.tar.gz` plus `dataset.csv` splits without the raw PCB/golden CSV source. Do not synthesize a fake PCB dataset. In model validation reports, mark dataset conversion as `not run: preconverted dataset provided` rather than failed or blocked when only converted data is available.

When using preconverted transfer archives locally, verify the extracted directory before writing specs. The archives may unpack an `images/` wrapper directory; point `dataset.*.images_dir` at the inner directory that contains `golden/` and the board/image folders referenced by `dataset.csv`, for example `.../<split>/images/images`, not the outer wrapper.

Product-side follow-ups remain: publish a licensed, trainable sample dataset and
package the existing container `dataset_convert` entrypoint as a skill action
with its own schema/template. Neither is implemented by this documentation and
preflight change.

## Eval Dataset

Optional. Eval dataset uses same format (images + CSV).

## Important Parameters

- **model.model_type**: Siamese variant. Options include Siamese, Siamese_3.
- **model.model_backbone**: Default custom.
- **model.embedding_vectors**: Number of embedding dimensions. Default 5.
- **train.optim.lr**: Learning rate. Default 5e-4.
- **dataset.batch_size**: Training batch size. Must be greater than 1; use `2` or higher for minimal smoke runs.
- **dataset.num_input**: Number of input images per comparison.
- **dataset.input_map**: Mapping of input channels / image pairs.

## Multi-GPU / Multi-Node

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

| Spec Key | Description | Default |
|----------|-------------|---------|
| `train.num_gpus` | Number of GPUs | 1 |
| `train.gpu_ids` | GPU device indices | [0] |

- Strategy: `auto` (Lightning picks best strategy automatically)
- No explicit `num_nodes` or `distributed_strategy` config — single-node only
- Lightweight Siamese network, single GPU typically sufficient

## Hardware

Minimum 1 GPU(s), recommended 1 GPU(s). 8GB+ VRAM per GPU. Siamese networks for inspection are lightweight. Single GPU sufficient.

## Error Patterns

**CSV format error**: Require `input_path,golden_path,label,object_name`, then
run `scripts/validate_dataset.py` with the spec's `images_dir`, lighting map,
`num_input`, concatenation, grid, and image extension before launch.

**Extracted image root mismatch**: If train, evaluate, or inference cannot find paths from `dataset.csv`, inspect the extracted `images.tar.gz` tree. The TAO-ready root must contain `golden/` plus the board folders referenced in the CSV. For transferred archives this can be one level below the extraction target, such as `images/images`.

**Training batch size assertion**: The Optical Inspection dataloader rejects
`dataset.batch_size: 1` for train. Keep the template default of 8 for normal
runs, or set `dataset.batch_size: 2` for minimal AutoML smoke validation.

**PyTorch checkpoint load failure on downstream actions**: For checkpoints
produced by the same trusted TAO train/AutoML workflow, set
`TORCH_FORCE_NO_WEIGHTS_ONLY_LOAD=1` for evaluate, inference, export, and
resume jobs if the current PyTorch default blocks loading the full checkpoint.
Do not use this env var for untrusted checkpoints.

## 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 `optical_inspection.config.json`:

| Action | Spec Field | Inference Function | Meaning |
|---|---|---|---|
| evaluate | `encryption_key` | `key` | encryption key |
| evaluate | `evaluate.checkpoint` | `parent_model` | model file inferred from the parent job results folder |
| evaluate | `results_dir` | `output_dir` | current job results directory |
| export | `encryption_key` | `key` | encryption key |
| export | `export.checkpoint` | `parent_model` | model file inferred from the parent job results folder |
| export | `export.onnx_file` | `create_onnx_file` | output ONNX path |
| export | `results_dir` | `output_dir` | current job results directory |
| inference | `encryption_key` | `key` | encryption key |
| inference | `inference.checkpoint` | `parent_model` | model file inferred from the parent job results folder |
| inference | `inference.trt_engine` | `parent_model` | model file inferred from the parent job results folder |
| inference | `results_dir` | `output_dir` | current job results directory |
| train | `encryption_key` | `key` | encryption key |
| train | `results_dir` | `output_dir` | current job results directory |
| train | `train.pretrained_model_path` | `ptm_if_no_resume_model` | PTM when no resume checkpoint exists |
| train | `train.resume_training_checkpoint_path` | `resume_model` | model file inferred from the current job results folder |

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-optical-inspection](references/tao-deploy-optical-inspection.md)