tao-train-action-recognition

The full skill.

Original instructions from the publisher’s SKILL.md

# Action Recognition

> **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).

Action recognition from video sequences. Supports RGB, optical flow, and joint (multi-stream) input types for classifying temporal actions in video clips.

Set model.pretrained_model_path for pretrained backbone weights.

## Quick Start (docker run)

Docker-native launch — no TAO SDK and no Python on the host. Use the local
Docker/platform skill instead when it gives a stricter environment-specific
command (non-root UID mapping, cache redirects, remote daemons).

```bash
TAO_PYT_IMAGE_DEFAULT=nvcr.io/nvidia/tao/tao-toolkit:7.2.0-pyt  # versions-key: images.tao_toolkit.pyt
TAO_PYT_IMAGE="${TAO_PYT_IMAGE:-$TAO_PYT_IMAGE_DEFAULT}"
RUN_ROOT="${RUN_ROOT:-$PWD}"
DOCKER_COMMON=(
  --rm --gpus all --shm-size=8g
  --shm-size=8g
  --ulimit memlock=-1
  --ulimit stack=67108864
  -v "$RUN_ROOT/data:/data:ro"
  -v "$RUN_ROOT/specs:/specs:ro"
  -v "$RUN_ROOT/results:/results"
)
```

Train:

```bash
docker run "${DOCKER_COMMON[@]}" "$TAO_PYT_IMAGE" \
  action_recognition train -e /specs/train.yaml
```

Evaluate:

```bash
docker run "${DOCKER_COMMON[@]}" "$TAO_PYT_IMAGE" \
  action_recognition evaluate -e /specs/evaluate.yaml
```

Inference:

```bash
docker run "${DOCKER_COMMON[@]}" "$TAO_PYT_IMAGE" \
  action_recognition inference -e /specs/inference.yaml
```

Export:

```bash
docker run "${DOCKER_COMMON[@]}" "$TAO_PYT_IMAGE" \
  action_recognition export -e /specs/export.yaml
```

Every action takes its spec with `-e`; `results_dir` is set in the spec or
overridden on the command line. Mount any pretrained-weights directory the spec
references, and keep every in-container path consistent across actions.

## 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:** action_recognition
- **Formats:** default
- **Training monitoring metrics:** `val_loss`, `val_acc`
- **Evaluate task metrics:** `accuracy`, `m_accuracy`
- **AutoML metric contract:** for the required evaluation-backed baseline and
  final comparison, use `accuracy` with maximize direction and run `evaluate`
  through `eval_fn` for every recommendation. `val_loss` is suitable only for
  an explicitly accepted training-proxy run because `evaluate` does not emit
  it. Scratch runs with no starting checkpoint require a minimal default train
  job followed by evaluation of its exact epoch/step checkpoint for the
  baseline.

### Per-Action Dataset Requirements

| Action | Spec Key | Source | Files | List? |
|---|---|---|---|---|
| evaluate | evaluate.test_dataset_dir | train_datasets | test/ extracted from test.tar.gz | No |
| inference | inference.inference_dataset_dir | train_datasets | test/smile/ extracted from test/smile.tar.gz | No |
| train | dataset.train_dataset_dir | train_datasets | train/ extracted from train.tar.gz | No |
| train | dataset.val_dataset_dir | train_datasets | test/ extracted from test.tar.gz | No |

### 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
LOCAL_DATA = "/workspace/data/extracted"
```

If the source dataset is provided as the TAO sample archives
`train.tar.gz`, `test.tar.gz`, or `test/smile.tar.gz`, download and extract
them before launching the TAO container. The action-recognition entrypoints
expect directory paths and fail with `NotADirectoryError` when these spec keys
point at `.tar.gz` files.

**train (mandatory data sources):**
```python
{
    "train.num_epochs": 30,
    "train.checkpoint_interval": 10,
    "train.validation_interval": 10,
    "train.num_gpus": 1,
    "dataset.label_map": {
        "catch": 0,
        "smile": 1
    },
    "dataset.batch_size": 2,
    "dataset.train_dataset_dir": f"{LOCAL_DATA}/train",
    "dataset.val_dataset_dir": f"{LOCAL_DATA}/test",
}
```

**evaluate (mandatory data sources):**
```python
{
    "dataset.label_map": {
        "catch": 0,
        "smile": 1
    },
    "evaluate.test_dataset_dir": f"{LOCAL_DATA}/test",
}
```

**inference (mandatory data sources):**
```python
{
    "dataset.label_map": {
        "catch": 0,
        "smile": 1
    },
    "inference.inference_dataset_dir": f"{LOCAL_DATA}/smile_infer/smile",
}
```

**export (mandatory checkpoint + output path):**
```python
{
    "export.checkpoint": "<selected train checkpoint>",
    "export.onnx_file": "<results_dir>/action_recognition.onnx",
}
```

For direct local-docker chaining without the SDK resolver, select the concrete
checkpoint produced by training, for example
`model_epoch_000_step_00005.pth`, and pass that exact file to `evaluate`,
`inference`, and `export`. Do not use the `ar_model_latest.pth` symlink unless
the user explicitly requests latest-checkpoint behavior. For resume training,
set `train.resume_training_checkpoint_path` to the exact epoch/step checkpoint
being resumed.
## Eval Dataset

Optional. Test dataset may be distributed as `test.tar.gz` separate from
training; extract it and point the spec to the extracted `test/` directory.
TAO training emits `val_loss` and `val_acc` for the packaged sample data, while
the evaluate action emits `accuracy` and `m_accuracy`. Use `accuracy` with
maximize direction for the normal evaluation-backed AutoML workflow. Use
`val_loss` with minimize direction only when the user explicitly accepts a
training-only proxy without the required impact baseline.

## Important Parameters

- **model.model_type**: Input type: rgb, of (optical flow), or joint (multi-stream).
- **model.backbone**: Default resnet_18. Used as the spatial feature extractor.
- **dataset.label_map**: Dictionary mapping class names to indices.
- **model.rgb_seq_length**: Number of frames per clip for RGB input.
- **model.of_seq_length**: Number of frames for optical flow input.
- **train.optim.lr**: Learning rate. Default 5e-4.

## 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 oriented

## Hardware

Minimum 1 GPU(s), recommended 2 GPU(s). 16GB+ VRAM per GPU. Memory depends on sequence length and input resolution. batch_size=2 is conservative for video data.

## Error Patterns

**Sequence length mismatch**: Ensure video clips have enough frames for the configured rgb_seq_length or of_seq_length.

**Evaluate/inference missing label map**: Downstream actions rebuild the
ActionRecognitionModel before loading the checkpoint, so they need the same
`dataset.label_map` used during training. Include it with every evaluate or
inference spec; otherwise model construction fails before the checkpoint can be
validated.

## 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 `action_recognition.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 | `results_dir` | `output_dir` | current job results directory |
| train | `encryption_key` | `key` | encryption key |
| train | `model.of_pretrained_model_path` | `ptm_if_no_resume_model` | PTM when no resume checkpoint exists |
| train | `model.rgb_pretrained_model_path` | `ptm_if_no_resume_model` | PTM when no resume checkpoint exists |
| train | `results_dir` | `output_dir` | current job results directory |
| 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.