# cog/keypoint — Keypoint training, preview, export

Train a generic keypoint detector, preview augmented data, run a dry-run, or
export the latest checkpoint to ONNX.

## Usage

```bash
# Preview a batch of augmented patches
s6 cog keypoint --config ./cfg/ds.json --preview-data

# Dry‑run (log graph and run one training step)
s6 cog keypoint --config ./cfg/ds.json --dry-run

# Train for 50 epochs and export an ONNX model
s6 cog keypoint --config ./cfg/ds.json --train -e 50 -b 16 -lr 1e-3 \
  --deploy model.onnx

# Restore latest checkpoint before continuing training
s6 cog keypoint --config ./cfg/ds.json --train --restore latest
```

## How it works
- Loads dataset/augmentation from a JSON config via
  `s6.app.cog.augmented_dataset.AugmentedKeypointDataset.from_json`.
- Uses `s6.nn.backbones.keypoints.GenericKeypointModel(in_channels=1)` with
  Smooth L1 loss on normalized keypoints in `[0, 1)`.
- Logs to TensorBoard under `logs/cog` unless `--no-tb` is set.
- Saves checkpoints under `checkpoints/` and optionally exports ONNX with
  outputs named `keypoints` and `heatmaps`.

## Key flags
- `--config` — path to dataset config JSON (required)
- `-t/--train` — run training; `--dry-run` for one step + graph
- `-e/--epochs`, `-b/--batch_size`, `-lr/--learning_rate`
- `--deploy` — ONNX filename to export after training
- `--restore` — checkpoint path or `latest` to resume
- `--no-tb` — disable TensorBoard logging; `--log-dir` to change root