# attached_modules

Python API: `luxonis_train.attached_modules`

Losses, metrics, and visualizers that attach to a node.

An attached module reads the output packet of one node and the labels of the batch.
[BaseLoss](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/attached_modules/losses/base_loss.md),
[BaseMetric](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/attached_modules/metrics/base_metric.md),
and
[BaseVisualizer](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/attached_modules/visualizers/base_visualizer.md)
are the base classes of the three kinds.
[BaseAttachedModule](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/attached_modules/base_attached_module.md)
is the base class of all three. A config lists the modules of a node under the `losses`, `metrics`, and `visualizers` keys of the
node.

The parameter names of `forward`, or of `update` for a metric, select the inputs. For example, `predictions` gets the main
prediction of the task of the node, and `target` gets the only label of the task.
[BaseAttachedModule.get_parameters](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/attached_modules/base_attached_module.md)
gives all the rules. For the label formats, see
[luxonis_train.loaders](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/loaders.md).

Prediction packets use these shapes in evaluation:

 * Classification: `[B, n_classes]` logits.
 * Segmentation and anomaly detection: `[B, C, H, W]` logits.
 * Embeddings: `[B, D]`.
 * OCR: `[B, T, n_classes]` logits.
 * Bounding boxes: one `[M_i, 6]` tensor per image, with rows `[x1, y1, x2, y2, score, class]` in pixels.
 * Keypoints: one `[M_i, n_keypoints, 3]` tensor per image, with `(x, y, confidence)` in pixels.
 * Instance masks: one `[M_i, H, W]` tensor per image.
 * FOMO: `[B, n_classes, H_f, W_f]` heatmap logits.

Detection heads skip non-maximum suppression during training, so their packets do not yet contain the final boxes, keypoints, or
instance masks. The heads document their packet keys;
[luxonis_train.loaders](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/loaders.md)
documents the matching label formats.

## Child Pages

 * [losses](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/attached_modules/losses.md)
 * [metrics](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/attached_modules/metrics.md)
 * [visualizers](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/attached_modules/visualizers.md)
 * [base_attached_module](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/attached_modules/base_attached_module.md)