# model

Python API: `luxonis_train.config.predefined_models.fomo.v1.model`

The FOMO detection model.

## Classes

### FOMOModel

FOMO, which predicts one point for each object instead of a box.

The head predicts a heatmap with one channel for each class. Each cell of the heatmap covers a patch of the image, and a high
value marks the center of an object. The dataset needs bounding box labels, and the loss takes the centers of the boxes as
targets.

 * `Throughput:`: Frames per second at 384x512. * `light`: 140 on RVC2, 243 on RVC4
    * `heavy`: 34 on RVC2, 230 on RVC4

> **Notes**
> `object_weight` in `loss_params` multiplies the loss of the cells that hold an object, and its default is `500`. A larger value raises the recall and also the number of false positives.

`use_nms` in `head_params` is on by default. At evaluation, the head then keeps only the cells that are the maximum of their 3x3
window. This removes some false positives, and also some true positives when two objects sit close together. At export, the head
applies a 3x3 max pooling to the heatmap.

`attach_index` in `head_params` selects the backbone stage the head reads, and its default is `1`. `0` reads the stage before it.
That stage gives a larger heatmap and a more precise position, and the model runs slower.

> **Example**
> The `model` section of a config:

```yaml
model:
  predefined_model:
    name: FOMOModel
    params:
      variant: light
```

 * `Components:`: * Nodes:
   [EfficientRep](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/backbones/efficientrep/efficientrep.md)
   ->
   [FOMOHead](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/heads/fomo_head.md)
    * Losses:
      [FOMOLocalizationLoss](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/attached_modules/losses/fomo_localization_loss.md)
    * Metrics:
      [ConfusionMatrix](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/attached_modules/metrics/confusion_matrix/confusion_matrix.md)
    * Visualizers:
      [FOMOVisualizer](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/attached_modules/visualizers/fomo_visualizer.md)
    * Main metric:
      [ConfusionMatrix](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/attached_modules/metrics/confusion_matrix/confusion_matrix.md)
    * Variants: * `light`
       * `heavy`

#### Methods

##### init

```python
def __init__(**kwargs):
```

Initialize the model with its default components.

`ConfusionMatrix` is already the main metric, so set `enable_confusion_matrix=False` to avoid adding a second one.

##### get_variants

```python
def get_variants() -> tuple[str, dict[str, Params]]:
```

Get the default variant name and the available variants.

The default is `light`. The variants differ in the backbone:

 * `light`:
   [EfficientRep](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/backbones/efficientrep/efficientrep.md)
   with the `"n"` variant;
 * `heavy`:
   [MobileNetV2](https://docs.luxonis.com/software-v3/ai-inference/model-source/training/luxonis-train/luxonis-train-api-reference/nodes/backbones/mobilenetv2.md)
   without a variant.

Both variants set `n_conv_layers` to `2` and `conv_channels` to `16` in `head_params`. A `head_params` given in the config
replaces the whole dictionary of the variant. The head then takes its own defaults for the keys the dictionary leaves out.

> **Example**
> ```pycon
>>> default, variants = FOMOModel.get_variants()
>>> default
'light'
>>> variants["heavy"]
{'backbone': 'MobileNetV2',
 'head_params': {'n_conv_layers': 2, 'conv_channels': 16}}
```

Returns

 * `tuple[str, dict[str, Params]]`: `"light"` and the two variants with their constructor arguments.