Training the model
==================

With a labelled dataset in hand, training is a guided flow on the
**Train** page: lock in a dataset version, pick an architecture, and
hand the run off to Roboflow's servers.

The dataset version
-------------------

Before training, Roboflow builds a dataset *version* -- a frozen
snapshot of the images plus two transforms applied on the way in:

* **Preprocessing** resizes every image to the resolution the model
  trains at. Keep that resolution small: the camera runs small
  models, and a detector trained at a modest resolution fits the
  camera's memory and runs fast.
* **Augmentation** synthesizes extra training images by perturbing
  the originals -- flips, brightness and exposure shifts, blur,
  noise. Each augmentation teaches the model to tolerate a real
  variation it will meet on the camera, which stretches a small
  hand-captured dataset much further.

.. figure:: figures/augmentation.jpg
   :class: framed
   :width: 100%
   :alt: Roboflow's saturation augmentation settings, previewing the original image alongside reduced and increased saturation versions

   An augmentation preview: each option shows what it
   does to a sample image before you commit it to the
   version.

Match the augmentations to variations the camera will actually see.
Brightness and exposure earn their place -- lighting changes
constantly. Skip ones that never happen in your setup; a camera
bolted in place never sees a vertical flip, so flip augmentation
only dilutes the dataset.

Choosing an architecture
------------------------

Next, pick the model architecture. Roboflow offers several, each
with a size selector trading accuracy against speed.

.. figure:: figures/architecture.jpg
   :class: framed
   :width: 100%
   :alt: Roboflow's Select Architecture page with Roboflow RF-DETR, YOLO26, Roboflow 3.0, and YOLOv11 options, each with a model-size dropdown

   The architecture choices -- each with a size selector
   trading accuracy against inference speed.

For the camera, pick **Roboflow 3.0**. It is YOLOv8 under the hood,
and the camera ships a YOLOv8 post-processor in
:mod:`ml.postprocessing.ultralytics`, so its output decodes with no
extra code on your side. Choose the **Fast** size -- it fits the
camera's memory and runs at a usable frame rate.

Running the training
--------------------

Start the run and training happens on Roboflow's servers -- usually
well under an hour for a small dataset, with an email when it is
done.
The version page then shows the training graphs and the accuracy
metrics: mAP, precision, and recall.

.. figure:: figures/metrics.jpg
   :class: framed
   :width: 100%
   :alt: Roboflow's trained-model version page showing the metrics panel with mAP, precision, recall, and F1, above the Deploy Your Model section

   The trained model with its accuracy metrics. From
   here, the *Visualize* page also runs it on test
   images or a webcam for a quick sanity check.

If the numbers are good, the model is ready to deploy. If not, the
fix is usually more or more varied data -- capture another clip,
label it, and train a new version.