Deploying to the camera
=======================

The trained model lives on Roboflow's servers. Getting it onto the
camera takes one download, then the same steps as loading any other
model.

Downloading the weights
-----------------------

On the **Deployments** page, choose **Deploy to 3rd Party
Platforms** and select the **OpenMV** tab. It downloads the model's
weights as a single integer-quantized ``.tflite`` file, named after
the project and version -- the int8 format the camera's TFLite
engine runs.

.. figure:: figures/deploy.jpg
   :class: framed
   :width: 100%
   :alt: Roboflow's "Deploy to 3rd Party Platforms" dialog with the OpenMV tab selected and a Download Files button

   The OpenMV deploy target -- **Download Files** saves
   the camera-ready ``.tflite`` weights.

Loading it on the camera
------------------------

Add the ``.tflite`` file to the camera with the IDE's :doc:`ROMFS
editor <../ide/romfs>`, which converts it for the board's NPU when
the board has one, then load it in a script with ``ml.Model``.
Models also run from the camera's flash drive -- copy the file over
and point ``ml.Model`` at the path -- but ROMFS is the better home:
models there execute straight from flash without a RAM copy.

A detection model's raw output is a tensor of box coordinates and
class scores that still needs decoding. Roboflow's YOLO-family
detectors decode with the post-processors the camera ships in
:mod:`ml.postprocessing.ultralytics`, so a few lines wire the model
to its decoder and you have boxes and labels.

.. seealso::

   The :doc:`machine learning chapter <../../ml/index>` for running
   models with the :mod:`ml` module -- loading, the inference
   pipeline, and the :doc:`walkthrough of decoding YOLO-family
   output <../../ml/postprocessing/yolov8-walkthrough>`.