=========
OpenVINO
=========

We use the OpenVINO toolkit to run our YOLO models, such that the model can be used in the OpenVINO inference engine, which runs on the CPU, GPU or our Intel Neural Compute Stick2.

Convert Darknet to OpenVINO models
==================================

Convert weights to Tensorflow format
------------------------------------
1. Clone `GibHub:tensorflow-yolo-v3 <https://github.com/mystic123/tensorflow-yolo-v3>`_ to your PC (no OpenVINO needed).
2. Checkout commit ``ed60b90`` if you have problems.
3. Run the command::

    python3 convert_weights_pb.py --class_names <PATH to names.names file> --data_format NHWC --weights_file <PATH_TO_YOUR_CURRENT_YOLO_WEIGHTS_FILE> --tiny

The variables are the following:
   - The required ``names.names`` file consists of all class names, with each one in its own line.
   - The ``<PATH_TO_YOUR_CURRENT_YOLO_WEIGHTS_FILE>`` represents the Darknet ``.weights`` file.
   - The ``--tiny`` setting stands for ``tiny-yolo`` which is the currently used architecture.
   - The output file will be written to the current folder as a ``frozen_darknet_yolov3_model.pb`` file.


Install OpenVINO
----------------
You can skip this step, if you have OpenVINO installed or access to a robot.

Otherwise visit the `OpenVINO Install Page <https://docs.openvinotoolkit.org/latest/_docs_install_guides_installing_openvino_linux.html>`_ and follow the instructions.


.. Tip:: If you do not use Ubuntu, e.g. "Arch Linux", there is also a `OpenVINO Docker container <https://docs.openvinotoolkit.org/latest/openvino_docs_install_guides_installing_openvino_docker_linux.html>`_ provided by Intel.


Convert Tensorflow format to OpenVINO
-------------------------------------
1. Source OpenVINO in your shell. Run ``source /opt/intel/openvino/bin/setupvars.sh``.
2. Create a config like the `Example .json config`_ below.
3. Run::

    python3 /opt/intel/openvino/deployment_tools/model_optimizer/mo_tf.py --input_model <PATH_TO_YOUR_FROZEN_TENSORFLOW_MODEL> --output_dir <YOUR_OUTPUT_DIR> --data_type FP16 --batch 1 --tensorflow_use_custom_operations_config <PATH_TO_YOUR_CONFIG_JSON>

The variables are the following:
   - The tensorflow model path is set in the ``<PATH_TO_YOUR_FROZEN_TENSORFLOW_MODEL>`` part.
   - The model output path is saved in ``<YOUR_OUTPUT_DIR>``.
   - The precision is set by replacing ``FP16`` with e.g. ``FP32``.
   - Set the path to your freshly created ``.json`` config file (see `Example .json config`_) in ``<PATH_TO_YOUR_CONFIG_JSON>``.

Example ``.json`` config
------------------------

In Darknet, these values are located in the ``.cfg`` files.

.. code-block:: json

    [
    {
        "id": "TFYOLOV3",
        "match_kind": "general",
        "custom_attributes": {
        "classes": 2,
        "anchors": [17, 27, 8,140, 16,124, 37, 56, 18,211, 32,218, 83,105, 66,288, 142,185],
        "coords": 4,
        "num": 9,
        "masks": [[3, 4, 5], [0, 1, 2]],
        "entry_points": ["detector/yolo-v3-tiny/Reshape", "detector/yolo-v3-tiny/Reshape_4"]
        }
    }
    ]