Modernize OpenVINO-based Nuclio functions and allow them to run on Kubernetes (#6129)
Currently, OpenVINO-based functions assume that a local directory will be mounted into the container. In Kubernetes, that isn't possible, so implement an alternate approach: create a separate base image and inherit the function image from it. In addition, implement some modernizations: * Upgrade the version of OpenVINO to the latest (2022.3). Make the necessary updates to the code. Note that 2022.1 introduced an entirely new inference API, but I haven't switched to it yet to minimize changes. * Use the runtime version of the Docker image as the base instead of the dev version. This significantly reduces the size of the final image (by ~3GB). * Replace the `faster_rcnn_inception_v2_coco` model with `faster_rcnn_inception_resnet_v2_atrous_coco`, as the former has been removed from OMZ. * Ditto with `person-reidentification-retail-0300` -> `0277`. * The IRs used in the DEXTR function are not supported by OpenVINO anymore (format too old), so rewrite the build process to create them from the original code/weights instead.
Showing
想要评论请 注册 或 登录