Converting RetinaFace model to OpenVINO

I am trying to accelerate the face detection model RetinaFace on CPU. So I deployed the optimization tool OpenVINO. However, after successfully converting the mxnet model to the openvino mode (Intermediate Representation), the speed of the openvino model is much slower than that of the pure mxnet deployment. (~2s vs ~0.001s)

Is there anyone that has the experience converting this model to openvino? I would like to know the reason why.

No file, test, or entry point is named. Start by reproducing the RetinaFace conversion and CPU benchmarks described in the report, then compare the OpenVINO and MXNet deployment setups. Done means identifying and documenting the cause of the roughly 2-second versus 0.001-second runtime difference, with a verified correction if one is found.

Written by the indexing model from the issue text.