Converting MXNet RetinaFace-MobileNet0.25 model to IR produces different output results on OpenVINO

[Gouveia__César]

Hi,

I wanted to convert MXNet RetinaFace-MobileNet0.25 model (https://github.com/deepinsight/insightface/tree/master/RetinaFace#third-party-models) to OpenVINO IR and I got that by using the following command:
python deployment_tools\model_optimizer\mo_mxnet.py --input_model mxnet_models\RetinaFace\mnet.25-0000.params --input_shape (1,3,640,640) --output_dir openvino_models
and it was successfully converted.

After that I tested inference on both mxnet model and openvino model to check if the outputs were the same. However the output sizes for MXNet are:
* 88 for width and height;
* 44 for width and height;
* 22 for width and height.
And the OpenVINO output sizes are:
* 80 for width and height;
* 40 for width and height;
* 20 for width and height.

Thanks,
César.

[samontab]

The first thing I would check is to make sure all the layers of the model are supported by OpenVINO.

Here you can see the details about the MXNet layers

Then it would be a good idea to check layer by layer if the input/outputs are correct or at which point there's something wrong.

Good luck!

[Gouveia__César]

Thanks for your response @samontab .

All the layers are apparently supported otherwise the conversion of the model would not have been possible. I can provide both the json and params from mxnet and bin and xml from openvino if needed.

Thanks,
César.

[IntelSupport]

Hi Cesar,

MobileNet0.25 topology is not officially supported by OpenVINO and RetinaFace-MobileNet0.25 model has not been validated for inferencing performance as well.

More information is available at the following page:

https://docs.openvinotoolkit.org/2020.4/openvino_docs_MO_DG_prepare_model_convert_model_Convert_Model_From_MxNet.html#supported_topologies

Also, I would encourage you to try out Intel® Distribution of OpenVINO™ Toolkit version 2020.4, which is a vastly improved version with latest features and leading performance.

Regards,

Munesh

[qiaoq]

why the input_shape is (1,3,640,640)