OpenVINO and Detectron

[Shevchuk__Mikhailo]

Whoever tried to convert a model from Detectron,  can share how to do it.

I asked about this example: 

https://docs.openvinotoolkit.org/latest/_inference_engine_ie_bridges_python_sample_instance_segmentation_demo_README.html

[Shubha_R_Intel]

Dear Shevchuk, Mikhailo:

The document below tells you exactly how to generate IR for maskrcnn_r50_fpn_2x as described:

https://docs.openvinotoolkit.org/latest/_docs_MO_DG_prepare_model_convert_model_tf_specific_Convert_Object_Detection_API_Models.html

Please select your Mask RCNN model from one of OpenVino's validated models here:

https://docs.openvinotoolkit.org/latest/_docs_MO_DG_prepare_model_convert_model_Convert_Model_From_TensorFlow.html

Thanks !

Shubha

[Juj__Satyajith]

Dear Shubha,

That specific example is for a tensorflow model, right? If I remember correctly, model optimisation does not currently support the Detectron framework.

Correct me if I am wrong.

[Shubha_R_Intel]

Dear folks, This OpenVino documentation definitely mentions Detectron, so it is supported. The key though is to convert a *.pkl file to a *.pb first, since Model Optimizer does not support *.pkl.  There is a script in the Detectron github repo which enables conversion:

https://github.com/facebookresearch/Detectron/blob/master/tools/convert_pkl_to_pb.py

Thanks,

Shubha