Tensorflow Model Zoo item that I would to use on NCS2

[Mason__Mike]

Quick questions I would like to use one of the iNaturalist models (faster_rcnn_resnet101_fgvc) for bird recognition.  

1. Would this model run - if correctly run through the Model Optimizer, and

2. How would I go about creating the prototxt and model to gt it to run.  

(I'm struggling on the translation of models in the TF model zoo to IR format.

Thanks for any help.

[Shubha_R_Intel]

Dearest Mike,  OpenVino's Model Optimizer does support several varieties of Faster R-CNN. Please check the following online documentation and search for the text "Faster R-CNN" within the page.

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

Fortunately you don't need to create your own prototxt and trained model in most cases. You may simply convert an existing trained model into Model Optimizer IR. 

Under deployment_tools\intel_models there are also several models trained by Intel which you can use, though I don't think any of them are targeted toward bird recognition.

Using Model Optimizer is often as simple and straightforward as python mo_tf.py  --input_model <frozen.pb> but it's not always that simple. For NCS2 you'd need to add --data_type FP16.

Model Optimizer requires a frozen tensorflow protobuf as input to --input_model. There are a few ways to get a frozen.pb , but a common method is to simply use freeze_graph.py which comes with the Tensorflow distribution.

Be sure to always add --log_level DEBUG so that you can see detailed errors if MO fails !

And as always, feel free to post your findings here.

Thanks for using OpenVino !

Shubha