Community
cancel
Showing results for 
Search instead for 
Did you mean: 
Ling_L_Intel
Employee
53 Views

IR Model from tensorflow

The current version support MASKRCNN running on CPU.    I am trying to run a mobile-net backbone based mask RCNN.   the original source is @https://github.com/GustavZ/Mobile_Mask_RCNN , with Keras and tensor flow model,  I use cmd 

      python3 mo_tf.py --input_model model.pb  , then it reports "Graph contains a cycle" 

     I change the command to

     python3 mo_tf.py --input_model model.pb --tensorflow_use_custom_operations_config extensions/front/tf/mask_rcnn_support.json   

     it reports "Generated IR model"  , yet the .bin is empty .

Any suggestion on it?    

    If you need the pb model. I could send it out.

thanks,

Ling

0 Kudos
4 Replies
Severine_H_Intel
Employee
53 Views

Dear Liu, 

you are missing one argument in your command line --tensorflow_object_detection_api_pipeline_config, please look at our documentation on how we support Mask RCNN: computer_vision_sdk_2018.3.343/deployment_tools/documentation/docs/TensorFlowObjectDetectionAPIModels.html for more details

Also, if the model is not from the model zoo, you might want to try with the mask_rcnn_support_api_v1.7.json  file. 

Best, 

Severine

Ling_L_Intel
Employee
53 Views

Dear Habert:

    thanks​ for your quick response. I add the argument --tensorflow_object_detection_api_pipeline_config , and the model could be generated, yet the bin file is empty.  I use the tool to open my pb model and the official pb model from model zoo,  they two looks quite different.  It seems the official pb model has more detail layers and lots of customer defined layers.  Could you tell me where is the source of the official pb model? or how to generate the pb model like yours?

thanks,

Ling   

 

Ling_L_Intel
Employee
53 Views

attached the two model picture as below

Ling_L_Intel
Employee
53 Views

Dear Habert:

     I read the document computer_vision_sdk_2018.3.343/deployment_tools/documentation/docs/TensorFlowObjectDetectionAPIModels as you suggested.  If I understand correctly,  now we only support Tensorflow opject detection API models from the object detection model zoo.  Currently the MASK-RCNN is reset101/50/inception backboned.  Our goal is to implement a mobilenet backbone MASK RCNN, my question is:

   If I use tensorflow object detection API to build the mobilenet backbone  mask RCNN, does model optimizer could transfer it correctly and we could run it on OPENVINO?

    Since the ssd-mobilenet-* network could be supported now, I guess Mask_RCNN_mobilenet* should be worked, right?  DO we have any plan to support it?   

thanks,

Ling