Cannot convert Tensorflow models to IR

[rudakov__mikhail]

I'm trying to convert and use the Object detection models such as faster_rcnn_inception.. etc. But, with the exception for  a single SSD model, I get various error messages from Model Optimizer.

The model is faster_rcnn_inception_v2_coco_2018_01_28 which I've downloaded from the Model Zoo.

python ./mo_tf.py --input_model C:\downloads\faster_rcnn_inception_v2_coco_2018_01_28.tar\faster_rcnn_inception_v2_coco_2018_01_28\frozen_inference_graph.pb   --tensorflow_use_custom_operations_config extensions/front/tf/faster_rcnn_support.json   --output="num_detections,detection_boxes,detection_classes,detection_scores" --tensorflow_object_detection_api_pipeline_config C:\downloads\faster_rcnn_nas_coco_2018_01_28.tar\faster_rcnn_nas_coco_2018_01_28\pipeline.config 

Output:

[ ERROR ]  Cannot infer shapes or values for node "add".

[ ERROR ]  <lambda>() missing 1 required positional argument: 'b'

[ ERROR ]  Stopped shape/value propagation at "add" node.

However, I've successfully converted and run a SSD model. 

I'm completely uncommon with the error messages the MO is producing. Could you please provide a valid command for converting any of the object detection models into IR?

[Cary_P_Intel1]

Hi,

I've tested the same pre-trained model as you attached on R4, there is no issue on both MO/IE. With R3, I saw you need to cut off the output, but it has been supported by R4 with automatic sub-graph replacement, please refer to the documents attached.