Running with Yolov3 models

[amkurup]

Hi, I am trying to run this package with yolo-v3 as given in this tutorial from intel. I have faced no problem running on the CPU with the ssd300 model downloaded from intel.

But when I try the the sample on the coco dataset, the program quits with this error: [ERROR] object detection network should have only one output. As I understand, the ssd300 model has 21 outputs. If so, what could be causing this issue?

The files I have tried are exactly as mentioned in the link above: yolov3.weights file with coco.names. I first used tensorflow-yolov3 to convert to frozen model using: python3 convert_weights_pb.py --class_names coco.names --data_format NHWC --weights_file yolov3.weights and then converted to IR using: python3 mo_tf.py --input_model /path/to/yolo_v3.pb --tensorflow_use_custom_operations_config $MO_ROOT/extensions/front/tf/yolo_v3.json

My system: Ubuntu16.04 & Ros Kinetic ros_openvino_toolkit commit: fcf0f9d dldt checkout: 2018_R5 open_model_zoo: 2018_R5

[amkurup]

Not sure if this is relevant but I tried adding --tensorflow_object_detection_api_pipeline_config pipeline.config as mentioned in issue (https://github.com/intel/ros_openvino_toolkit/issues/6#issuecomment-457992911) but I do not have a pipeline.config file.

[chaoli2]

@amkurup This exception is thrown here. Seems that the output size here is not the network output size, but the pipeline output. See the flow chart for more details. I'd like to try to reproduce your issue from my side and will let you know if there is any update.

[amkurup]

Hi @chaoli2 I was able to trace the exception to its source, but as you know, the basic yolo-v3 model has 3 output yolo layers. Seems like they would always throw such an exception! Yet, there are users who claim this toolkit to be working with yolo-v3 generated models. Could you please elaborate the workaround?

I am guessing I am using mo_tf.py in the wrong way. My usage is: python3 mo_tf.py --input_model /path/to/yolo_v3.pb --tensorflow_use_custom_operations_config $MO_ROOT/extensions/front/tf/yolo_v3.json

[chaoli2]

@amkurup The command we use to convert is:

python3 mo_tf.py --input_model /path/to/frozen_darknet_yolov3_model.pb --tensorflow_use_custom_operations_config 
$MO_ROOT/extensions/front/tf/yolo_v3.json --input_shape=[1,416,416,3]

After that, we've tried the generated model with OpenVINO native sample code _(/opt/intel/computer_vision_sdk/inference_engine/samples/object_detection_demo_yolov3async) and it works well, but not tested in this ROS package yet.

[amkurup]

Thanks @chaoli2 I will try to use the above commands and the sample application as a test. I am eager to find out your results with the ROS package.

[amitrege]

@amkurup @chaoli2 Has this been resolved yet?

I am using this command python3 mo_tf.py --input_model /path/to/frozen_darknet_yolov3_model.pb --tensorflow_use_custom_operations_config $MO_ROOT/extensions/front/tf/yolo_v3.json --input_shape=[1,416,416,3] to generate IR.

Then I make a change to the pipeline_object_topic.yaml file and change the model parameter to point to the generated .xml file.

I get the same error: Object Detectionnetwork should have only one input

[amkurup]

@amitrege As far as my knowledge goes, seems like @chaoli2 and team will have to look into this. The fundamental issue is the use of Yolov3 (since it has 3 output layers). For my project, I was able to use OpenVINO's demo code for Yolov3 and bind it with this package (for ROS support)

[amitrege]

Thanks @amkurup . Is your code open source by any chance?