An error when i try to convert YOLO to IR

[Dimant__Shalom]

Hi, i have 2 problems:

1. when i try to convert from yolov3 weights to IR i need to specify the input to [1,412,412,3]. but the original YOLOv3 is 612x612 (with better result then 412x412). can i change it?

2. when i run the command:

./object_detection_demo -i ~/splited.mp4 -m ~/intel/openvino_2019.1.144/deployment_tools/model_optimizer/frozen_darknet_yolov3_model.xml

i get the following error msg:

[ ERROR ] Error reading network: in Layer detector/darknet-53/Conv_1/Conv2D: trying to connect an edge to non existing output port: 2.1

i look in the frozen_darknet_yolov3_model.xml file and i think it correpted because at the end of the file there is a list of connection between the layers. and there is a line in there: 

<edge from-layer="2" from-port="1" to-layer="3" to-port="0"/>

but THERE IS NO LAYER NAMED "2"!

also 

<edge from-layer="4" from-port="1" to-layer="5" to-port="0"/>

but there is no layer named "4"!

and so on.

can somone help me plz?

thx

Shalom

The XML file is attached

[Shubha_R_Intel]

Dear Dimant, Shalom,

Indeed I confirmed your observations about missing layers when I looked at the contents of your zip file. May I know exactly which command you used to build the Yolo V3 IR ? I can tell you that I followed Converting YOLO* models in 2019 R1 and I definitely didn't see the problems you are seeing. 

OpenVino R1.1 was just released. Can you kindly try it ? Please post your results here.

Thanks,

Shubha

[Dimant__Shalom]

Hi Shubha thanks for your response,

I alredy use OpenVino R1.1 

I follow the orders in Converting YOLO* models.

Just in case, i did it again.

When i run the follow command: 

python3 /home/ws/intel/openvino_2019.1.144/deployment_tools/model_optimizer/mo_tf.py --input_model frozen_darknet_yolov3_model.pb --tensorflow_use_custom_operations_config ~/intel/openvino_2019.1.144/deployment_tools/model_optimizer/extensions/front/tf/yolo_v3.json

I get the first error i wrote:

Model Optimizer version:        2019.1.1-83-g28dfbfd
[ ERROR ]  Shape [ -1 416 416   3] is not fully defined for output 0 of "inputs". Use --input_shape with positive integers to override model input shapes.
[ ERROR ]  Cannot infer shapes or values for node "inputs".
[ ERROR ]  Not all output shapes were inferred or fully defined for node "inputs".
 For more information please refer to Model Optimizer FAQ (<INSTALL_DIR>/deployment_tools/documentation/docs/MO_FAQ.html), question #40.
[ ERROR ]
[ ERROR ]  It can happen due to bug in custom shape infer function <function tf_placeholder_ext.<locals>.<lambda> at 0x7f11ffba8620>.
[ ERROR ]  Or because the node inputs have incorrect values/shapes.
[ ERROR ]  Or because input shapes are incorrect (embedded to the model or passed via --input_shape).
[ ERROR ]  Run Model Optimizer with --log_level=DEBUG for more information.
[ ERROR ]  Exception occurred during running replacer "REPLACEMENT_ID" (<class 'extensions.middle.PartialInfer.PartialInfer'>): Stopped shape/value propagation at "inputs" node.
 For more information please refer to Model Optimizer FAQ (<INSTALL_DIR>/deployment_tools/documentation/docs/MO_FAQ.html), question #38.

 so i change the command to:

python3 mo_tf.py  --input_shape [1,416,416,3] --input_model ~/convertYoloV3ToIR/tensorflow-yolo-v3/frozen_darknet_yolov3_model
.pb --tensorflow_use_custom_operations_config "/home/ws/intel/openvino_2019.1.144/deployment_tools/model_optimizer/extensions/front/tf/yolo_v3.json"

and then i run:

./object_detection_demo -i ~/splited.mp4 -m ~/intel/openvino_2019.1.144/deployment_tools/model_optimizer/frozen_darknet_yolov3_model.xml

and get the msg:

[ INFO ] InferenceEngine:
        API version ............ 1.6
        Build .................. custom_releases/2019/R1.1_28dfbfdd28954c4dfd2f94403dd8dfc1f411038b
Parsing input parameters
[ INFO ] Files were added: 1
[ INFO ]     /home/ws/splited.mp4
[ INFO ] Loading plugin

        API version ............ 1.6
        Build .................. 23780
        Description ....... MKLDNNPlugin
[ INFO ] Loading network files:
        /home/ws/intel/openvino_2019.1.144/deployment_tools/model_optimizer/frozen_darknet_yolov3_model.xml
        /home/ws/intel/openvino_2019.1.144/deployment_tools/model_optimizer/frozen_darknet_yolov3_model.bin
[ ERROR ] Error reading network: in Layer detector/darknet-53/Conv_1/Conv2D: trying to connect an edge to non existing output port: 2.1

the JSON file:

[
  {
    "id": "TFYOLOV3",
    "match_kind": "general",
    "custom_attributes": {
      "classes": 80,
      "coords": 4,
      "num": 9,
      "mask": [0, 1, 2],
      "entry_points": ["detector/yolo-v3/Reshape", "detector/yolo-v3/Reshape_4", "detector/yolo-v3/Reshape_8"]
    }
  }
]

Thanks!

[Shubha_R_Intel]

Dear Dimant, Shalom,

It looks like you did everything correctly.

A github forum poster recently discovered github issue 151 during inference on tiny yolo v3. And I reproduced it on 2019R1.1. I regret to say that there may be a yolo v3 bug as well. Your error looks distinctly different but who knows - maybe even the non-tiny yolo v3 is broken. I shall reproduce it and report back here. I'm really sorry about the inconvenience.

Thanks,

Shubha

[Dimant__Shalom]

Thanks Shubha,

I'd love you to let me know when you know

[Shubha_R_Intel]

Dear Dimant, Shalom,

Yolo V3 works fine on 2019R1.1. Tiny Yolo is broken, however, per that github issue I posted earlier. You should not be running object_detection_demo  for yolo. Instead, try the python yolov3 demo:

C:\Users\sdramani\Downloads\tensorflow-yolo-v3>python "c:\Program Files (x86)\IntelSWTools\openvino_2019.1.148\deployment_tools\inference_engine\samples\python_samples\object_detection_demo_yolov3_async\object_detection_demo_yolov3_async.py" -i c:\users\sdramani\Downloads\sample-videos\person-bicycle-car-detection.mp4 -m frozen_darknet_yolov3_model.xml -l c:\users\sdramani\Documents\Intel\OpenVINO\inference_engine_samples_build\intel64\Release\cpu_extension.dll

It works !

Hope it helps and sorry that it took so long,

Shubha

[Dimant__Shalom]

Hi Shubha, thank for your answer.

Unfortunately this answer didn't solve the problem :-(

When i execute:

python3 object_detection_demo_yolov3_async.py -i ~/20190522_110655_ra71.mp4 -m ~/convertYoloV3ToIR/tensorflow-yolo-v3/frozen_darknet_yolov3_model.xml

the output is: 

[ INFO ] Loading network files:
        /home/ws/convertYoloV3ToIR/tensorflow-yolo-v3/frozen_darknet_yolov3_model.xml
        /home/ws/convertYoloV3ToIR/tensorflow-yolo-v3/frozen_darknet_yolov3_model.bin
Traceback (most recent call last):
  File "object_detection_demo_yolov3_async.py", line 349, in <module>
    sys.exit(main() or 0)
  File "object_detection_demo_yolov3_async.py", line 175, in main
    net = IENetwork(model=model_xml, weights=model_bin)
  File "ie_api.pyx", line 271, in openvino.inference_engine.ie_api.IENetwork.__cinit__
RuntimeError: Error reading network: in Layer detector/darknet-53/Conv_1/Conv2D: trying to connect an edge to non existing output port: 2.1

Notice that the XML file is corrupted, therfor i don't understend why it should work anyway.

Can you please send me you'r XML file?

[Shubha_R_Intel]

Dear Dimant, Shalom,

Make sure you are not using Tensorflow 1.13. Model Optimizer only supports up to 1.12 and in fact YoloV3 requires a fairly recent version too - 1.12 should work fine. But 1.13 will break Model Optimizer.

Attached is a *.zip file containing the generated IR XML.

Hope it helps,

Thanks,

Shubha

[Evgenya_S_Intel]

Dear Dimant Shalom,

I would like to answer your initial questions:

1. Yes, you do can use YOLO v3 model with bigger input shape (recommended size is 608x608). To do so, please add --size key in convert_weights_pb.py script like so:

python3 convert_weights_pb.py --class_names coco.names --data_format NHWC --weights_file yolov3_608.weights --size 608

2. Edges are connected according to layer ids in .xml, not by names. 

I didn't manage to reproduce the issue you reported, but I would like to try harder.

For futher investigation of your issue, please provide as much as you can of:

- IR (.xml and/or .bin)

- TensorFlow .pb file

- DarkNet files used to generate .pb (.cfg and .weights)

- TensorFlow version you use for DarkNet -> TensorFlow and TensorFlow -> IR conversions.

I'm really sorry for your inconvenience. 

Thanks,

Evgenya

[gao__yuangang]

Excuse me. Now,openvino2019.1.148 is installed in win10. Is the tensorflow version 1.14.0 needed to be installed? 

[gao__yuangang]

excuse me. now openvino2019.1.148 support tensorflow1.14.0 ? 

I successfully convert yolov3 model to tensorflow(pb model) , to openvino (IR model) in tensorflow1.14.0 on windows10. but failed on others tensorflow version.