Converting Public Model yolo-v4-tf to IR

[at93]

Hi Everyone,

I've just downloaded the Intel public model yolo-v4-tf. Can someone explain to me how to convert the darknet weights/cfg file to IR.  I've seen the scripts included in the download but not familiar to keras.

Thanks.

[at93]

I am able to convert the darknet weights/cfg to keras file using the convert.py script. However, when I tried to convert from keras to tensorflow using the keras_to_tensorflow.py, I am getting the following errors:

C:\Users\MeAtPlay\Documents\Intel\public\yolo-v4-tf\keras-YOLOv3-model-set\tools\model_converter>python keras_to_tensorflow.py --input_model="C:\Users\MeAtPlay\Documents\Intel\public\yolo-v4-tf\yolov4.h5" --output_model="C:\Users\MeAtPlay\Documents\Intel\public\yolo-v4-tf\yolov4.pb"
WARNING:tensorflow:From C:\Users\MeAtPlay\AppData\Roaming\Python\Python36\site-packages\tensorflow_core\python\compat\v2_compat.py:68: disable_resource_variables (from tensorflow.python.ops.variable_scope) is deprecated and will be removed in a future version.
Instructions for updating:
non-resource variables are not supported in the long term
Traceback (most recent call last):
File "keras_to_tensorflow.py", line 188, in <module>
app.run(main)
File "C:\Users\MeAtPlay\AppData\Roaming\Python\Python36\site-packages\absl\app.py", line 303, in run
_run_main(main, args)
File "C:\Users\MeAtPlay\AppData\Roaming\Python\Python36\site-packages\absl\app.py", line 251, in _run_main
sys.exit(main(argv))
File "keras_to_tensorflow.py", line 134, in main
model = load_input_model(FLAGS.input_model, FLAGS.input_model_json, FLAGS.input_model_yaml, custom_objects=custom_object_dict)
File "keras_to_tensorflow.py", line 67, in load_input_model
model = load_model(input_model_path, custom_objects=custom_objects)
File "C:\Users\MeAtPlay\AppData\Roaming\Python\Python36\site-packages\tensorflow_core\python\keras\saving\save.py", line 143, in load_model
return hdf5_format.load_model_from_hdf5(filepath, custom_objects, compile)
File "C:\Users\MeAtPlay\AppData\Roaming\Python\Python36\site-packages\tensorflow_core\python\keras\saving\hdf5_format.py", line 160, in load_model_from_hdf5
model_config = json.loads(model_config.decode('utf-8'))
AttributeError: 'str' object has no attribute 'decode'

[at93]

I am able to fix the error "AttributeError: 'str' object has no attribute 'decode'" by downgrading h5py package to 2.10.0.

So now that I have the frozen TF, how do I use the model optimzer to convert it to IR. The yolov4.json file is not included in the models zoo? Any help is appreciated.

[Adli]

Hi at93,

Thank you for reaching out to us. Please try and run the following command:

python "<openvino_dir>\deployment_tools\model_optimizer\mo.py" --framework=tf --data_type=FP32 --output_dir=<output_dir> --model_name=yolo-v4-tf --input_shape=[1,608,608,3] --input=image_input --scale_values=image_input[255] --reverse_input_channels --input_model="<frozen_file_directory>\<frozen_file_name>.pb"

Besides, you can run the following command to convert public model 'yolo-v4-tf' into the IR model:

python converter.py --name yolo-v4-tf

Regards,

Adli

[at93]

Thanks Adli,

I was able to convert using

python converter.py --name yolo-v4-tf

 However, when I run the object_detection_demo it does not seems to like the model file. The following is the output of the program.

C:\Users\MeAtPlay\Documents\Intel\OpenVINO\omz_demos_build\intel64\Release>object_detection_demo -at "yolo" -m "C:\DVCell\Common64\models\yolo-v4-tf.xml" -i "C:\DVCell\Common64\Media\WaikikiBeach.mp4" -d "GPU"
[ INFO ] InferenceEngine: API version ......... 2.1
Build ........... 2021.2.0-1877-176bdf51370-releases/2021/2
[ INFO ] Parsing input parameters
[ INFO ] Reading input
MFX: Can't initialize session
[ INFO ] Loading Inference Engine
[ INFO ] Device info:
[ INFO ] GPU
clDNNPlugin version ......... 2.1
Build ........... 2021.2.0-1877-176bdf51370-releases/2021/2
Loading network files
[ INFO ] Batch size is forced to 1.
[ INFO ] Checking that the inputs are as the demo expects
[ INFO ] Checking that the outputs are as the demo expects
[ ERROR ] [ ERROR ] Invalid output type: Add. RegionYolo expected

[at93]

In addition, I tried the object_detection_demo again with yolov3 model and it worked.

[Adli]

Hi at93,

If possible, could you run the 'object_detection_demo.py' with the 'yolo-v4-tf' IR model? The demo is located in 'python_demos\object_detection_demo' directory. 

Regards,

Adli

[at93]

Hi Adli,

The tried the 'object_detection_demo.py' with the 'yolo-v4-tf' IR model. It does run without errors BUT I don't think it inferred properly. The bounding boxes are everywhere as you can see in the image here. 

[at93]

I also tried the same python script with the Yolo3 model and it is OK. Please see image below.

 

[Adli]

Hi at93,

Thank you for your prompt response. We are investigating this issue and will get back to you soon.

Regards,

Adli

[Adli]

Hi at93,

The object detection demo (python) has not been updated to support 'yolo-v4-tf' yet. The supported models for this demo can be found on the 'models.lst' file:

https://github.com/openvinotoolkit/open_model_zoo/blob/master/demos/python_demos/object_detection_demo/models.lst

I apologize for the inconvenience.

Regards,

Adli

[at93]

Thank Adli,

Are there other sample code or documentation on how to use the Openvino2021.2 library to infer the yolov4 model or I have to wait till 2021.3 version?

Thanks.

[Vladimir_Dudnik]

You can run accuracy validation for yolo-v4-tf model using Open Model Zoo Accuracy Checker tool (you'll also need MSCOCO validation dataset for that.

[Adli]

Hi at93,

This thread will no longer be monitored since we have provided a solution. If you need any additional information from Intel, please submit a new question.

Regards,

Adli