Get custom trained yolov3 to run on NCS2

[MILLET__Florian]

Hello everyone,

I succeded in training a YoloV3 model with my own dataset, to detect only one class of object. It properly works and do pretty great on the ./darknet detect images, but now I'd like to make it run on the NCS2, to detect on live camera. I already succeeded in making the official yolov3.weights to run on it, by converting it with https://github.com/mystic123/tensorflow-yolo-v3 's ./convert_weights_pb.py, and the openvino model optimizer, but when I try to use it with my own final.weights I get the error :

        File "C:\Users\VIDOC\Desktop\tensorflow-yolo-v3-master\utils.py", line 115, in load_weights

    (shape[3], shape[2], shape[0], shape[1]))

ValueError: cannot reshape array of size 283154 into shape (256,128,3,3)

So now my question is : is it even possible to convert custom trained yolov3 model ? And if so, do I need to prepare it before training to fit the official yolov3 model strucutre ? Do I need to create the converter by my own ?

The only changes I got from the real yolov3 model is the cfg file, where I put the classes to 1 and the filters to 18 (as I only need one class, if I understood correctly), the .names file where I only wrote my one class on the first line, and the classes = 1 line in the .data file.

Thank you for your answers,

Florian Millet

[JesusE_Intel]

Hi Florian,

Take a look at this forum discussion where I provide my input for tiny YoloV3. It should be similar to the full version of YOLOv3, you will just need to use a different configuration file (json). Could you share the model optimizer command that you used?

Regards,

Jesus

[MILLET__Florian]

Hi,

I can't get to the model optimizer step, I don't manage to convert the .weights file to .pb in the first place (the cannot reshape error), but when I got to the model optimizer with the official yolov3 weights, I used the command provided in the official Intel guide :

python3 mo_tf.py

--input_model /path/to/yolo_v3.pb

--tensorflow_use_custom_operations_config $MO_ROOT/extensions/front/tf/yolo_v3.json

--batch 1

And my custom yolov3.json look like :

     [
  {
    "id": "TFYOLOV3",
    "match_kind": "general",
    "custom_attributes": {
      "classes": 1,
      "coords": 4,
      "num": 9,
      "mask": [3,4,5],
      "jitter":0.3,
      "ignore_thresh":0.7,
      "truth_thresh":1,
      "random":1,
      "anchors":[10,13,16,30,33,23,30,61,62,45,59,119,116,90,156,198,373,326],
      "entry_points": ["detector/yolo-v3/Reshape", "detector/yolo-v3/Reshape_4", "detector/yolo-v3/Reshape_8"]
    }
  }
]

Regards,

Florian

[HemanthKum_G_Intel]

Hi Florian,

I have shared my experiments on Yolov3 here - https://github.com/vijnasu/OpenVINO-Experiments

Feel free to take and modify the Jupyter notebook and let me know if that solves your issue.

[MILLET__Florian]

Hemanth Kumar G. (Intel) wrote:

Hi Florian,

I have shared my experiments on Yolov3 here - https://github.com/vijnasu/OpenVINO-Experiments

Feel free to take and modify the Jupyter notebook and let me know if that solves your issue.

Hi, thank you, but I already manage to perform this on the official yolov3 weight, my problem is when I try it with custom trained model,

regards

[HemanthKum_G_Intel]

Hi Florian,

Could you try adding the input_shape argument to the model optimizer? Something like

--input_shape [1,416,416,3]

as per your input dataset shape used for training?

[MILLET__Florian]

Hemanth Kumar G. (Intel) wrote:

Hi Florian,

Could you try adding the input_shape argument to the model optimizer? Something like

--input_shape [1,416,416,3]

as per your input dataset shape used for training?

Thank you, but as I said earlier, I'm stuck at the convert step, I don't even have a .pb file to give to the model optimizer for now,

Regards

[HemanthKum_G_Intel · Accepted solution]

Hi Florian,

It seems to be not an issue of OpenVINO. Could you raise this issue in YoloV3 forums and get the help from them to generate the frozen graph such that OpenVINO can be used thereafter?

[yu__jia]

Hi Florian,

Would you like to ask if you have solved this problem? I have the same problem as you.

[pang__pih]

Hi,

I am using OpenVINO 2020.1, and I have the same problem.

The convert_weights_pb.py works on coco.names and yolov3.weights,

but not on my 2 class custom trained model. What are the requirement of the yolov3.cfg file of custom trained model, for it to work with convert_weights_pb.py?

Hope somebody has the answer to this issue now!

Thanks

Pih Lung

Error output:

(venv) plp@pihlungs-laptop:~/Documents/code/asl_cust_pt_demo$ python3 convert_weights_pb.py --class_names classes.names --data_format NHWC --weights_file asl_2class_1800.weights

Traceback (most recent call last):

File "convert_weights_pb.py", line 56, in <module>

tf.app.run()

File "/home/plp/.local/lib/python3.6/site-packages/tensorflow/python/platform/app.py", line 40, in run

_run(main=main, argv=argv, flags_parser=_parse_flags_tolerate_undef)

File "/home/plp/.local/lib/python3.6/site-packages/absl/app.py", line 299, in run

_run_main(main, args)

File "/home/plp/.local/lib/python3.6/site-packages/absl/app.py", line 250, in _run_main

sys.exit(main(argv))

File "convert_weights_pb.py", line 46, in main

load_ops = load_weights(tf.global_variables(scope='detector'), FLAGS.weights_file)

File "/home/plp/Documents/code/asl_cust_pt_demo/utils.py", line 115, in load_weights

(shape[3], shape[2], shape[0], shape[1]))

ValueError: cannot reshape array of size 283925 into shape (256,128,3,3)

(venv) plp@pihlungs-laptop:~/Documents/code/asl_cust_pt_demo$

Have tried these "solutions" without any luck!!

https://github.com/mystic123/tensorflow-yolo-v3/issues/15

https://github.com/mystic123/tensorflow-yolo-v3/issues/64