What Requirements are needed to train customer Darknet Yolo v3 & v4 model for Openvino 2020 R3 LTS ?

[CBell1]

We train custom Darknet Yolo v3 & v4 models with customized datasets.

We work in a Ubuntu environment with Tensorflow and we want to use OpenVino 2020 R3 LTS for inference with NCS.

What Requirements are needed to train Darknet Yolo v3 & v4 for Openvino 2020 R3 LTS ?

Is it correct to use an environment with Ubuntu 18.04 LTS and Python 3.6?

What TensorFlow version do you suggest?

What weights to pb conversion tools do you suggest?

Thank you

[Iffa_Intel]

Greetings,

First and foremost, please take note that the OpenVINO version 2020 R3 LTS is an outdated version. However, you may choose to use or not.

You may refer to this official documentation for detailed infos regarding compatibility of YOLO model with Openvino: https://docs.openvinotoolkit.org/2021.1/openvino_docs_MO_DG_prepare_model_convert_model_tf_specific_Convert_YOLO_From_Tensorflow.html

Ubuntu 18.04 is indeed validated to works with OpenVINO and Python 3.6 seems good so far (I'm using these too).

If you go through the documentation, there are details on weight conversion and Tensorflow that would help you to decide.

Sincerely,

Iffa

[CBell1]

Many thanks Iffa for your useful feedback

Last question, what version of TF do you suggest?

Best regards

[Iffa_Intel]

Personally, based on my own development experience, I prefer Tensorflow 1 but it won't hurt to explore the newer TF, which is TF2 .

Sincerely,

Iffa

[CBell1]

TF 1.11 & 1.14 are working well

I just finished to train a customized Yolo v3 model using a customized dataset.

I converted the weights to PB and to IR.

All following you suggestions and the info in the link you have shared

It works

Many Thanks!

[Iffa_Intel]

Glad to hear that!

Intel will no longer monitor this thread since this issue has been resolved. If you need any additional information from Intel, please submit a new question

Sincerely,

Iffa