Raspian-Stretch Conversion of pretrained Tensorflow files to intermediate xml/bin for Yolov3

[Lawton__John]

Hi

I have been using reference A and reference B to prepare to convert tensorflow pretrained files to xml/bin intermediate format.

1) I have installed Tensorflow generally as per:

sudo apt install libatlas-base-dev

pip3 install tensorflow

2) Executed the python3 test program:

# Python

import tensorflow as tf

tf.enable_eager_execution()

hello = tf.constant(‘Hello, TensorFlow!’) print(hello)

3) Execution returns:

Runtimewarning:compiletime version 3.4 of module 'tensorflow.python.framework.fast_tensor_u does not match runtime version 3.5 Runtimewarning: builtins.type size changed, may indicate binary incompatibility. Expected 432; got 412

Note Raspian has been updated and upgraded.

Any advice ion how to remedy this is  appreciated so that I can proceed to execute the conversion python3 script.

Many Thanks 

                                 john

Reference A

https://software.intel.com/en-us/articles/OpenVINO-Using-TensorFlow

Reference B

https://medium.com/tensorflow/tensorflow-1-9-officially-supports-the-raspberry-pi-b91669b0aa0

[Shubha_R_Intel]

Hi John. Neither of the warning errors you've posted have anything to do with OpenVino. They are Tensorflow issues. I found some links which may help you:

https://stackoverflow.com/questions/47225210/tensorflow-install-fails-with-compiletime-version-3-5-of-module-does-not-match

https://stackoverflow.com/questions/48135310/tensorflow-on-raspberry-pi-error

Thanks,

Shubha

[Lawton__John]

Thank you for your comment.

I have now revisited this tensorflow conversion to OpenVino issue and whilst I agree it is not an OpenVino issue

it makes it very difficult to create the xml and bin files because of the tensorflow python 3.4/ 3.5 issue.

I have tried installing python and tensorflow under python virtual environments but have not yet succeeded in 

obtaining compatible versions. I do not see a  tensorflow for version for python 3.5 on the tensorflow site.

Is it possible to obtain the tensor flow format architecture and weight files directly for submission to the OpenVino conversion process to xml and bin files?

Would appreciate guidance!

        Thanks

                                    john

[Shubha_R_Intel]

Dearest John, 

Please forgive me but I do not understand this :

Is it possible to obtain the tensor flow format architecture and weight files directly for submission to the OpenVino conversion process to xml and bin files?

Can you perhaps ask it in a different way ?

Thanks,

shubha

[Lawton__John]

Hi shubha

1) Very sorry I was not clear. What I was asking was "can I obtain the xml and bin files directly to use in the OpenVino YoloV3 sample example program"

2) The following points are pertinent to the challenges I have encountered.

a) The model optimiser support directories do not appear to be supplied in the Raspian installation.

b) With reference to:

 https://software.intel.com/en-us/articles/OpenVINO-Using-TensorFlow#yolov3-overview

"Convert YOLOv3 Model to IR"

I have used the git content at  https://github.com/mystic123/tensorflow-yolo-v3 in order to access the tensorflow implementation

and attempt to perform the conversion to form the xml and bin files.

c) Incompatability between the python tensorflow model and the raspian distro version of python has led me to try using python virtual environments to find compatible  python/tensorflow versions. Unfortunately I have not succeeded thus far. Ideally I would like to use tensorflow compatible with python 3.5 ( as per raspian distro).

3) I would of course like the flexibility and intrinsic development of understanding in generating the xml and bin files from the reference sources.

Appreciate your comments.

                                         John