[Error 5] Toolkit Error: Stage Details Not Supported: BatchNormWithGlobalNormalization

[idata]

I am creating custom Inception model as describe https://tensorflow.org/ . when i am testing this model in NCS then i am getting following error

imran@imran-ThinkPad-L440:~/movidius/ncsdk/examples/tensorflow/custom$ mvNCCompile output/model_ckpt.meta -in=Mul -on=final_result -s12

mvNCCompile v02.00, Copyright @ Movidius Ltd 2016

/usr/local/lib/python3.5/dist-packages/tensorflow/python/util/tf_inspect.py:45: DeprecationWarning: inspect.getargspec() is deprecated, use inspect.signature() instead

if d.decorator_argspec is not None), _inspect.getargspec(target))[Error 5] Toolkit Error: Stage Details Not Supported: BatchNormWithGlobalNormalization

So when i check my tensorflow version in my system , it was 1.3.0 but when i am printing tensorflow version in retrain.py script of the above link i am getting 1.4.0 version of the tensorflow. So please how can i downgrade my bazel version of tensorflow.

[idata]

Can you try

pip3 install tensorflow==1.3this must set Tensorflow to exact version that you cpecify in the command.

[idata]

@georgievm_cms.

It did not work for me .

pip3 install tensorflow==1.3

and then run the command

bazel-bin/tensorflow/examples/image_retraining/retrain --image_dir ~/tensorflow/MachineryTool it still printing 1.4.0 version of the tensorflow. i cant attach image here , otherwise i would have shown you the error.

[idata]

@imran You installed two different versions of the TensorFlow tool. One you built through source, the other you installed through Python Pip. If you use TensorFlow through Python you should have version 1.3 now.

[idata]

Thanks @tsteiner2

This is the issue i am facing because bazel install tensorflow 1.4 while bazel build(I don't know how to pass tensorflow version in bazel build). In my local system i have installed tensorflow version 1.3. I am using bazel to retraining inception v3 model for my custom object detection(Otherwise inception model is working fine )

[idata]

Hi, @tsteiner2

By implementing an inception v3 model, I encounter the same issue.

Both my colleague and I only have one version of python.One installed with pip3(3.5version)One installed in a 3.6 python virtual_env.

We are usint this version of inception delivered by tensorflow.

As you can see in the tar file, there is pnly the .pb an .pbtxt provided that I used instead of the .meta in mvNCompiler.

Maybe the issue comes from there ?

As i am new to this field, do you have any trick to pass from .pb to .met + .ckpt file ?

Regards