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

[idata]

Hi guys

So I am quite new to the NCS.

However, I have a ResNet 18 network which I have trained myself, and I would like to run it on the NCS.I sadly just get to the compile stage when i hit the error: "[Error 5] Toolkit Error: Stage Details Not Supported: Switch". I use this command: "mvNCCompile resnet_18_5.meta -w resnet_18_5 -in input/X -on FullyConnected/BiasAdd -is 480 270 -s 12".

I have found that "Switch" refers to an operation in my Tensorflow graph used by several BatchNormalization layers.

However, since it states in your documentation that you support BN i thought it would work.

I am using Ubuntu on a VM on Win7.

Hope some of you great people in here can help. :smile:

Best Regards

Kristian Jensen

[idata]

@krnj Can you provide your network so that I can reproduce and help debug your issue?

[idata]

Hey @Tome_at_Intel , i figured out one more thing that is relevant. My code is using the tf.cond operation which is where the switch occur and I am not sure if you support?

[idata]

@krnj I am not seeing the tf.cond in TensorFlowParser.py. I wonder if you can replace the tf.switch or tf.cond with python conditional statements. If you can provide a link to your network, it would be helpful to us. Thanks.

[idata]

Hi @Tome_at_Intel, I've been facing the same error. Replacing tf cond with python conditional statements wouldn't work because tf.cond is evaluated lazily on the Tensorflow computational graph at runtime whereas python conditionals are evaluated at graph construction time. Please address this issue as soon as possible, as this layer seems to be an integral part of custom networks.

[idata]

Hi @Tome_at_Intel, I am also facing the same error. I am following the code [https://github.com/bsrinivas8687/WBC_Segmentaion] and trying to implement it on my NC. I have installed ncsdk-2.05.00.02 on my Ubuntu 16.04. The said github code was written using Keras. Therefore, I have converted the .hdf5 file to .pb file.

The model architecture, weight file and pb file are shared here: https://www.dropbox.com/sh/8eeb2mtxhie4sg6/AAC1CSUJontj2XCVTfbxMSiYa?dl=0

Kindly help me.

Thanking you.

With Regards

Maitreya M

[idata]

Hi @Tome_at_Intel I am facing the same issue trying to use a retrained tiny-yolo-v2 model from darkflow.

I just used my personal voc data and followed the tutorial.Here is the full output: magaly@elter-magaly ~ mvNCCompile darkflow/ckpt/tiny-yolo_elter-250.meta -s 12 -in input -on output -o tiny-v2_elter.graph /usr/local/lib/python3.6/dist-packages/skimage/__init__.py:60: DeprecationWarning: the imp module is deprecated in favour of importlib; see the module's documentation for alternative uses import imp /usr/local/bin/ncsdk/Controllers/Parsers/TensorFlowParser/Convolution.py:44: SyntaxWarning: assertion is always true, perhaps remove parentheses? assert(False, "Layer type not supported by Convolution: " + obj.type) /usr/local/bin/ncsdk/Controllers/Parsers/Phases.py:501: DeprecationWarning: invalid escape sequence \ """ /usr/local/bin/ncsdk/Controllers/Parsers/Phases.py:747: DeprecationWarning: invalid escape sequence \ """ mvNCCompile v02.00, Copyright @ Intel Corporation 2017 /usr/local/bin/ncsdk/Controllers/TensorFlowParser.py:173: DeprecationWarning: invalid escape sequence \d /usr/local/lib/python3.6/dist-packages/tensorflow/python/util/tf_inspect.py:45: DeprecationWarning: inspect.getargspec() is deprecated, use inspect.signature() or inspect.getfullargspec() shape: [1, 416, 416, 3] [Error 5] Toolkit Error: Stage Details Not Supported: Switch

Regards