frame

Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!

Sign In

Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!

Tensorflow Support on NCS.

As we got to know that now we can use tensorflow created model in NCS. As in there document they have specify that for custom created model we will need graph, stat.txt and categories.txt file in which categories.txt and stat.txt file is ok. But my question is that as we know that while training on tensorflow we get model.ckpt.meta, model.ckpt.index , model.ckpt.data and greaph.pb file. so when i am giving this graph.pb file in NCS api . its giving me error. So from how we can create graph file without weight (not graph.pb)? what is the difference between this graph.pb file and the graph file used in NCS.

Comments

  • 15 Comments sorted by Votes Date Added
  • @Tome_at_Intel,
    As you mention above link of tensorflow for NCS. for compilation code given is

    mvNCCompile output/inception-v1.meta -in=input -on=InceptionV1/Logits/Predictions/Reshape_1 -s12

    So i wanted to know what is -in=input and -on here. for ssd_mobilenet_v1_coco_11_06_2017 what are the change i need to do. or temporary you are supporting inception only

  • hi @Tome_at_Intel,

    i am still waiting for your reply, as you told me to forward my model which i have already given you or you can download from here
    https://1drv.ms/u/s!AufXe8jWDlK5h0e7_yyGU3dOdZaP
    Please let me know the proceedure, This model is trained on ssd_mobilnet

  • Hello @Tome_at_Intel , I have similar issue as @imran and please let me know when you guys have a solution.
    Thanks so much.

  • @imran The protobufs (pb) file and the graph file used by the NCS are similar is nature, because they both contain information from the weights file and the prototxt file but the formats are different. You can test freezing the pb model along with the meta and ckpt files. For more information on Tensorflow support on the NCS, please visit: https://movidius.github.io/ncsdk/TensorFlow.html

  • Got my stick and followed the instructions. First I installed a fesh copy of Ubuntu 16.04 on a raspberry pi. Then I followed the instructions here:

    https://developer.movidius.com/start

    The last step in Step 2 is make install. I got the following error:

    ...
    Movidius Neural Compute Toolkit Setup.
    Checking OS and version...
    Installing on Ubuntu 16.04

    Installer NCSDK version: 1.09.00.04
    ...
    Checking whether tensorflow is installed...
    looking for tensorflow CPU version...
    You are using pip version 8.1.1, however version 9.0.1 is available.
    You should consider upgrading via the 'pip install --upgrade pip' command.
    looking for tensorflow GPU version...
    You are using pip version 8.1.1, however version 9.0.1 is available.
    You should consider upgrading via the 'pip install --upgrade pip' command.
    Couldn't find a supported tensorflow version, installing tensorflow v1.3.0
    Collecting tensorflow==1.3.0
    Could not find a version that satisfies the requirement tensorflow==1.3.0 (from versions: )
    No matching distribution found for tensorflow==1.3.0
    You are using pip version 8.1.1, however version 9.0.1 is available.
    You should consider upgrading via the 'pip install --upgrade pip' command.
    Installation failed. Error on line 303

  • @Tome_at_Intel where i can send you that. becouse it is very big file.

  • edited October 2017 Vote Up0Vote Down

    @Tome_at_Intel,
    this is the final .pb file directory i got.
    or you can get it from onedrive here
    https://1drv.ms/u/s!AufXe8jWDlK5h0e7_yyGU3dOdZaP

  • Does NCS support the training of Tensorflow models?

  • @dtyu The NCS is designed to be used in conjunction with edge devices to offload inference-related processing from a neural network. Training must be done on a separate host machine.

  • Hi everyone. I would also appreciate any update on how to get a custom TF network to work with Movidius, as I'm experiencing the same problem while compiling the network using mvNCCompile.

  • similar issue here. @Tome_at_Intel please let us know if its not supported then we can start working on something else.

  • @smit.idevices Early versions of NCSDK were supporting only Caffe models and since TF is more complex framework it is time consuming to develop full support for it. Hopefully in next NCSDK versions more layers and TF features will be supported until fully done. Currently there are TF models with less complexity which are working properly on NCS.

  • @georgievm_cms Thank you for the reply. I can understand. I am really looking forward for upcoming SDKs and full tensorflow support.

  • Hi guys. I am wondering whether a custom tensorflow model is supported by Movidius. I first got error 34 but managed to solve it. Now I got error 5 toolkit error stage details not supported realdiv. I believe this is not something I could solve. I guess that RealDiv is an operation supported by tensorflow but not by Movidius yet. All in all, I really want to hear some progress on this topic.

  • @Yang Current release 1.11 of NCSDK has RealDiv in the support list. If you are on older release please update and try again with your model.

Sign In or Register to comment.