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!

Alert: This site has migrating to our new home at Please visit us there!

Technical FAQ

Q: I get an error "No module named Caffe"
A: Check that you have updated ~/.bashrc with PYTHONPATH set to the Caffe python directory (default is: /opt/movidius/caffe/python) and have either sourced ~/.bashrc or started a new terminal.

Q: I noticed that the Toolkit installs Caffe. Why does it require Caffe?
A: The Toolkit uses Caffe to create graph files and run comparison checks between the NCS and Caffe.

Q: I already have Caffe installed on my system. Can I use that instead?
A: Caffe is installed as part of the the Toolkit installation. Although it's possible to comment out the Caffe installation part of the Toolkit setup script, we recommend you let the Toolkit install its own instance of Caffe. The Toolkit sets PYTHONPATH to the Caffe python installation path (default is /opt/movidius/caffe/python), but if you need to work on your own Caffe installation, update PYTHONPATH back to your old Caffe python installation path.

Q: Which Python version can I use with the Toolkit and the API?
A: The SDK (the API and the Toolkit) require Python 3. Any Python 2.7 code should be ported to Python 3.


  • 4 Comments sorted by Votes Date Added
  • Is it also possible to run OpenCL programs against the compute stick?

  • The Neural Compute Stick (NCS) and accompanying SDK are designed for validation and deployment of Deep Neural Network (DNN) inference applications. The development model is that the developer tunes and compiles a Caffe network with the Toolkit and then builds an application that loads the compiled (graph) file to the NCS using the API. When the application requires an inference it will offload that to the NCS, and the NCS will do all the processing to return the result. OpenCL is not executed on the NCS but nothing would prevent your application from using OpenCL for other tasks.

  • Thanks, I understand the intended workflow. I'd like to evaluate adding support in TensorFlow via OpenCL which fits the intended purpose and would allow use of a wider variety of network architectures. On the Movidius web site ( it looks like the Myriad SDK supports OpenCL, is there a way to get MDK for the NCS?

  • That isn't available at this time.

This discussion has been closed.