How to setup my Ubuntu PC for these Deep Learning frameworks?

I need to use both NVIDIA Digits and tensorflow Object Detection API for different deep learning purposes. I am a bit lost with the compatibility issues, especially for what concerns CUDA. So which version should I choose for DIGITS, Tensorflow-gpu, CUDA and Object Detection API in order to make everything work?

Topic object-detection nvidia gpu tensorflow

Category Data Science

1
answered at

An easy way to run different versions of multiple frameworks alongside each other is to use Docker. With Docker, you can create a new "Container", kind of a light-weight virtual machine, for each project. Docker handles all hardware abstraction and separates the different Frameworks.

Note: To use the GPU within a Docker container, you need to install Nvidia-Docker, as a "normal" Docker installation can not share the GPU with a container.

In your example, you could have a Docker container for Digits, and a separate one for TensorFlow - then the two frameworks will be completely isolated from each other.

This approach comes with multiple advantages:

  • You can run the same container on a different machine without installing e.g. TensorFlow - all you need is (Nvidia-)Docker.
  • You can easily use different (TensorFlow) versions for different projects, as each project gets it's own container which is separated from the system.
  • Containers can use different CUDA versions, e.g. CUDA 8 for Digits and CUDA 9 for TensorFlow.
  • There are pre-built containers for almost all existing deep learning frameworks. Nvidia even provides some GPU-optimized containers with their free NGC (Nvidia GPU Cloud) service.

About

Geeks Mental is a community that publishes articles and tutorials about Web, Android, Data Science, new techniques and Linux security.