This project will allow you to build from scratch, package into a jar file, and install into the local maven repository, a TensorFlow distribution's Java extensions built for the GPU. It will also package the JNI native libraries, along with the entire native TensorFlow libraries, into a jar file that can be retrieved using ai.kognition.pilecv4j.utils.NativeLibrary.
* Note: This currently ONLY builds a 64-bit linux (Ubuntu) install using Docker. *
To run the build, you will need to have an NVidia GPU with the drivers installed.
This install script executes the build using a Docker container that's already set up with the appropriate CUDA NVidia libraries, so you don't need to install all of that on your machine to get it to build. However, you wont be able to run an program you write against this TensorFlow build without having the appropriate version of CUDA installed on your environment. The easiest way to do this will be to run whatever you build in a Docker container that's alread set up with these things.
All of this means basically there's two more prerequisites:
- You'll need Docker installed
- You'll need the NVidia Docker runtime,
nvidia-dockerinstalled
By default the container used for the build has Ubuntu 18.04, CUDA 9.2, and CUDNN 7. If you want something different you can modify these scripts and supply a different base container from NVidia's Docker images in the docker.io registry. I can't guarantee that the build will work with a different setup and so just changing the script may not be the only change necessary. As it is, it works on my machine.