This library implements the C TensorFlow API adapted for the Scala Native platform.

Scala Native is a unique platform that marries the high level language of Scala but compiles to native code with a lightweight managed runtime which includes a state of the art garbage collector. The combination allows for great programming and the ability to use high performance C language libraries like TensorFlow.

Scala Native uses the Scala compiler to produce NIR (Native Intermediate Representation) that is optimized and then converted to LLVM IR. Finally LLVM code is optimized and compiled by Clang to produce a native executable.

You can use the Gitter8 template stensorflow.g8 to get started. The linked directions will send you back here to install the Tensorflow library below but it should be easier overall.

If you are already familiar with Scala Native you can jump right in by adding the following dependency in your sbt build file. Refer to the TensorflowTest.scala source in this repository or in the template referred to above for an example.

libraryDependencies += "org.ekrich" %%% "stensorflow" % "x.y.z"

To use in sbt, replace x.y.z with the version from Maven Central badge above. All available versions can be seen at the Maven Repository.

Otherwise follow the Getting Started instructions for Scala Native if you are not already setup.

• Use version 0.5.0 for Scala Native 0.5.x.
• Use version 0.3.0 for Scala Native 0.4.9+.

Note: Refer to release notes for older versions of Scala and Scala Native

The TensorFlow C library is required and the current version is 2.16.1 with 2.5.0 pre-built for Linux.

• Linux/Ubuntu can TensorFlow following the following directions:

https://www.tensorflow.org/install/lang_c

The essential directions are repeated here replacing <version> with the above:

$ FILENAME=libtensorflow-cpu-linux-x86_64-<version>.tar.gz
$ wget -q --no-check-certificate https://storage.googleapis.com/tensorflow/libtensorflow/${FILENAME}
$ sudo tar -C /usr/local -xzf ${FILENAME}
$ sudo ldconfig /usr/local/lib

• macOS can install TensorFlow using Homebrew which will install into the /usr/local/Cellar/libtensorflow/<version> directory.

Note: macOS 12 or greater is recommended to install TensorFlow via Homebrew and is used in CI. Tensorflow 2.16.1 is built for macos 13.1 so you will get a linking warning if using an older OS version.

$ brew install libtensorflow

• Other OSes need to have libtensorflow available on the system.

Reference the link above for Scaladoc. The documentation is a little sparse but hopefully will improve with time.

After sbt is installed and any other Scala Native prerequisites are met you can use the following Gitter G8 template instructions to get a fully functional Scala Native application with an example in the body of the main program.

$ sbt new ekrich/stensorflow.g8
$ cd <directory entered after the prompt>
$ sbt run

In addition, look at the stensorflow unit tests for other examples of usage.

TensorFlow Website
TensorFlow for JVM using JNI

The Language Server clangd can be used to help development using VSCode or other editors. For VSCode see the clangd plugin on the Visual Studio Marketplace for more info.

Add a compile_flags.txt with the following content to the stensorflow/src/main/resources/scala-native directory.

# Tensorflow Setup
# Standard path on macOS arm
-I
/opt/homebrew/include

Change the path to match your include path. There is a small amount of official documentation that contains info about compile_flags.txt. Otherwise some other info can be found online.

Release 0.5.0 - (2024-04-11)
Release 0.4.0 - (2024-03-01)
Release 0.3.0 - (2022-11-29)
Release 0.2.0 - (2021-12-13)
Release 0.1.0 - (2021-07-02)