Using GPU with TensorFlow

[imagej]

Using GPU with TensorFlow

[ctrueden]

Maven is a build tool. It does not fully solve deployment. You do not need libtensorflow_jni_gpu to build. What is needed to make GPU work at runtime is a deployment that includes the correct native artifacts. The reason to declare things in a Maven POM at runtime scope is to document that they are necessary and/or desirable for the project to function optimally at runtime. Maven has some limited support for synthesizing a simple runtime environment based on the project's dependencies, so that e.g. you can run from an IDE or via (in the case of SciJava projects) mvn -Pexec on the command line. But these are not the intended primary deployment vehicles. In the case of GPU support for TensorFlow within ImageJ, we just need to ship the correct natives in the correct jars/<platform> and/or lib/<platform> directories. The only real "problem" to solve, IMHO, is how to make the imagej-maven-plugin's copy-jars goal smart enough to populate an ImageJ installation in a cross-platform way. But this problem is not specific to imagej-tensorflow.

[hadim]

Oh, I see so it seems to be simpler to solve than what I was thinking.

Since it's not specific to this repo, I close here.