Feature/TensorFlow GPU

[Spiess]

Implements additional build flag -PGPU=true to use GPU version of TensorFlow.

Closes #293.

Will be made concrete once #294 is merged.

[silvanheller]

Have you tested that this is actually just a drop-in replacement and uses the GPU (e.g. using nvidia-smi)

[Spiess]

Please excuse, I forgot to mention that I did test this and could determine that the GPU is used, but since different item extractions are not batched, this could still be improved in the future.

[silvanheller]

I tested this on ubuntu and can confirm it works. There is however not really any noticeable speedup for CLIP. One request: Could the jar have a flag indicating whether it was built with gpu support?

[lucaro]

I tested this on ubuntu and can confirm it works. There is however not really any noticeable speedup for CLIP. One request: Could the jar have a flag indicating whether it was built with gpu support?

I think that would also make sense for the application build. Alternatively, we might just add a CLI command that lists the currently available GPU compute resources. That would also serve as a debug feature.

[lucaro]

I just tried what happens if you include tensorflow-core-platform-gpu on a machine without cuda support. It tries to load cuda, fails, and falls back to the CPU. Apart from some additional error messages in the log and a slightly larger size of the dependencies, there is, therefore, no downside in always loading the gpu-enabled version. The switch is not even necessary. The same behavior can be observed with tensorflow-core-platform-mkl. Interestingly enough, tensorflow-core-platform-mkl-gpu which contains both the intel and the nvidia accelerator support did not work on my machine, otherwise I'd have suggested that as a default.

[lucaro]

I removed the switch and set it directly to use the GPU-enabled dependency. We can merge this now.