Object Recognition: We should default to GPU enabled tensorflow during install.

[pburma]

In order to enable TensorFlow's use of GPU it is necessary today to manually add in the jni_gpu jar file into the libs directory.

I am probably doing this with unnecessary extra steps because I haven't spent the time to determine how the dependencies impact each other, but today what I do is change the gradle.build file for the objectRecognition project. I have specified the latest version of tensorflow and added the gpu jar reference:

compile "org.tensorflow:tensorflow:1.12.0"
compile "org.tensorflow:libtensorflow_jni:1.12.0"
compile "org.tensorflow:libtensorflow_jni_gpu:1.12.0"
compile "org.apache.commons:commons-math3:3.6.1"

I am pretty sure you need both the _jni.jar and the _jni_gpu.jar but I could be wrong about that.

After the jars are built I will go into the libs directory under the build folder and rename "libtensorflow_jni.jar" to "libtensorflow_jni.original" and then rename the "libtensorflow_jni_gpu.jar" to "libtensorflow_jni.jar".

When I startup the objectRecognition process tensorflow is now using GPU.

IMO - we should be including GPU support by default. This will require having the tensorflow compatible nvidia libraries as specified here:

The following NVIDIA® software must be installed on your system:

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

NVIDIA® GPU drivers —CUDA 10.0 requires 410.x or higher.
CUDA® Toolkit —TensorFlow supports CUDA 10.0 (TensorFlow >= 1.13.0)
CUPTI ships with the CUDA Toolkit.
cuDNN SDK (>= 7.4.1)
(Optional) TensorRT 5.0 to improve latency and throughput for inference on some models.

The support for CUDA 10.0 seems to be recent, up until today I have had to always install CUDA 9.0.

Back to my original request, that we should either run GPU by default or have a flag to make it easy to run in GPU mode so I don't have to go and modify the gradle build file or rename the jars after the install.