Inte Xeon CPU E5-2643 v4 @ 3.4 GHz (two) 64 GM RAM
Windows10 Enterprise version 1803, build 17134.590
NVidia GeForce 1080 graphics cards with GPU (three)
Updated Nvidia graphics driver to 419.35 (driver only, express installation option...not custom)
Used admin privs to instal Nvidia CUDA toolkit 10.1.105_418.96_win10.exe from here https://developer.nvidia.com/cuda-downloads?target_os=Windows&target_arch=x86_64&target_version=10&target_type=exelocal. At first, I did not see any evidence that CUPTI was installed, but it looks like everything needed is in the extras directory.
Following instructions here: https://developer.nvidia.com/rdp/cudnn-download. Downloaded the cuDNN libraries as cudnn-10.1-windows10-x64-v7.5.0.56.zip, and put the zip file in C:\Program Files\NVIDIA GPU Computing Toolkit. This required admin privleges.
Following instructions at https://www.tensorflow.org/install/gpu, I set the paths for C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.0\bin C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.0\extras\CUPTI\libx64
I did not know what this referred to, and did not set it: C:\tools\cuda\bin
Downloaded CUPTI cupti-win64-2019.1.1.1.zip from https://developer.nvidia.com/gameworksdownload#?dn=cuda-profiler-tools-interface-cupti-2019-1 and put the zip file in the same directory. Turns out this was unessecary...the libraries appeard in
Per https://docs.nvidia.com/cuda/cuda-installation-guide-microsoft-windows/
C:\Users\csherwood>nvcc -V nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2019 NVIDIA Corporation Built on Fri_Feb__8_19:08:26_Pacific_Standard_Time_2019 Cuda compilation tools, release 10.1, V10.1.105
After that, I followed the instructions here: https://towardsdatascience.com/tensorflow-gpu-installation-made-easy-use-conda-instead-of-pip-52e5249374bc, which boil down to: conda create --name tf_gpu tensorflow-gpu
That seems to work, because the following test works correctly:
sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))
However, it did not run Dan's Optical Wave Gauge because it could not find the MobilenetV2 app. This may have been a mistake on my part when trying to follow Dan's instructions on moving the model apps.
For Dan's Optical Wave Gauge, I installed his owg environment. Then I did this
conda remove tensorflow
conda install tensorflow-gpu==1.11.0
conda install keras=2.2.4
conda install cudnn
(keras gets uninstalled as a side-effect of removing tensorflow)