R Gui crashes with Keras Tensorflow windows cuda gpu - r

Question

edit: solved. I was not using the proper version of
cudnn. After downloading and copying over cuDNN v5.1 Library for Windows 10, everything works fine so far. I should have followed tensorflow windows guidelines more
closely.
I just spent some time to set up a windows system to run
r-tensorflow with keras and tensorflow as a backend.
I followed these blog instructions, and I am able to get the hello world example to work. However, when I continue to some R keras examples
, several will just crash and exit RGui during running the training phase,
e.g. mnist_cnn fails at the training model block.
# train and evaluate
model %>% fit(
x_train, y_train,
batch_size = batch_size,
epochs = epochs,
verbose = 1,
validation_data = list(x_test, y_test)
)
A Visual Studio JIT popup shows "An unhandled win32 exception occured in Rgui.exe[10456]." And VS editor shows, "Unhandled exception at 0x00007FFFEA2734BE (ucrtbase.dll) in Rgui.exe: Fatal program exit requested."
There is very little load on CPU and or GPU during runtime
(less than 30%). Any ideas on how to further debug are appreciated.
Hardware/Software/Setup:
windows 10 Pro 64 bit OS x64 cpu
intel i7-7700 #3.60GHz
64G ram
R 64 bit 3.4.1
Anaconda 3/Python 3.5.4
Nvidia 1060 6G
CUDA 8.0 V8.0.60
Visual Studio 2015 community edition
using
library(tensorflow)
use_condaenv("r-tensorflow")
library(keras)
I also just ran in Rstudio w/ rdesktop log diagnostics on the file and got the following log after crash (I see it cannot find file, but not certain what file).
20 Aug 2017 01:29:32 [rdesktop] ERROR system error 2 (The system cannot find the file specified);
OCCURRED AT: virtual void rstudio::core::http::NamedPipeAsyncClient::connectAndWriteRequest() C:/Users/Administrator/rstudio/src/cpp/core/include/core/http/NamedPipeAsyncClient.hpp:84;
LOGGED FROM: void rstudio::desktop::NetworkReply::onError(const rstudio::core::Error&) C:\Users\Administrator\rstudio\src\cpp\desktop\DesktopNetworkReply.cpp:288
20 Aug 2017 01:29:42 [rdesktop] ERROR system error 2 (The system cannot find the file specified);
OCCURRED AT: virtual void rstudio::core::http::NamedPipeAsyncClient::connectAndWriteRequest() C:/Users/Administrator/rstudio/src/cpp/core/include/core/http/NamedPipeAsyncClient.hpp:84;
LOGGED FROM: void rstudio::desktop::NetworkReply::onError(const rstudio::core::Error&) C:\Users\Administrator\rstudio\src\cpp\desktop\DesktopNetworkReply.cpp:288
edit: stepping through several more exampes, several seem to work fine -- it could just be specific to the particular code and my setup. I don't know enough right now to troubleshoot the code ideosyncracies. Those that crash also seem to do so at the model fitting stage. I'll update a list on crash/no-crash as I test. pass = no-crash, fail = crash and exit
mnist_hierarchical_rnn # pass
mnist_transfer_cnn # fail
variational_autoencoder_deconv # fail
stateful_lstm # pass
variational_autoencoder # pass
reuters_mlp # pass
mnist_mlp # pass
mnist_irnn # pass
mnist_cnn # fail
mnist_antirectifier # fail