I used to create a shiny-app to conduct image classification using Tensorflow on DigitalOcean, however it didn’t work on my own PC station. The main problem is: RStudio-server and Shiny-server can’t find the shared object: libcudart.so.8.0 which located under /usr/local/cuda/lib64 and should be binded to environment variable: “LD_LIBRARY_PATH”. Recently, I noticed that the R package tensorflow has released, so I tried that package, and fixed the previous problem.

There are three scenarios: R console, RStudio-server and Shiny-server, each has its only LD_LIBRARY_PATH value.

R console

To open R console, just type R in terminal. Then use the R command: Sys.getenv(“LD_LIBRARY_PATH”) to check LD_LIBRARY_PATH’s value. You will see that “/usr/local/cuda/lib64” is included if your Tensorflow has been setup correctly. Then, we can conduct the image classification in R console using command: system(“python /home/tian/models/tutorials/image/imagenet/classify_image.py –image_file /tmp/imagenet/cropped_panda.jpg”, intern = TRUE)

RStudio-server

If we run command Sys.getenv(“LD_LIBRARY_PATH”) from RStudio-server, you may get result like: “/usr/lib/R/lib::/lib:/usr/lib/x86_64-linux-gnu:/usr/lib/jvm/default-java/jre/lib/amd64/server”, which not include “/usr/local/cuda/lib64”. To add the directory to LD_LIBRARY_PATH, we need add a line rsession-ld-library-path=/usr/local/cuda/lib64 in “/etc/RStudio/rserver.conf”, then restart RStudio-server.

sudo nano /etc/RStudio/rserver.conf
## rsession-ld-library-path=/usr/local/cuda/lib64
sudo rstudio-server restart

Now, run command Sys.getenv(“LD_LIBRARY_PATH”) again, you will find the path “/usr/local/cuda/lib64” has included. So we can run image classification in RStudio-server using command: system(“python /home/tian/models/tutorials/image/imagenet/classify_image.py –image_file /tmp/imagenet/cropped_panda.jpg”, intern = TRUE)

Shiny-server

Though we can driver Tensorflow from R console and RStudio-server, however the shiny-app still doesn’t work. Include a line in shiny-app like: writeLines(Sys.getenv(“LD_LIBRARY_PATH”), “log.txt”) will tell us the value of “LD_LIBRARY_PATH” for Shiny-server, and we will find that the paths not include “/usr/local/cuda/lib64”. I tried to find the configure file to setup “LD_LIBRARY_PATH” for Shiny-server, but didn’t succeed. Luckily, we can setup using R command: Sys.setenv(), the only inconvenience is that we always need to include that command for Tensorflow related shiny-app. For example, include this line into the first line of server.R

Sys.setenv(LD_LIBRARY_PATH="/usr/local/cuda/lib64:/usr/lib/R/lib:/usr/lib/x86_64-linux-gnu:/usr/lib/jvm/default-java/jre/lib/amd64/server")

After above setting, we may encounter the problem that the shiny-app still doesn’t work. The problem is that directory /tmp is not accessible for shiny. We can change the access permissions of that folder to let user shiny can access.

sudo chmod 777 -R /tmp