R with TensorFlow 2.0 on Debian/sid

I recently posted on getting TensorFlow 2.0 with GPU support running on Debian/sid. At that time I didn’t manage to get the tensorflow package for R running properly. It didn’t need much to get it running, though.

The biggest problem I faced was that the R/TensorFlow package recommends using install_tensorflow, which can use either auto, conda, virtualenv, or system (at least according to the linked web page). I didn’t want to set up neither a conda nor virtualenv environment, since TensorFlow was already installed, so I thought system would be correct, but then, I had it already installed. Anyway, the system option is gone and not accepted, but I still got errors. In particular because the code mentioned on the installation page is incorrect for TF2.0!

It turned out to be a simple error on my side – the default is to use the program python which in Debian is still Python2, while I have TF only installed for Python3. The magic incantation to fix that is use_python("/usr/bin/python3") and one is set.

So here is a full list of commands to get R/TensorFlow running on top of an already installed TensorFlow for Python3 (as usual either as root to be installed into /usr/local or as user to have a local installation):

devtools::install_github("rstudio/tensorflow")

And if you want to run some TF program:

library(tensorflow)
use_python("/usr/bin/python3")
tf$math$cumprod(1:5)

This gives lots of output but mentioning that it is running on my GPU.

At least for the (probably very short) time being this looks like a workable system. Now off to convert my TF1.N code to TF2.0.