Keras: Keras tries do reinstall tensorflow when I only have the cpu version installed

[cderv]

Hi,

We are working on servers with no GPU. Hence, our users use the CPU version of tensorflow.

What we did to install is

library(tensorflow)
install_tensorflow(envname="r-tensorflow",
                   version = "cpu",
                   python = "/applis/Python/3.6.10/bin/python3", 
                   ignore_installed = FALSE)

Everything is fine and we get tensorflow-cpu installed.

Collecting tensorflow-cpu==2.2.0

But when trying to test the installation with the keras tutorial

library(keras)
# load data (we are on offline servers so they are cached)
c(c(x_train, y_train), c(x_test, y_test)) %<-% dataset_mnist(path = "/home/share/echange/data/mnist.npz")

keras will reinstall tensorflow

Using virtual environment '/home/bureau.si.interne/dervieuxchr/.virtualenvs/r-tensorflow' ...
Looking in indexes: https://devin-depot.rte-france.com/repository/pypi-all/simple
Collecting tensorflow

Looking with pip list in the virtualenv I have both

tensorflow                2.2.0
tensorflow-cpu         2.2.0

This will cause tensorflow to be used, and I have what I think to be GPU related errors.

With tensorflow-cpu only, everything is working fine.

note that this happens only the first time I use

c(c(x_train, y_train), c(x_test, y_test)) %<-% dataset_mnist(path = "/home/share/echange/data/mnist.npz")

in a R session.

I don't really know what happens 🤔 Is there a check somewhere for tensorflow when fist executing a Keras function ?

[cderv]

In fact it will happen with any first keras function I use... There seems to be a check when keras is first loaded, and this check will not look for the tensorflow-cpu version and so will reinstall it using the default tensorflow version. Am I right ?

(I tried to skip the data_mnist step by loading the variables from a RData, but I got the same situation using

x_train <- array_reshape(x_train, c(nrow(x_train), 784))

executing this line the first time will install tensorflow even if I have tensorflow-cpu in the same version in my virtualenv.

Is this intended ?

[dfalbel]

This behavior is a bug.
What happens is the: https://rstudio.github.io/reticulate/reference/configure_environment.html
and in the meantime you can disable it by setiing the RETICULATE_AUTOCONFIGURE env var to FALSE

[cderv]

Thanks for the workaround, I did not know that.

However, when I set this variable, I got this error

library(keras)
Erreur : package or namespace load failed for ‘keras’: .onLoad a échoué dans loadNamespace() pour 'tensorflow', détails : appel : py_module_import(module, convert = convert) erreur : ModuleNotFoundError: No module named 'tensorflow'

I correctly have tensorflow-cpu in my environment (from what I see from pip list)

[dfalbel]

What happens if you import tensorflow directly from python ins this environment?
Many times this message is misleading and the problem is that there was an error loading tensorflow.

[cderv]

I get an error

>>> import tensorflow
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ModuleNotFoundError: No module named 'tensorflow'

let me try to reinstall tensorflow-cpu in a clean virtualenv 🤔

[cderv]

Ok my bad. The virtualenvironment was corrupted by the automatic installation then the removal.
I redid an installation of tensorflow-cpu and with the env var set to FALSE, I don't have the automatic installation behavior.
Thanks, this workaround will help while waiting for a fix on CRAN for this bug.

[dfalbel]

Glad it worked! Yes, will work on a fix for this!

[dfalbel]

Tracking here now: https://github.com/rstudio/reticulate/issues/791