Keras: Documentation should mention that dev. version of reticulate needs to be installed

[setgree]

keras depends on tensorflow which depends on reticulate, but if you just have the CRAN version of reticulate installed, you get the following error message:

  namespace ‘reticulate’ 1.8 is being loaded, but >= 1.8.0.9002 is required
ERROR: lazy loading failed for package ‘tensorflow’
* removing ‘/Library/Frameworks/R.framework/Versions/3.5/Resources/library/tensorflow’
Installation failed: Command failed (1)

update.packages('reticulate') doesn't solve it because the latest on CRAN right now is 1.8.0.

Sorry I didn't open a pull request, I tried to find the underlying Rmd doc. that generated the main page but didn't succeed . Thanks!

[jjallaire]

It's right here: https://keras.rstudio.com/#installation

We currently instruct people to install with devtools::install_github("rstudio/keras"). This will automatically cause reticulate to also be installed from GitHub (due to this in the description: https://github.com/rstudio/keras/blob/master/DESCRIPTION#L28).

Let us know if this is not working as intended for you.

[scottcame]

I am running into the same issue as @setgree, using devtools::install_github("rstudio/keras"). The install from github worked fine 2 days ago. All of a sudden today it has stopped working. I'll dig into it a bit and see if I can fix locally...

Further update: The only way I can get a successful install of the keras package is to install reticulate, then tensorflow, then keras, all via install_github(), in that order. BTW I am doing this in a clean docker container, from image rocker/r-ver:3.5.0, with the tf/keras dependencies (e.g., python, pip, etc.) installed first.

[jjallaire]

I just made a change to the keras DESCRIPTION file to explicitly require the development version of reticulate. I have verified that this works w/ install_github (after removing TF, keras, and reticulate). Hopefully this works for you as well!

[scottcame]

I am still getting a failed install. Here's a dockerfile that replicates the problem:

https://gist.github.com/scottcame/5cbab4725b11a240e9373a705a880b3f

Uncommenting the explicit github install of reticulate and tensorflow results in a successful install of keras.

FYI, I tried doing the image build in play-with-docker, and it stalls out...probably too much to download and install.

But you can replicate locally by creating a new, empty directory somewhere and:

curl -s https://gist.githubusercontent.com/scottcame/5cbab4725b11a240e9373a705a880b3f/raw/3d951a5744b54006600ce5bbbbf7432d1c23a478/Dockerfile | docker build -t r-keras-tmp -f - .

[setgree]

Hi all, thanks for checking in on it @jjallaire. Indeed when I used devtools::install_github("rstudio/keras"), this was unsuccessful at installing the dev. version of retiuclate. It looks like this commmit (https://github.com/rstudio/keras/commit/1488ef8ce44958705ac7c6a8e8b618b073c31f93#diff-35ba4a2677442e210c23a00a5601aba3) is intended to address this, thanks for looking into it!

[jjallaire]

Maybe you are running an older version of devtools that doesn't respect the Remotes entry in the DESCRIPTION file? (although this was added 2+ years ago).

[scottcame]

Looks like the docker image has devtools 1.13.6...the current CRAN version

[jjallaire]

Okay, I just changed the DESCRIPTION to point to the CRAN version rather than dev versions (the dev versions are there for a bugfix but if this prevents installation on some systems it's not worth it). I have no idea why the Remotes field is not being respected here.

[setgree]

Thanks, this works flawlessly now!

[scottcame]

Confirmed...the dockerfile in my gist linked above now results in a successful keras package install. Thanks!