Installation Issue: tensor flow in rocker/ml docker image, ModuleNotFoundError: No module named 'tensorflow'

[sneumann]

[thanks @t-kalinowski for your response, I was since editing my issue and adding the workaround]
Hi,
I am installing (and testing) tensorflow using

R -e 'install.packages("reticulate"); remotes::install_github("rstudio/tensorflow", upgrade="never"); library(tensorflow); install.packages("keras"); library(keras); install_keras(); model.sla <- keras_model_sequential(input_shape = c(NULL, 12906, 1))'

This works nicely in an older rocker/rstudio image based on Ubuntu22.04.4, with R-python 3.10.12, R-4.4.1
This fails on the rocker/ml images based on Ubuntu22.04.3 with nvidia cuda stuff installed, also with R-python 3.10.12 and R-4.4.1

I have rocker-org/rocker-versioned2#880 as initially I thought that would be a rocker/ml issue.
I have a workaround now to install as shown above, and for usage

## Does not work:
R -e 'tensorflow::as_tensor("Hello World")' ; 

## Workaround:
R -e 'reticulate::use_virtualenv("/root/.virtualenvs/r-tensorflow", required = TRUE); library(tensorflow); tensorflow::as_tensor("Hello World")'

So all left now is possibly to add your suggestions to https://tensorflow.rstudio.com/install/

I also leave below my output from debugging. I was curious about the error for tf_config(),
this comes because python_version can be print()ed, but not cat()ed.

Yours,
Steffen

> reticulate::py_config()
python:         /root/.virtualenvs/r-tensorflow/bin/python
libpython:      /usr/lib/python3.10/config-3.10-x86_64-linux-gnu/libpython3.10.so
pythonhome:     /root/.virtualenvs/r-tensorflow:/root/.virtualenvs/r-tensorflow
version:        3.10.12 (main, Nov 20 2023, 15:14:05) [GCC 11.4.0]
numpy:          /root/.virtualenvs/r-tensorflow/lib/python3.10/site-packages/numpy
numpy_version:  1.26.4

# R -e 'reticulate::use_virtualenv("/root/.virtualenvs/r-tensorflow", required = TRUE); library(tensorflow);  tensorflow::tf_config()'
> tensorflow::tf_config()
TensorFlow v2.15.1 ()
Python vError in cat("Python v", x$python_version, " (", aliased(x$python), ")\n",  : 
  argument 2 (type 'list') cannot be handled by 'cat'

Digging into the tf_config() error above:

> cat(tensorflow::tf_config()$python_version)
Error in cat(tensorflow::tf_config()$python_version) : 
  argument 1 (type 'list') cannot be handled by 'cat'
> print(tensorflow::tf_config()$python_version)
[1] ‘3.10’
> str(tensorflow::tf_config()$python_version)
Classes 'package_version', 'numeric_version'  hidden list of 1
 $ : int [1:2] 3 10

[t-kalinowski]

[Sorry, premature send, there was a new-style GitHub send button, which I mis-interpreted as "save draft". I'll update and polish this issue, until then please consider as WIP]

At a glance, it seems this is a "reticulate is finding the wrong Python installation" issue. Consulting https://rstudio.github.io/reticulate/articles/versions.html#order-of-discovery is usually the quickest way to resolve these.

Also, it looks like you're using {keras} with the latest tensorflow. Note that the latest tensorflow depends on {keras3}, and new code today should generally prefer to use {keras3}. If you really want to use "legacy" keras with the latest tensorflow, you have to take some extra steps (py_install("tf-keras"); Sys.setenv(TF_USE_LEGACY_KERAS=1) probably some other steps too (not tested, but it wouldn't surprise me if import("tf_keras", as = "keras") is also needed).