How to find installed conda path for tensorflow package in R?

Question

I tried to run keras package from R in RStudio as shown in this tutorial . However, I get stuck by loading data from mnist dataset with the command mnist <- dataset_mnist() as I get the error:

Error: Python module tensorflow.python.keras was not found.

Detected Python configuration:

python:         C:\PROGRA~3\ANACON~1\python.exe
libpython:      C:/PROGRA~3/ANACON~1/python37.dll
pythonhome:     C:\PROGRA~3\ANACON~1
version:        3.7.3 (default, Mar 27 2019, 17:13:21) [MSC v.1915 64 bit (AMD64)]
Architecture:   64bit
numpy:          C:\PROGRA~3\ANACON~1\lib\site-packages\numpy
numpy_version:  1.16.2
tensorflow:     [NOT FOUND]

python versions found: 
 C:\PROGRA~3\ANACON~1\python.exe
 C:\ProgramData\Anaconda3\python.exe

It seems not to find tensorflow package, which is already installed as installed.packages() shows:

           Package      LibPath                                              Version        Priority
keras      "keras"      "C:/Users/NikolasVarela/Documents/R/win-library/3.5" "2.2.4.1.9001" NA      
tensorflow "tensorflow" "C:/Users/NikolasVarela/Documents/R/win-library/3.5" "1.13.1"       NA    

I previously manually installed Anaconda for python 3.7 and downgraded in the Anaconda Navigator to 3.6 as shown in this udemy tutorial as mentioned that tensorflow only runs on python 3.6. After that I downloaded all needed packages including tensorflow and keras on the anaconda navigator. I assume that this manuell change makes the problem in finding tensorflow. But I didn't find a solution on the web.

Further, I tried to set the path of the dataset_mnist() function as mentioned in the description . However, I don't find a ~/.keras/datasets on my directory. Any ideas what I'm doing wrong? Thanks

Answer 1

I had the same problem. I finally resolved it as follows:

1. Downgrade to python 3.6 in the conda environment (you've already mentioned it, but I'm showing how here): conda create -y --name r-tensorflow tensorflow-gpu keras-gpu python=3.6.8 . (Note: I'm not sure it is necessary to install the keras-gpu package here.)
2. Restart R session since it has probably already initialized with the wrong version of Python.
3. Force R to use the conda environment with the downgraded Python: keras::use_condaenv("r-tensorflow", required = TRUE) .

4. Now it should work. Try the following if you don't mind downloading the MNIST database:

    library(keras) mnist <- dataset_mnist() 

PS. I've first tried installing via R as follows:

install.packages("keras")
library(keras)
install_keras(method = c("conda"),
              conda = "auto", version = "default", tensorflow = "gpu")

PPS. Of course you need to already have Python, Anaconda/Miniconda, CUDA and cuDNN installed as well as some other requirements. I'm assuming you have and that you don't have version conflicts.