R: How to get a Long short term memory (LSTM) neural network running on Windows 10: installation guide and an example program in R - PLC-Programmer/R GitHub Wiki

keywords: R, Windows 10, python, miniconda, tensorflow, keras

Wikipedia: what is a LSTM? --> Long short-term memory

(A) what are the prerequisites?

  • the host CPU must support AVX2 instructions. So, a very old CPU cannot run tensorflow because trying to load the native TensorFlow runtime will fail. At this point you are required to look for another piece of hardware.

(B) on some preconditions:

  • an existing python installation or environment respectively, no matter in what form, doesn't hurt. So leave it as it is and do not try to get it running from your R installation
  • the idea here is that the needed python environment will be installed (efficiently) while executing the R code (here R version 4.X.X)
  • so, execute the R code (from the "LSTM test.R" program) until you come to this line:

model <- keras_model_sequential()

(C) installing the needed python environment:

When above R code line has been entered (to be executed), R probably cannot find the path to a suitable python environment and throws out these messages:

No non-system installation of Python could be found.

Would you like to download and install Miniconda?

...

Would you like to install Miniconda? [Y/n]:

Now answer "Y".

After a hopefully successful installation of Miniconda, a corresponding environment with name "r-reticulate" should exist in this Windows directory:

C:\Users<user>\AppData\Local\r-miniconda\envs\r-reticulate

(D) setting up the corresponding python environment:

Most probably the "tensorflow" package is still missing from the python environment with name "r-reticulate".

So, you now leave your R environment for a while and turn to this python environment. You do this for example by doing a Windows search for "miniconda" and open the App "Anaconda Prompt (Miniconda 3)".

Now you activate the "r-reticulate" environment:

> conda activate r-reticulate

If this doesn't work, then try with the absolute path:

> conda activate C:\Users<user>\AppData\Local\r-miniconda\envs\r-reticulate

Now you install tensorflow:

(r-reticulate) C:\Users<user>> pip install tensorflow

A note: tensorflow might not like the latest python version!!

  • at the moment Python 3.7.7 is OK

At this point you can make some very simple tests for tensorflow in this python environment:

(r-reticulate) C:\Users<user>> python

>>> import tensorflow as tf

=> this import command must not fail!

A reason for failure could be a missing Visual C++ Redistributable on your Window host.

Installing "Visual C++ Redistributable 2019, 64Bit" will fix this problem. So, install on your Window host:

https://support.microsoft.com/en-us/help/2977003/the-latest-supported-visual-c-downloads

Now try again:

>>> import tensorflow as tf

(do not worry here for a potential error message like this: "Could not load dynamic library 'cudart64_101.dll'; dlerror: cudart64_101.dll not found". This only works on a host with a GPU. If not available, tensorflow will use the CPU.)

>>> import tensorflow.compat.v1 as tfc

>>> sess = tfc.InteractiveSession()

>>> my_tensor1 = tfc.random_uniform((4, 4), 0, 1)

>>> print(my_tensor1)

>>> my_tensor2 = tfc.constant('Hello World')

>>> print(my_tensor2)

Now exit python (quit()) and return to the miniconda prompt to install some other packages:

(r-reticulate) C:\Users<user>> conda install libpython

(r-reticulate) C:\Users<user>> conda install m2w64-toolchain

(r-reticulate) C:\Users<user>> conda install pip install --upgrade keras

At last return to python and test Keras:

>>> from keras import backend

>>> dir(backend)

(this dir command should result in a meaningful answer)

>>> quit()

At last deactivate the current python environment: (r-reticulate) C:\Users<user>> conda deactivate

Now you can close this miniconda prompt.

(E) finally make it all happen in R!

Whenever you run R code which makes calls to the "r-reticulate" python environment, it happens automatically (from now on).

Or in other words: from now on you can concentrate on the R code. There's no need to open miniconda again and manually activate the "r-reticulate" python environment or operations like this.

Now return to your R environment and run again above command and after that just go on with the remaining "LSTM test.R" program:

model <- keras_model_sequential()

=> it should all sail smoothly on:

Hardcopy plot of a program run (#1)

Hardcopy plot of a program run (#2)

Hardcopy plot of a program run (#3)

__end

⚠️ **GitHub.com Fallback** ⚠️