tensordocker

This repository provides a docker installation of R, rstudio, BioConductor together with maplet and autonomics in a CUDA enabled keras/tensorflow image. The github page for the image is here.

Running the docker image

To install the image from the command line:

docker pull ghcr.io/karstensuhre/tensordocker:latest

To run the docker image (adapt the -v option to mount the required local directory):

docker run -v /home:/home/rstudio/host -it --detach --name tensor -p8888:8888 -p8787:8787 ghcr.io/karstensuhre/tensordocker:latest
docker exec tensor rstudio-server start

or using WSL

docker.exe run -v "C:\\Users":/home/rstudio/host -it --detach --name tensor -p8888:8888 -p8787:8787 ghcr.io/karstensuhre/tensordocker:latest
docker.exe exec tensor rstudio-server start

Note that this command mounts the entire home directory (-v option). You may want to change this to a more limited scope.

To access the rstudio server:

• http://localhost:8787 (rstudio interface, user: rstudio, password: pwd)

To access the tensorflow jupyter notebook:

• http://localhost:8888 (tensorflow jupyter interface, token required - see below)

To obtain the jupyter login token:

docker exec tensor jupyter notebook list

To obtain a shell in the container:

docker exec -it tensor /bin/bash

Creating the docker image from scratch

Note: this part is ONLY needed if you wish to recreate your own docker image. It is NOT needed if you download the docker image from github.io (as explained above). The script make_tensordocker.sh can be used to generate the docker image from scratch. Check out the head of make_tensordocker.sh where I left useful comments and links.

To run the script interactively in steps and answer Y/N:

./make_tensordocker.sh

To run the entire script without interactive prompting to create the image in one go:

./make_tensordocker.sh all

Working with the docker image

Caution: When using library(keras) in R for the first time, DO NOT install a python library. It is already there. Answer NO to the install prompt!

To test tensorflow:

tensorflow::tf_gpu_configured()

You can run mnist_example.R for testing keras/tensorflow.

To update/install maplet from the latest commit:

devtools::install_github(repo="krumsieklab/maplet", subdir="maplet")

Using GPU support

To use NVIDIA GPU support, CUDA needs to be installed first. Then run the following as root (once after boot):

sudo nvidia-persistenced

TO check if CUDA is working alright:

nvidia-smi

To start the docker image with GPU support use the --gpus option:

docker run --gpus all -v `pwd`:/home/rstudio/host -it --detach --name tensor -p8888:8888 -p8787:8787 ghcr.io/karstensuhre/tensordocker:latest

Using rstudio you can then pull this github repository using the GIT functionality of R and then run mnist_example.R for testing the performance of the GPU. FOr comparision, there is also a python version that performs the same actions mnist_example.py. It can runs be executed inside rstudio.