gpuDL-docker - Docker container for Deep Learning in R and Python with GPU support

This Dockerfile sets up a complete environment for experimenting with R, Python and the most popular Deep Learning libraries (Tensorflow, Keras, Theano).

It installs:

• R 4.0.3 (r-base)
• Python 3.7
• Miniconda 3
• Jupyter notebook for Python

It additionally installs the following packages

Python packages

• h5py
• pandas
• pygpu==0.6.2
• nose
• mkl
• six
• pyyaml
• keras==2.2.4
• tensorflow-gpu==1.13.1
• ipyparallel
• jupyter
• matplotlib
• seaborn
• scikit-learn
• scikit-cuda
• plotly

R Packages

• r-base-dev and [https://packages.ubuntu.com/xenial/r-recommended r-recommended] Ubuntu packages
• yhatr
• forecast
• stringr
• randomForest
• lubridate
• rpart
• e1071
• kknn
• ggplot2
• plyr
• reshape2
• devtools
• dse
• autoencoder
• pls
• MTS
• rnn
• feather
• data.table
• dplyr
• ranger
• zoo
• plotly
• gmatrix
• HiPLARM
• HiPLARb
• onlinePCA
• gputools
• gbonte/gbcode (Github)
• rstudio/keras@2.2.4.1 (Github)
• IRKernel/IRKernel (Github)

CUDNN support

The image is based on the nvidia/cuda:10.0-cudnn7-devel-ubuntu18.04 from the official NVIDIA CUDA Dockerhub including CUDA 10.0 and cuDNN7, with the respective headers.

Quickstart

git clone https://github.com/jdestefani/gpuDL-docker.git
cd gpuDL-docker
nvidia-docker run -it -v `pwd`/docker_volume:/root/shared_data -p #PORT#:8888 jdestefani/gpu_dl:latest

Note:

• Important The docker depends on nvidia-docker to access the underlying NVIDIA hardware. Running with regular docker will not allow to use GPU.
• The "-v pwd:/root/shared_data" shares the folder docker_volume on your computer (the 'host') with the container in the '/root/shared_data' folder
• Ports are shared as follows:
  • 8888 bridges to the Jupyter Notebook
• #PORT# should be replaced with the port on which the default jupyter port (8888) should be redirected.

Build and running the container from scratch

Clone this repository

git clone https://github.com/jdestefani/gpuDL-docker.git

Build

From Dockerfile folder, run

docker build --rm=true -t gpu_dl .

It may take about 30 minutes to complete.

Run

nvidia-docker run -it -v `pwd`/docker_volume:/root/shared_data -p #PORT#:8888 gpu_dl

How to verify that the Docker is working properly?

0. Make sure that git and nvidia-docker are installed.
1. Run docker container in interactive mode (-it cf. Quickstart)
2. Run the following commands and check the correct execution of the code:

cd shared_data/samples
python keras_CNN_MNIST.py
Rscript keras_CNN_MNIST.R