Installing and Using Tensorflow with R

This tutorial implements the basic MNIST classification tutorial using the R interface to Keras with Tensorflow backend. In particular, this guide is aimed at using the GPU nodes available on Princeton's HPC Clusters (Adroit and Tiger).

Read the tutorial

The original tutorial is available here under the "MNIST Example" heading.

Clone the repository

Log into a head node on either Adroit or Tiger for GPU node access. Clone this repository to your home directory with

git clone https://github.com/simpsondl/mnist_r_tutorial.git

This creates a folder mnist_r_tutorial.

Create a conda environment with Tensorflow

This will create a conda environment with Tensorflow and Keras available (note that specifying Python 3.6 is necessary to work around a conflict with R).

module load anaconda3
conda create —n mnist python=3.6
conda activate mnist
pip install tensorflow-gpu keras
conda deactivate

Once this finishes, you will have a conda environment with tensorflow that can be accessed when needed. Note that this environment is not active by default, and in order to have R use this environment, we need to add a line to our R scripts.

Configure R environment and download data

From your home directory, run

module load anaconda3
cd mnist_r_tutorial
conda activate mnist
Rscript mnist_download.R

This will check if you have the packages keras, reshape2, and ggplot2 installed. If they are not installed, then the script will stop with an error message indicating which packages need to be installed. Open an R terminal, and install missing packages with

install.packages(c("keras", "reshape2", "ggplot2"))

If you already have some of these installed, then the command can be changed by removing these. If this is your first time installing an R package, you will be asked if you would like to create a personal library in your home directory at a given path--select yes both times. You will also be prompted to select a CRAN mirror; choose one in the US (for example, 64). Once the missing packages have been installed, re-run the mnist_download.R script. If this completes, it should create a new RData file containing the MNIST dataset (mnist_data.RData). Notice that the conda environment we created earlier is specified in the R script using the command use_condaenv("mnist").

Run the classification

Now that the environments have been created and the data has been downloaded, the actual classification can be performed. In mnist.s, remove the space in the --mail-user line and include your Princeton NetID. From the mnist_r_tutorial directory, run

sbatch mnist.s

Notice that this loads two modules, cudatoolkit/10.0 and cudnn/cuda-10.0/7.50; these specific versions are required for R to work correctly. There may be other versions that also work, but this configuration has been tested.

This will request a GPU node for 5 minutes and queue the job. You can check whether or not your job is running with squeue -u yourusername. Once the job has completed, you should have several output files: slurm-XXXXXXX.out containing any printed output from the job (this will include training progress and a model summary), training_metrics.pdf containing a plot of the accuracy and loss values during training, and output.pdf containing MNIST example images with their labels and predictions. Download the PDF files to your computer for viewing with

scp yourusername@hpcclustername.princeton.edu:~/mnist_r_tutorial/*pdf .

Note that this command should be run from your computer (not the cluster), and you will need to change yourusername to your Princeton NetID and hpcclustername to adroit or tiger depending on which HPC cluster you are using. Finally, you can change the download location by changing . to a path of your choice.