singularity-pytorch-gpu

Singularity image for a deep learning (pytorch) environment + GPU support (cuda-10.2). Contains libraries to perform common ML tasks. Openslide is included to manipulate whole-slide histology images, imagemagick for general image manipulation. JupyterLab and code-server (VS Code) are also included in the image. This image has been tested in an HPC (SGE) with distributed pytorch applications.

Installing singularity

To install singularity, see the official docs.

Building/downloading the image

To build an image called torchenv.sif based on the definition file Singularity.1.0.0, an NVIDIA GPU and cuda-10.2 drivers must be available on the host system. Clone this repository, move into it and run the singularity build command.

git clone https://github.com/manuel-munoz-aguirre/singularity-pytorch-gpu.git && \
cd singularity-pytorch-gpu && \
sudo singularity build torchenv.sif Singularity.1.0.0

Otherwise, the image can be pulled directly from singularity hub:

singularity pull torchenv.sif shub://manuel-munoz-aguirre/singularity-pytorch-gpu:1.0.0

Using the container

To spawn an interactive shell within the container, use the command below. The --nv flag setups the container to use NVIDIA GPUs (read more here).

singularity shell --nv torchenv.sif

To run a script (for example, script.py) using the container without starting an interactive shell:

singularity exec --nv torchenv.sif python3 script.py

The container can also be launched and used on a system without a GPU, but upon startup it will display a warning about missing NVIDIA binaries on the host.