pytorch-gpu-data-science-project

Repository containing scaffolding for a Python 3-based data science project with GPU acceleration using the PyTorch ecosystem.

Creating a new project from this template

Simply follow the instructions to create a new project repository from this template.

Project organization

Project organization is based on ideas from Good Enough Practices for Scientific Computing.

Put each project in its own directory, which is named after the project.
Put external scripts or compiled programs in the bin directory.
Put raw data and metadata in a data directory.
Put text documents associated with the project in the doc directory.
Put all Docker related files in the docker directory.
Install the Conda environment into an env directory.
Put all notebooks in the notebooks directory.
Put files generated during cleanup and analysis in a results directory.
Put project source code in the src directory.
Name all files to reflect their content or function.

Installing NVIDIA CUDA Toolkit

Installing the NVIDIA CUDA Toolkit manually is only required if your project needs to use the nvcc compiler. Note that even if you have not written any custom CUDA code that needs to be compiled with nvcc, if your project uses packages such as PyTorch Geometric that include custom CUDA extensions for PyTorch then you will need nvcc installed in order to build these packages.

If you don't need nvcc, then you can skip this section as conda will install a cudatoolkit package which includes all the necessary runtime CUDA dependencies (but not the nvcc compiler).

Workstation

You will need to have the appropriate version of the NVIDIA CUDA Toolkit installed on your workstation. For PyTorch you should install NVIDIA CUDA Toolkit 10.1 (documentation).

After installing the appropriate version of the NVIDIA CUDA Toolkit you will need to set the following environment variables.

$ export CUDA_HOME=/usr/local/cuda-10.1
$ export PATH=$CUDA_HOME/bin:$PATH
$ export LD_LIBRARY_PATH=$CUDA_HOME/lib64:$LD_LIBRARY_PATH

Ibex

Ibex users do not neet to install NVIDIA CUDA Toolkit as the relevant versions have already been made available on Ibex by the Ibex Systems team. Users simply need to load the appropriate version using the module tool.

$ module load cuda/10.1.243

Building the Conda environment

After adding any necessary dependencies that should be downloaded via conda to the environment.yml file and any dependencies that should be downloaded via pip to the requirements.txt file you create the Conda environment in a sub-directory ./envof your project directory by running the following commands.

export ENV_PREFIX=$PWD/env
conda env create --prefix $ENV_PREFIX --file environment.yml --force

Once the new environment has been created you can activate the environment with the following command.

conda activate $ENV_PREFIX

Note that the ENV_PREFIX directory is not under version control as it can always be re-created as necessary.

If you wish to use any JupyterLab extensions included in the environment.yml and requirements.txt files then you need to activate the environment and rebuild the JupyterLab application using the following commands to source the postBuild script.

conda activate $ENV_PREFIX # optional if environment already active
. postBuild

For your convenience these commands have been combined in a shell script ./bin/create-conda-env.sh. Running the shell script will set the Horovod build variables correctly, create the Conda environment, activate the Conda environment, and built JupyterLab with any additional extensions. The script should be run from the project root directory as follows. follows.

./bin/create-conda-env.sh

Listing the full contents of the Conda environment

The list of explicit dependencies for the project are listed in the environment.yml file. To see the full lost of packages installed into the environment run the following command.

conda list --prefix $ENV_PREFIX

Updating the Conda environment

If you add (remove) dependencies to (from) the environment.yml file or the requirements.txt file after the environment has already been created, then you can re-create the environment with the following command.

$ conda env create --prefix $ENV_PREFIX --file environment.yml --force

Using Docker

In order to build Docker images for your project and run containers with GPU acceleration you will need to install Docker, Docker Compose and the NVIDIA Docker runtime.

Detailed instructions for using Docker to build and image and launch containers can be found in the docker/README.md.

yunhaj47/pytorch-gpu-data-science-project