GOES 16 Lightning Count Prediction Benchmark

The GOES 16 Lightning Count Prediction benchmark is a deep learning benchmark for HPC systems used for atmospheric science problems.

Contributors

• David John Gagne
• Bill Anderson
• Gunther Wallach
• Bill Petzke

Requirements

The code is designed to run on Python 3.6 and 3.7. It requires the following Python libraries:

• numpy
• scipy
• pandas
• xarray
• tensorflow>=2.0.0
• scikit-learn
• pyproj
• dask distributed (for data processing)
• ipython
• cartopy
• pyyaml
• jupyter (for interactive visualization of neural networks)

Setup from Scratch

• Install Python 3.7 on your machine. I recommend the Miniconda Python installer available here.

• Create the benchmark environment: conda env create -f environment.yml This will automatically install all dependencies, including the latest version of tensorflow.

• Once the environment is installed activate it on your machine: source activate goes

• Make sure the CUDA kernel and CUDA toolkit are installed on your system and know the path and versions.

Run Benchmark Script

• Clone the goes16ci git repository to your home directory.

cd ~
git clone https://github.com/NCAR/goes16ci.git
cd goes16ci

• Install the goes16ci library

pip install .

• Download the GOES16 patch files. You will need about 8 GB free to download and untar the data.

python download_data.py

• Run the benchmark script. The script will output trained neural networks and a yaml file with the timing information for each step.

python goes16_deep_learning_benchmark.py

• If you want to modify the neural network or other properties of the script, you can make a copy of benchmark_config_default.yml and modify it. To run the script with the new config file:

python goes16_deep_learning_benchmark.py -c benchmark_config_default.yml

Setup on Cheyenne/Casper

• Clone the git repo to your home directory

cd ~
git clone https://github.com/NCAR/goes16ci.git
cd goes16ci

• Create a link to the patch data on GLADE

ln -s /glade/p/cisl/aiml/dgagne/goes16_nc/ABI_patches_20190315 data

• Modify the goes16_benchmark_casper.sh script with your account number.

• Submit the benchmark script to the casper queue: sbatch goes16_benchmark_casper.sh