README.md

CosmoFlow

Link to the application: https://portal.nersc.gov/project/m3363/

References:

• https://arxiv.org/abs/1808.04728

How to use (training)

Setup and download dataset

1. Get the SBI-FAIR repository

   git clone --depth 1 https://github.com/DSC-SPIDAL/sbi-fair
   SBI_FAIR_DIR=${PWD}/sbi-fair

2. Create a directory for downloading datasets and store results

   mkdir cosmoflow
   cd cosmoflow
   mkdir output

3. Get the datasets for training

   ${SBI_FAIR_DIR}/tools/scripts/load_dataset.py ${SBI_FAIR_DIR}/datasets/cosmoflow/datasets.yaml cosmoUniverse_2019_05_4parE_tf_v2_mini

   You can use any of the following datasets:

   • cosmoUniverse_2019_05_4parE_tf_v2_mini
   • cosmoUniverse_2019_05_4parE_tf_v2

4. Create a file with parameters

   # Few epochs for testing
   echo 'epochs: 2' > options.yaml 

   We will update the list of available options here, in the meantime please refer to the original repository https://github.com/sparticlesteve/cosmoflow-benchmark for the list of all options.

Using Docker

1. Build Docker container

   cd ${SBI_FAIR_DIR}/models/cosmoflow
   ./build_docker.sh
   cd - # Go back to results directory 

2. Run Training

   GPU_SWITCH='--runtime=nvidia --gpus all' # or '' for CPU workloads
   # Mount the directories with the dataset
   VOLUME_MOUNTS='-v ./cosmoUniverse_2019_05_4parE_tf_v2_mini/default:/input/train_dataset -v ./output:/output -v ./options.yaml:/input/options.yaml'
   docker run ${GPU_SWITCH} ${VOLUME_MOUNTS} cosmoflow run train

Using Apptainer

1. Build Apptainer container

   cd ${SBI_FAIR_DIR}/models/cosmoflow
   ./build_apptainer.sh
   cd - # Go back to results directory 

2. Run Training

   GPU_SWITCH='--nv' # or '' for CPU workloads
   # Mount the directories with the dataset
   VOLUME_MOUNTS='--bind ./cosmoUniverse_2019_05_4parE_tf_v2_mini/default:/input/train_dataset --bind ./output:/output --bind ./options.yaml:/input/options.yaml'
   apptainer run --app train ${GPU_SWITCH} ${VOLUME_MOUNTS} ${SBI_FAIR_DIR}/models/cosmoflow/cosmoflow.sif

Results

The outputs of the run will be available in ./output.