CosmoFlow TensorFlow Keras benchmark implementation
You can find the previous TensorFlow implementation which accompanied the CosmoFlow paper at https://github.com/NERSC/CosmoFlow
The dataset we use for this benchmark comes from simulations run by the ExaLearn group and hosted at NERSC. The following web portal describes the technical content of the dataset and provides links to the raw data.
For this benchmark we currently use a preprocessed version of the dataset which generates crops of size (128, 128, 128, 4) and stores in TFRecord format. This preprocessing is done using the prepare.py script included in this package. We describe here how to get access to this processed dataset, but please refer to the ExaLearn web portal for additional technical details.
Globus is the current recommended way to transfer the dataset locally. There is a globus endpoint at:
The contents are also available via HTTPS at:
MLPerf HPC v1.0 preliminary dataset
Preprocessed TFRecord files are available in a 1.7TB tarball named
cosmoUniverse_2019_05_4parE_tf_v2.tar. It contains subfolders for
train/val/test file splits.
In this preparation, there are 524288 samples for training and 65536 samples for validation. The TFRecord files are written with gzip compression to reduce total storage size.
MLPerf HPC v0.7 dataset
The pre-processed dataset in TFRecord format is in the
cosmoUniverse_2019_05_4parE_tf folder, which contains training and validation
subfolders. There are 262144 samples for training and 65536 samples
for validation/testing. The combined size of the dataset is 5.1 TB.
For getting started, there is also a small tarball (179MB) with 32 training
samples and 32 validation samples, called
Running the benchmark
Submission scripts are in
scripts. YAML configuration files go in
Running at NERSC
sbatch -N 64 scripts/train_cori.sh