Performance benchmarking of ML/AI workloads ResNet, DeepCam, and CosmoFlow; work in support of Architectural Requirements for Deep Learning Workloads in HPC Environments (PMBS21).
Imagenet data: download training and validation datasets from http://www.image-net.org/challenges/LSVRC/2012/downloads
Untar the data with:
tar xf ILSVRC2012_img_val.tar -C $IMAGENET_DATA_HOME/validation
tar xf ILSVRC2012_img_train.tar -C $IMAGENET_DATA_HOME/train
Data must be converted to TFRecords format; this can be done with the script https://github.com/tensorflow/tpu/blob/master/tools/datasets/imagenet_to_gcs.py
python imagenet_to_gcs.py \
--raw_data_dir=$IMAGENET_DATA_HOME \
--local_scratch_dir=$IMAGENET_DATA_HOME/tf_records \
--nogcs_upload
Resnet directory: https://code.ornl.gov/olcf-analytics/summit/distributed-deep-learning-examples
DeepCam directory: https://github.com/sparticlesteve/mlperf-deepcam/tree/nersc-dev
CosmoFlow directory: https://github.com/sparticlesteve/cosmoflow-benchmark