applying pointnet++ to ATLAS data.
Running on JLSE 4 V100s using
TRIAL=0 mpirun -hostfile $COBALT_NODEFILE -n 4 python run_pointnet.py -c configs/jlse.json --logdir logdir/$(date "+%Y-%m-%d")/$TRIAL --horovod
After 30 epochs, reached 60% accuracy on training data, 58% on testing data. Image throughput is about 21 images/second/gpu.
With a single process (no MPI), I see about 23 images/second throughput.
The dataset used is found here:
https://opendata-qa.cern.ch/record/15009
After the file is unpacked you can generate the training/validation file lists via:
ls -d /path/to/zej/* | head -n 80000 > training_filelist.txt
ls -d /path/to/zej/* | tail -n 20000 > validation_filelist.txt
Then update the config/jlse.json
config file to use these filelists instead of the existing ones.