Skip to content

neuston_sbatch

Sidney Batchelder edited this page Mar 15, 2021 · 1 revision

The neuston_sbatch.py module allows a user to create and submit a neuston_net.py TRAIN or neuston_net.py RUN sbatch script on a SLURM enabled system.

usage: neuston_sbatch.py [-h] [--job-name STR] [--email EMAIL] [--walltime HH:MM:SS] 
                         [--gpu-num INT] [--cpu-num INT] [--mem-per-cpu MB] 
                         [--slurm-log-dir DIR] [--ofile OFILE] [--dry-run] 
                         [--batch SIZE] [--loaders N] {TRAIN,RUN} ...

SLURM SBATCH auto-submitter for neuston_net.py

positional arguments:
  {TRAIN,RUN}          These sub-commands are mutually exclusive. The sub-commands are identical to 
                       the TRAIN and RUN commands from "neuston_net.py"
                       Note: optional arguments (below) must be specified before "TRAIN" or "RUN"
    TRAIN              Train a new model.
    RUN                Run a previously trained model.

optional arguments:
  -h, --help           show this help message and exit

SLURM Args:
  --job-name STR       Job Name that will appear in slurm jobs list. Defaults is "NN"
  --email EMAIL        Email address for slurm notifications. The default is "{USERNAME}@whoi.edu"
  --walltime HH:MM:SS  Set Slurm Task max runtime. Default is "24:00:00"
  --gpu-num INT        Number of GPUs to allocate per task. Default is 1
  --cpu-num INT        Number of CPUs to allocate per task. Default is 4
  --mem-per-cpu MB     Memory to allocate per cpu in MB. Default is 10240MB
  --slurm-log-dir DIR  Directory to save slurm log file to. 
                       Defaults to OUTDIR (as defined by TRAIN or RUN subcommand)
  --ofile OFILE        Save location for generated sbatch file. 
                       Defaults to "{OUTDIR}/{PID}.{JOB_NAME}.sbatch"
  --dry-run            Create the sbatch script but do not run it

NN Common Args:
  --batch SIZE         Number of images per batch. Defaults is 108
  --loaders N          Number of data-loading threads. 4 per GPU is typical. Default is 4