Introduction

Introduction#

CONUS 404-based and FastEddy-based limited area models have been added to credit. The models are developed and supported by NCAR/RAL.

RAL GWC AI model updates#

  • The following scripts are added to credit.datasets

les_singlestep.py # single-step training  / inference datasets for FastEddy data
wrf_singlestep.py # single-step training / inference datasets for the CONUS 404 data
wrf_multistep.py # multi-step training dataset for the CONUS 404 data

  • credit.trainers

trainerLES.py # single-step training routine
trainerWRF.py # single-step training routine
trainerWRF_multi.py # multi-step training routine

  • credit.losses

les_loss.py # training loss class

  • credit.models

swin_wrf.py # CONUS 404 model for prediction
dscale_wrf.py # CONUS 404 model for creating initialization

Test new features#

Go to the ksha branch, copy the following two folders to your directory, modify the directory information, and you can start testing new features.

/glade/work/ksha/DWC_runs/TEST_WRF/
/glade/work/ksha/DWC_runs/TEST_LES/

Regional AI models#

  • Single-step training using a single Casper GPU:

credit_train_wrf -c /glade/work/ksha/DWC_runs/TEST_WRF/model_single_none.yml

  • Single-step training using FSDP on 4 Derecho GPUs:

qsub /glade/work/ksha/DWC_runs/TEST_WRF/launch_single.sh

  • Multi-step training using FSDP on 4 Derecho GPUs:

qsub /glade/work/ksha/DWC_runs/TEST_WRF/launch_multi_01.sh

  • Inference using a single Casper GPU:

credit_rollout_wrf -c /glade/work/ksha/DWC_runs/TEST_WRF/model_predict_none.yml

  • Inference using FSDP on 4 Derecho GPUs:

qsub /glade/work/ksha/DWC_runs/TEST_WRF/launch_predict.sh

LES models#

  • Single-step training using a single Casper GPU:

credit_train_les -c /glade/work/ksha/DWC_runs/TEST_LES/model_single_none.yml

  • Single-step training using FSDP on 4 Derecho GPUs:

qsub /glade/work/ksha/DWC_runs/TEST_LES/launch_single.sh

  • Inference using a single Casper GPU:

credit_rollout_les -c /glade/work/ksha/DWC_runs/TEST_LES/model_predict_none.yml

  • Inference using FSDP on 4 Derecho GPUs:

qsub /glade/work/ksha/DWC_runs/TEST_LES/launch_predict.sh