ADCIRC ROM (Reduced-Order-Modeling) is suite of tools for developing surrogate machine learning models of storm surge. The tools can be used both in an HPC and a single-threaded environment.
-
Start a Jupyter lab session.
-
Use Jupyter lab to launch a terminal, and in the terminal run the following:
pip install adcirc-rom
arom dataset setup
This will create a data folder with the subdirectories datasets, storms and models.
These subdirectories are needed for the model development workflow. The datasets directory is used for storing
machine-learning ready datasets. The storms directory will contain the raw ADCIRC input (note when run within the DesignSafe
environment this directory will be prepopulated with a dataset of 446 synthetic ADCIRC simulations).
Finally, the models dataset is used for storing saved ML models and predictions.
- To generate a dataset, run the command
arom dataset create default
This will create a dataset named 'default' in the directory data/datasets.
This dataset can be used to train machine learning models.
The dataset.py script takes a number of options that control the size and scope of the generated dataset,
as well as the included features.
Note: with the default settings, dataset creation in the Designsafe environment will take a few hours due to the lack of MPI support and the size of the data to be processed. The dataset generation script supports parallization with MPI - and is significantly faster when run on HPC resources such as TACC.
- To train and save a new model named 'xgb_base' using XGBoost for both classification and regression, and using the dataset named default, run the command
arom model train --modelname=xgb_base --dataset=default --regressor=xgb250 --classifier=xgb250
This will create a new model named 'xgb_base'. During, training, a portion of the dataset is set aside for testing purposes - predictions are generated for the test dataset and saved alongside the model binary. Additional model training parameters can be passed to the script.
Finally, to generate predictions on a new dataset using a saved model, run
arom model predict [modelname] --dataset=[datasetname]
All predictions can be accessed in the folder data/datasets/[modelname].
Please reach out with any questions or bug reports to Benjamin Pachev [email protected].