Ptype-Physical

This package contains code for generating machine learning estimates of winter precipitation type from vertical profiles of the atmosphere.

Installation

1. Clone the repository from github git clone git@github.com:ai2es/ptype-physical.git
2. Within the terminal, go to the top-level directory with cd ptype-physical.
3. Install miniconda, then create a ptype environment with the following command: conda env create -f environment.yml.
4. Alternatively, add the dependencies to an existing environment with conda env update -f environment.yml.
5. Activate the environment by running conda activate ptype or source activate ptype.
6. Install the package directly: pip install . or link the package if you are debugging pip install -e .

Train a classifier model to predict ptypes

Train a multi-layer perceptron model via

python applications/train_mlp.py -c config/ptype.yml

Upon completion, the model will save predictions on training, validation and test splits to file. Next compute metrics on the splits and case studies via

python applications/evaluate_mlp.py -c config/ptype.yml

Two types of approaches are supported: (1) A standard MLP trained with cross-entropy, and (2) evidential approach trained with Dirichlet loss.

• Option (1) may be selected by setting the loss to "categorical_crossentropy" and the output_activation to "softmax" in ptype.yml.

• Option (2) may be selected by setting the loss to "dirichlet" and the output_activation to "linear" in ptype.yml.

For more details,

python applications/train_mlp.py --help
python applications/evaluate_mlp.py --help

Active training

Active training can be called with Option (1) via

python applications/active_training.py -c config/ptype.yml -i 20 -p "random"

which will perform 20 iterations with a random policy. The number of cross-validation steps at each iteration is controlled using "n_splits" in the config file.

One may also select "mc-dropout" for the policy, and use the parser option -s to set the number of Monte Carlo iterations:

python applications/active_training.py -c config/ptype.yml -i 20 -p "mc-dropout" -s 100

For option (2), the supported policies are

"evidential", "aleatoric", "epistemic", "random", "mc-dropout"

The script can also be run on multiple GPU nodes for either option via

python applications/active_training.py -c config/ptype.yml -i 20 -p "random" -l 1 -n 20

which will launch 20 PBS jobs to GPU nodes.

Once all iterations have completed, the results are currently viewed using notebooks/compare_active_training.ipynb

For more details,

python applications/active_training.py --help

Config file

Inference

The P-type models can currently be run historically on using the High Resolution Rapid Refresh (HRRR), Rapid Refresh (RAP), or Global Forecast System (GFS) models. All data is downloaded in GRIB format and automatically deleted from the users space. Data is downloaded using the Herbie-data package. Data can be processed in parallel using either standard Python multiprocessing or using dask (better for large historical runs).

The output format for the prediction files is:

/glade/scratch/username/ptype_output/{nwp_model}/{init_day}/{init_hour}/ptype_predictions_{nwp_model}{init_hour}z_fh{forecast_hour}.nc

The content of each file consists of the following (can be slightly modified in configuration file)

• Probability of precipitation for rain, snow, sleet, and freezing rain.
• ML categorical precipitation type (max probability)
• NWP categorical precipitation type (from model used for input)
• Orography
• 2m temperature and dewpoint, 10m winds (other variables can be added in configuration file)

Configuration file for inference: /config/inference.yml

• model
  • supports "hrrr", "gfs", "rap"
• ML_model_path
  • Path to ML model
• out_path
  • Base save path
• drop_input_data
  • Boolean (whether or not to save vertical profile data used for mdoel input)
  • Setting this to True greatly increase file size
• n_processors
  • Number of processors to use if using standard Python Multiprocessing (ignored if use_dask=True)
• use_dask
  • Boolean (True uses Dask for parallelizaion, False uses standard Multiprocessing)
• dates
  • Dictionary for starting and ending model initialization times (inclusive)
• forecast_range
  • Dictionary for forecast hours (inclusive) for each model initialzation time
• height_levels
  • Dictionary for model height levels needed for input into the ML model
• variables
  • Dictionary of variables and input needed to load data for each of the NWP models
  • All NWP models don't all have the same variables so different ones are needed (specifically ones to derive dewpoint)
  • "product" comes from the herbie API
  • Other variables can be added
• dask_params
  • Parameters to be used if use_dask=True
  • Specific for NCARs Casper Cluster