pps-mw-training

a tool for training of pps-mw algorithms

Quickstart

Installation

$ mamba env create -f pps_mw_training_environment.yml

$ conda activate pps-mw-training

$ pip install .

Run tests

The package contains a test suite can be run by tox:

$ pip install tox

$ tox

Training pipelines

The package currently contains two training "pipelines":

• iwp_ici - for the training of a multilayer perceptron quantile regression neural network for the retrieval of ice water path and related quantities from EPS-SG ICI data,
• pr_nordic - for the training of a U-Net convolutional quantile regression neural network for the retrieval of precipitation rate over the nordic region and from JPSS1/ATMS and N-SPP/ATMS data.

The two pipelines differ a bit in complexity. The iwp_ici pipeline is less complex and uses a fixed training dataset, and this package does not contain the tools to generate this training dataset. However, tools for the generation of a training dataset are included for the pr_nordic pipeline.

Run package scripts

The package contains three scripts:

• regrid - for regridding of data to produce a training dataset,
• reformat - for reformating of data files to produce a training dataset,
• train - for training of neural networks.

You only should run the train script for the iwp_ici pipeline, but you need to run all three type of scripts for the pr_nordic pipeline. Each script has a command line interface, and you can e.g. run the train script as described below:

$ train --help

usage: train [-h] {pr_nordic,iwp_ici} ...

Run the pps-mw-training app.

positional arguments:
  {pr_nordic,iwp_ici}
    pr_nordic          Run the Nordic precip training pipeline.
    iwp_ici            Run the IWP ICI training pipeline.

options:
  -h, --help           show this help message and exit

$ train iwp_ici --help

usage: train [-h] [-a ACTIVATION] [-b BATCH_SIZE] [-d DB_FILE] [-e EPOCHS] [-l N_HIDDEN_LAYERS] [-m MISSING_FRACTION] [-n N_NEURONS_PER_HIDDEN_LAYER]
               [-o] [-t TRAIN_FRACTION] [-u TEST_FRACTION] [-v VALIDATION_FRACTION] [-w MODEL_CONFIG_PATH]

Run the pps-mw training app for the training of a single quantile regression neural network, handling multiple quantiles and retrieval parameters, and
missing data, to retrieve ice water path and other associated parameters from ICI data.

options:
  -h, --help            show this help message and exit
  -a ACTIVATION, --activation ACTIVATION
                        Activation function to use for the hidden layers, default is relu
  -b BATCH_SIZE, --batchsize BATCH_SIZE
                        Training batch size, default is 4096
  -d DB_FILE, --db-file DB_FILE
                        Path to ICI retrieval database file to use as training data, default is /home/a002491/ici_retrieval_database.nc
  -e EPOCHS, --epochs EPOCHS
                        Number of training epochs, default is 256
  -l N_HIDDEN_LAYERS, --layers N_HIDDEN_LAYERS
                        Number of hidden layers, default is 4
  -m MISSING_FRACTION, --missing-fraction MISSING_FRACTION
                        Set this fraction of observations to a fill value, in order to allow for the network to learn to handle missing data, default is 0.1
  -n N_NEURONS_PER_HIDDEN_LAYER, --neurons N_NEURONS_PER_HIDDEN_LAYER
                        Number of hidden layers, default is 128
  -o, --only-evaluate   Flag for only evaluating a pretrained model
  -t TRAIN_FRACTION, --train-fraction TRAIN_FRACTION
                        Fraction of the training dataset to use as training data, default is 0.7
  -u TEST_FRACTION, --test-fraction TEST_FRACTION
                        Fraction of the training dataset to use as test data, default is 0.15
  -v VALIDATION_FRACTION, --validation-fraction VALIDATION_FRACTION
                        Fraction of the training dataset to use as validation data, default is 0.15
  -w MODEL_CONFIG_PATH, --write MODEL_CONFIG_PATH
                        Path to use for saving the trained model config, or to read from for an evaluation purpose, default is /home/a002491/work/pps-mw-
                        training/saved_model

Prior to running the actual training of the pr_nordic pipeline one must create the training dataset, i.e:

• regrid - regrid ATMS data onto positions of the grid of the BALTRAD data, note that this is a quite computational expensive operation,
• reformat - reformat BALTRAD data, saves the BALTRAD composites in a new file format only including the data needed for the training. This step, in practise, only reduces the file size of the BALTRAD composites. Currently, the train script only handles reformatted BALTRAD composites, so, at least for now, it is necessary to perform this processing.