AromeDownscaling-unet-keras

This project is a part of an internship at Meteo France, which aims at downscaling meteorological fields obtained with the Arome model. It is fully usable as long as it is used on the data it has been created to work with.

It implements a Unet, and is isnpired by https://github.com/antoinedoury/RCM-Emulator. A second project implements a DDPM specifically for wind downscaling: https://github.com/lorisdanjou/AromeDownscaling-ddpm-pytorch

Prerequisites

Some packages (and specific versions of these packages) need to be installed to run the codes:

pip install -r requirements.txt

Structure of the project

The Data directory contains all the functions needed to load and preprocess data, unet contains all the necessary functions to define a Unet, a UResnet or a ResUnet, training contains all the functions needed to train the models (loss functions, etc.) and the results directory provides the necessary stuff to print results. Other useful functions are contained in core and utils.

All these functions are called by high-level scripts, defined to do a specific task.

Define and train a model

To define and train a model, you should use the sr.py script. Like all other scripts in this project, it is not necessary to change the code to change a hyperparameter. All the options are specified in a configuration .json file. For example, you can call:

python3 sr.py -c config/sr_example.jsonc

This script can also be used to load a pretrained model and resume training, with the weights option in the configuration file.

Import a trained model and make an inference

The infer.py file is made for that:

python3 infer.py -c config/sr_example.jsonc

Plot the results of an experiment

You can use the plot_results.py script to do that:

python3 plot_results.py -c config/sr_example.jsonc

Print the results of several experiments to compare them

You can use the plot_synthesis.py script for that task:

python3 plot_synthesis.py -c config/synthesis_example.jsonc