This repository has all the necessary to start doing video prediction specially focus in radar weather data from the NEXRAD network.
- https://ieeexplore.ieee.org/document/9542945
- https://journals.ametsoc.org/view/journals/atot/aop/JTECH-D-21-0012.1/JTECH-D-21-0012.1.xml
In the "examples" directory there are plenty of jupyter notebook examples to start on that.
The standard dimensions used are:
- dataset: (samples, frames, height, width);
if developing you will encounter that the models use: (samples, channels, frames, height, width), where channels should be 1 at the input and the output and only increase/decrease within the model.
GPU is necessary for training models
Create a separate conde environment (optiona):
conda create -n mlnowcasting
source activate mlnowcasting
Install the minimum requisites
conda config --append channels conda-forge
conda install numpy
if GPU: conda install pytorch torchvision cudatoolkit=10.1 -c pytorch (with the proper cuda version)
if CPU: conda install pytorch-cpu torchvision-cpu -c pytorch
Install the package
pip install .
At this point you should be able to run prediction, like in "examples/How to make predictions"
The following packages will allow you to use every functionality, if you don't want to go for all read above what each package is used for.
pip install nexradaws
conda install cartopy arm_pyart IPython pysteps hyperopt sh pillow imageio opencv scikit-learn jupyter
if GPU: conda install tensorflow-gpu keras
if CPU: conda install tensorflow keras