Deeprecip is a recurrent neuronal network with an architecture tailured to produces short term precipitation prediction. Images from weather radars serve as inputs.

The network is implement in Chainer, a deep learning framework based on the “Define-by-Run” scheme.

This is work in progress. See this slides for more information.

1. Install required system packages (instructions for Ubuntu)

$ apt install gcc-4.8
$ apt install python3-dev
$ apt install libhdf5-dev
$ apt install libfreetype6-dev
$ apt install graphviz
$ apt install python3-mpltoolkits.basemap

Note, currently a pip install of basemap seems not possible.

2. Install CUDA and CuNN. Make sure that you run gcc and g++ on version 4.8! Use update-alternatives --config to switch versions.

3. Clone package

git clone https://gitlab.com/scheidan/deeprecip.git

3. Create a virtual environment and activate it

pyvenv venv
source venv/bin/activate

Note pyenv requries python 3.

4. Download the required python packages

(venv) $ pip install -e .

The -e flag means "editable" and creates links to the python files of the package (instead of copying them). If you update the package, the new files become available immediately. (Thanks Uwe!)

See run.py for an example.