FamPredAI

Reservoir Computing for prediction of timeseries of food insecurity.

Installation process and dependencies

Running Reservoir Computing The current repo doesn't require any installation. It can be run directly after installing the required packages using:

python -m pip install -r requirements.txt

This assumes the version of python used is 3.9. Dependencies can be installed in a virtual/conda environment.

Furthermore, you will need to clone and install the rescomp package by running:

git clone https://github.com/GLSRC/rescomp.git
cd rescomp
pip install .

Complete instructions can be found in the docs for the package. Running Keras Based model (CNN, LSTM) The rescomp librabry is incompatible with keras because of the required version for nunmpy. To run the deep learning algorithms follow the instructions:

python -m pip install -r requirements_deep_learning.txt

Once again this assumes the version of python used is 3.9. Dependencies can be installed in a virtual/conda environment.

Data

Data to train the models can be found for each country in data/<country>/full_timeseries_daily.csv. The Model class is set up to read from these files when loading the training data. Therefore, shall the data need to be updated, we suggest keeping the same name and format of the current file.

Generating forecasts

Reservoir Computing To generate forecasts for a chosen country and model, just pass the following arguments to the function forecast in reservoir_computing.py:

• country: name of the chosen country;
• first_forecast: first day to be forecasted (NB: can be at max the day following last available day for target data);
• constants: list of discrete/categorical secondary data to be used to train the model;
• variables: list of continuous secondary data to be used to train the model;
• hyperparameters: dict of hyperparameters needed for the model;
• forecast_window: number of days to forecast for. CNN and LSTM Follow the steps in the forecast_from_file function found in cnn.py and lstm.py

A dictionary with example arguments for Haiti can be imported from the file parameters.py, to be passed to this function.

To generate forecasts for monthly test splits, using the best parameters found for the chosen model at each of these, run the function forecast_from_file in the reservoir_computing.py, cnn.py and lstm.py

• country: name of the chosen country;
• model: name of the model, i.e. RC, to be used to read from the correct file of parameters in folder best_hyperparameters.