nicolasfauchereau / DL4SEAS

Repository for experiments using Deep Learning for Seasonal Climate Forecasting over New Zealand

Deep Learning experiments for seasonal climate forecasting

content of the repository

• dl4seas: modules

  • IO: inputs / outputs, for now functions to create TF protobuf files from xarray datasets and a list of netcdf files
  • NN: custom layers to use in CNNs and DataGenerator Class
  • preprocessing: for now only a train_scaler function using dask-ml or scikit-learn to train a Standard Scaler on xarray Dataset objects
  • utils: some utility functions

• CNN

notebooks, scripts, outputs and figures for the core experiments building on Convolutional Neural Networks (incl. Auto-encoders)

• transfer_learning

notebooks, scripts, outputs and figures for the experiments specifically building on pre-trained NN architectures (Transfer Learning)

• wide_and_deep

notebooks, scripts, outputs and figures for the experiments building on "wide and Deep"* architectures, which include auxiliary inputs representing the current state and / or the recent evolution of the climate system from observational sources.

• Cheng, H.-T., and Coauthors, 2016: Wide & Deep Learning for Recommender Systems. arXiv:1606.07792 [cs, stat]s