Quickstart

Step 1

The first step is to download the subsampled low-resolution real-geography version of the data here. This contains subsampled and prenormalized data that was used for training, validation, and metrics for the ClimSim paper. It can be reproduced with the full version of the dataset using the preprocessing/create_npy_data_splits.ipynb notebook.

System requirements can be found here. This was contributed by STAT GU4243/GR 5243 Applied Data Science - Fall 2023 at Columbia University.

Training data corresponds to train_input.npy and train_target.npy. Validation data corresponds to val_input.npy and val_target.npy. Scoring data (which can be treated as a test set) corresponds to scoring_input.npy and scoring_target.npy. We have an additional held-out test set that we are using for our Kaggle competition.

Step 2

Install the climsim_utils python tools, by running the following code from the root of the GitHub repo:

pip install .

If you already have all climsim_utils dependencies (tensorflow, xarray, etc.) installed in your local environment, you can alternatively run:

pip install . --no-deps

Step 3

Train your model on the training data and validate using the validation data. If you wish to use something like a CNN, you will probably want to separate the variables into channels and broadcast scalars into vectors of the same dimension as vertically-resolved variables. Methods to do this can be found in the climsim_utils/data_utils.py script.

Step 4

Evaluation time! Use the evaluation/main_figure_generation.ipynb notebook to see how your model does! Use the calc_MAE, calc_RMSE, and calc_R2 methods in the climsim_utils/data_utils.py script to see how your model does on point estimates and use the calc_CRPS method to check how well-calibrated your model is if it’s stochastic. 😊