Producing accurate hydrological distributional predictions using Bayesian long-short term memory networks

What is this?

The repository contains the software and model code for Producing accurate hydrological distributional predictions using Bayesian long-short term memory networks. It features all relevant source code for running the Bayes-by-Backprop-algorithm with the CAMELS-US-dataset as an application. It is possible to train and evaluate models following all three architecture presented in the paper (many-to-one, many-to-one with discharge and many-to-many). Moreover, it is possible to tweak the proposed hyperparameters in order to experiment with different architectures.

How to use the model code

The source code is written for Python 3.8, therefore we recommend setting up a virtual environment with this version of Python. The source code depends on a couple of requirements which are listed in the requirements.txt-file.

A file names camels_run.py in der folder camels_us contains the necessary source code to train/evaluate a model. The file contains a number of changable variables which are described in the following:

• CAMELS_ROOT: points to a specific subfolder of the CAMELS-US-dataset that is used for this project. Should not be changed unless you wish to use an own version of the dataset.
• DEVICE: the device used for the computations. Should normally not be changed.
• PLOT_FREQUENCY: the frequency to evaluate the model, plot the results of the evaluation and save an instance of the trained model. For example: 10 means that the model is evaluated every 10 epochs.
• SAVE_MODEL: should be set to True if instances of the trained model should be saved. False otherwise.
• ONLY_EVAL: should be set to True if training has already been done and one only wishes to evaluate the model. Note that in this case a valid path for loading the model has to be given as described below.
• SAVE_PATH: path where instances of the model should be saved (if the option is enabled).
• LOAD_PATH: path of the model instance that should be loaded in case ONLY_EVAL is enabled.

Lines 70-74 can be modified to change the used architecture or hyperparameters.

Note that it is not necessary to download separate datasets as the repository already contains all necessary data from CAMELS-US.