Python library to train neural networks with a strong focus on hydrological applications.

This package has been used extensively in research over the last year and was used in various academic publications. The core idea of this package is modularity in all places to allow easy integration of new datasets, new model architectures or any training related aspects (e.g. loss functions, optimizer, regularization). One of the core concepts of this code base are configuration files, which lets anyone train neural networks without touching the code itself. The NeuralHydrology package is build on top of the deep learning framework Pytorch, since it has proven to be the most flexible and useful for research purposes.

We (AI for Earth Science group at Institute for Machine Learning, Johannes Kepler University, Linz, Austria) are using this code in our day-to-day research and will continue to integrate our new research findings into this public repository.

Note: We will gradually add more examples/documentation over the next couple of days/weeks.

• Documentation: neuralhydrology.readthedocs.io
• Research Blog: neuralhydrology.github.io
• Bug reports/Feature requests https://github.com/neuralhydrology/neuralhydrology/issues

We recommend to use Anaconda/Miniconda. With one of the two installed, a dedicated environment with all requirements installed can be set up from the environment files provided in environments.

If you have no CUDA capable GPU available run

conda env create -f environments/environment_cpu.yml

With a CUDA capable GPU available, check which CUDA version your GPU supports and then run e.g. (for CUDA 10.2)

conda env create -f environments/environment_cuda10_2.yml

If neither Minicoda/Anaconda are available, make sure to Python environment with all packages installed that are listed in one of the environment files.

For now download or clone the repository to your local machine and install a local, editable copy. This is a good idea if you want to edit the neuralhydrology code (e.g., adding new models or datasets).::

    git clone https://github.com/neuralhydrology/neuralhydrology.git
    cd neuralhydrology
    pip install -e .

Besides adding the package to your Python environment, it will also add three bash scripts: nh-run, nh-run-scheduler and nh-results-ensemble. For details, see below.

Training and evaluating models requires a dataset. If you're unsure where to start, a common dataset is CAMELS US, available at CAMELS US (NCAR). Download the "CAMELS time series meteorology, observed flow, meta data" and place the actual data folder (basin_dataset_public_v1p2) in a directory. This directory will be referred to as the "data directory", or data_dir.

One of the core concepts of this package is the usage of configuration files (.yml). Basically, all configurations required to train a neural network can be specified via these configuration files and no code has to be touched. Training a model does require a .yml file that specifies the run configuration. We will add a detailed explanation for within the next weeks that explains the config files and arguments in more detail. For now refer to the example config for a full list of all available arguments (with inline documentation). For an example of a configuration file that can be used to train a standard LSTM for a single CAMELS US basin, check 1_basin_config.yml.

To train a model, prepare a configuration file, then run::

    nh-run train --config-file /path/to/config.yml

If you want to train multiple models, you can make use of the nh-run-scheduler command. Place all configs in a folder, then run::

    nh-run-scheduler train --directory /path/to/config_dir/ --runs-per-gpu X --gpu-ids Y

With X, you can specify how many models should be trained on parallel on a single GPU. With Y, you can specify which GPUs to use for training (use the id as specified in nvidia-smi).

To evaluate a trained model on the test set, run::

nh-run evaluate --run-dir /path/to/run_dir/

If the optional argument --epoch N (where N is the epoch to evaluate) is not specified, the weights of the last epoch are used. You can also use --period if you want to evaluate the model on the train period --period train) or validation period (--period validation)

To evaluate all runs in a specific directory you can, similarly to training, run::

nh-run-scheduler evaluate --directory /path/to/config_dir/ --runs-per-gpu X --gpu-ids Y

To merge the predictons of a number of runs (stored in $DIR1, ...) into one averaged ensemble, use the nh-results-ensemble script::

nh-results-ensemble --run-dirs $DIR1 $DIR2 ... --save-file /path/to/target/file.p --metrics NSE MSE ...

--metrics specifies which metrics will be calculated for the averaged predictions.

If you have any questions regarding the usage of this repository, feature requests or comments, please open an issue. You can also reach out to Frederik Kratzert (kratzert(at)ml.jku.at) by email.