Preliminary steps to train a ConvLSTM model of ParFlow

Parflow-NN works on Python 3. Creating an Anaconda environment with Python 3.8 is the easiest option, especially if you want GPU support.

Install anaconda or miniconda for your platform and then use the supplied environment.yml to create and activate the environment. Install the parflow-nn package inside this environment.

conda env create -f environment.yml
conda activate parflow_nn
pip install -e .

We recommend installing the parflow-nn package in develop mode (the -e flag in pip install -e .) while this package is actively under development, so that you can tweak your local copy of the code easily if you need to, and observe the changes.

If you wish to leverage an available GPU, install the tensorflow-gpu package.

conda install tensorflow-gpu

If you prefer to use pip in a venv, create the virtual environment, activate it, and then install the package.

python3 -m venv env
source env/bin/activate
pip install -e .

We recommend installing the parflow-nn package in develop mode (the -e flag in pip install -e .) while this package is actively under development, so that you can tweak your local copy of the code easily if you need to, and observe the changes.

If you wish to leverage an available GPU, first install and configure the CUDA Toolkit for your platform. You're most likely to succeed if you use the 10.1 version of the toolkit, as this is the version that Tensorflow on PyPI seems to work best with.

Then install the tensorflow-gpu package through pip.

pip install tensorflow-gpu

After activating your Conda environment or venv, run the command

python -c "import tensorflow as tf; print(tf.test.is_gpu_available())"

and look at the last line of the output. If you see a True, you're all set!

cd washita/tcl_scripts
./model.tcl <run_name>

This folder generates the folder runs/<run_name> with the simulation results.

Execute the command:

python -m parflow_nn nn <run_dir>

This step requires you to point to the run_directory, i.e. the runs/<run_name> folder generated in the simulation step above. If not already present, an intermediate folder called nc_files is created inside the runs/<run_name> folder with the model .nc files.

• Hoang Tran ([email protected]; [email protected])