A framework to learn Shallow Water Equations with Deep Neural Networks.

Introduction

This framework aims to provide tools for experiments on physics-informed deep neural networks. It stems from the research conducted at the University of Parma from the group of prof. Alessandro Dal Palú, in collaboration with the HyLab research team.

• Wiki (in Italian)
• Experiments
• Models

Running experiments

Two scripts are available to either train or test the model:

• train.py: loads the dataset, initiates a model with random weights and performs training. The outputs are plotted for each epoch and saved in the session folder under runs/ (train_XXX)
• test.py: loads the dataset, loads the weights from a trained network, generates new samples from a test dataset. The results are plotted and saved ina the session folder under runs/ (eval_XXX)

Training

You can use python train.py -h to check out the parameters. Load a dataset from disk (PARFLOOD format), filter dynamic sequences and train deepSWE@32 filters:

python train.py -root /path -epochs 200 -future_frames 4 -filters 32 -dynamicity 0.5

Testing

You can use python test.py -h to check out the parameters. Load a dataset from disk (PARFLOOD format), filter dynamic sequences, train deepSWE@32 filters and apply 5cm of approximation for the accuracy:

python test.py -root /path -weights model.weights -filters 32 -dynamicity 0.5 -accuracy_threshold 0.05

Please note: the network hyperparameters (filters, lr) in the test script, must be compatible with the ones used during training.

Credits

• ndrplz/ConvLSTM_pytorch
• Po-Hsun-Su/pytorch-ssim
• Andreas Holm Nielsen