Skip to content

Latest commit

 

History

History
55 lines (34 loc) · 1.91 KB

README.md

File metadata and controls

55 lines (34 loc) · 1.91 KB
Raw

1D Advection

This directory contains the code to train and evaluate models on the 1D advection dataset. Make sure to install the requirements by running pip install -r requirements.txt from the root folder of this repository.

Data

The data can be obtained from here. Make sure to update the path of data_dir (line 40) in loaders.py to point to the directory containing the data.

Training

To train the model on the advection_1d dataset, run the following command:

python modelname_train.py # where modelname is one of the following: deeponet, nomad, fno, cvit

This will train the model and save the best model to ./model_states/modelname.pickle.

Evaluation

To evaluate the model on the advection_1d dataset, run the following command:

python modelname_test.py

This will load the best model from ./model_states/modelname.pickle and evaluate it on the test set, and output the results to ./model_outputs/modelname.h5.

Results

In the previous step, the results are saved to ./model_outputs/modelname.h5. To visualize the results, run the jupyter notebook plot.ipynb which will load the results and plot them. We state the obtained error metrics below.

Using Relative L2 Error

Model Mean Median Max
NOMAD 0.032787 0.0246571 0.1741
DeepONet 0.0574771 0.0454819 0.232457
FNO 0.0408418 0.0310558 0.191958
CViT 0.0335431 0.0275473 0.152318

Using TV

Model Mean Median Max
NOMAD 0.032787 0.0246571 0.1741
DeepONet 0.0574771 0.0454819 0.232457
FNO 0.0408418 0.0310558 0.191958
CViT 0.0335431 0.0275473 0.152318

Visualizations

adv