Please look at Project report.pdf for details about the project.
The repo consists of two parts. PDEBench (~1st month work) and PDEBench + L2L ( ~2nd month work).
To setup the environment run
bash create_env.sh
Below is a description of each folder added by me inside of pdebench/models. You can find here the original present directories.
Here you can find the .sh files which you can run using the sbatch command in commandline It contains all the scripts necessary to run different experiments. TODO: Figure out how to clean up this folder, there is a lot of redundant code.
Contains all the error files. An error file should be empty if the code ran correctly. You get an error file when you run a .sh file inside run_training.
Contains all the outputs files. The output files are the results of the experiments. You get an output file when you run a .sh file inside run_training.
Contains the pretrained models which can be found here. Please look at the README inside of ./pretrained_models/ for more instructions on how to run them.
Here I manually (by copy-pasting) placed all newly trained NNet models. Once training is finished you can find the newly trained model inside of e.g. ../../data/1D/Advection/Train/. The NN model should have a .pt and .pickle file.
Clone the repository. Then in command line run
cd pdebench-daad
bash create_couple_env.sh
Next we download a dataset. For example, lets download the Advection dataset. For more datasets please follow the instructions given here
# Activate the environment
. ../couple_env/bin/activate
# Download the Advection dataset
cd pdebench/data_download/
python download_direct.py --root_folder $proj_home/data --pde_name advection
Once the download finished, please check if the downloaded dataset(e.g. a file called 1D_Advection_Sols_beta4.0.hdf5) is inside a path like:
/pdebench-daad/pdebench/data/1D/Advection/Train/
If not, then please put 1D_Advection_Sols_beta4.0.hdf5 in such a path. This is also the path where later the trained model weights will be saved to.
To run training go to pdebench-daad/L2L/bin and run
sbatch launch_l2l_pdebench_Unet_es.sh
The errors are saved in pdebench-daad/L2L/bin/errors and outputs in pdebench-daad/L2L/bin/outputs
To plot the results with only one hyperparameter training go into data_processing.py. Here we need to change the folderPath valiable. To do this, open the generated outputs file in pdebench-daad/L2L/bin/outputs and somewhere from line 40 onward you should copy a path that ends with /simulation/work and set that as folderPath.
Many thanks to my mentor, Marcel Aach, without whom this project wouldn't have happened.
In development...