PyTorch Implementation of "ROMER: Reduced Order Modeling of Fluid Flows with Transformers" (paper)
Set up conda environment and clone the github repo:
# create a new environment
$ conda create --name ROMER python=3.7
$ conda activate ROMER
# install requirements
$ pip install -r requirements.txt
# clone the source code of ROMER
$ git clone https://github.com/BaratiLab/ROMER.git
$ cd ROMER
$ mkdir data
$ mkdir Results
The datasets are available here. After download, unzip the data in the data folder.
Train and test the models from the paper:
chmod +x ./training.sh
./training.sh
Train and test the models in the hyperprameter study:
chmod +x ./parameter_search.sh
./parameter_search.sh
Train a model with your choice of hyperparameters:
python Train.py -data data -down down -embed_dim embed_dim -dt dt -num_layers num_layers -num_heads num_heads -batch batch -ae_epoch ae_epochs -epochs epochs
where
data is the dataset number (1, 2 or 3) (default=3)
down is the number of encoder/decoder blocks in the autoencoder (default=4)
embed_dim is the feature dimension of the embedding tensor (default=128)
dt is the number of time steps that the transformer learns to propagate the system (default=1)
num_layers is the number of layers of the transformer (default=6)
num_heads is the number of heads in the multi-head attention in the transformer (default=8)
batch is the batch size used for training (default=64)
ae_epoch is the number of epochs to train the autoencoder (default=100)
epoch is the number of epochs to train the transformer (default=100)
You can find the recommended number of epochs for each dataset in training.sh.