in ./data/, prepare the dataloaders of the datasets.

in ./scripts/datasets.py, prepare the argument parser for the dataset.

Currently supported datasets: Burgers, Darcy, TorusLi, TorusVisForce

in ./models, prepare the models.

in ./scripts/models.py, prepare the argument parser for the model.

Currently supported models: FNO, T-FNO, LSM

in ./scripts/, you can define different settings of training, (e.g. train.py, train_multitask.py). for each setting, you may need some different arguments. set them in ./scripts/get_parser.py 's add_base_args() function. In addition, you may customize some LightningModules and Callbacks in ./lightning_modules/.

The process of each training scripts is:

First, in the training script, you should intialize a list of dataset parsers and model parsers that fits the kind of pipeline, say, $n$ datasets and $m$ models. Then the fetcher would create a parser of $ n\cdot m $ subparsers, correspondent to all the possible choices.

Don't forget to pass the type of task to the Fetcher().

Then one may customize his own training script concisely.

for linux environments, prepare your scripts in ./shell_scripts/.

To extract logs, run:

tb-reducer runs/YOUR_FOLDER -o OUTPUT_PATH/BLABLABLA.csv --handle-dup-steps mean