Collection of various PDE related work.
Use quarto preview pde-refiner.ipynb to preview the Reveal.js slides in a browser window while making changes.
Notes on attempt to reproduce PDE-Refiner for Kuramoto-Sivashinsky. Was not able to reproduce > 90s rollout with > 0.8 correlation, but this is a start.
To generate the data download this repo and install the conda environment: https://github.com/brandstetter-johannes/LPSDA
Then run ./generate-ks.sh. This takes 3 days to run on a M2 Pro. Then make a copy of the training dataset for each GPU you will train on.
Download https://github.com/pdearena/pdearena, install the conda environement inside of docker/ then add this to training config for Kuramoto-Sivashinsky so we get checkpoints:
+ - class_path: pytorch_lightning.callbacks.ModelCheckpoint
+ init_args:
+ monitor: 'train/loss_mean'
+ save_top_k: 1
+ verbose: True
+ save_last: True
+ dirpath: ${trainer.default_root_dir}/ckpts
+ filename: "epoch_{epoch:03d}"
+ auto_insert_metric_name: False
+ save_on_train_epoch_end: True
Then can run ./train.sh this takes 3 hours on 8x A100s or 2 days on V100.
Can use evaluation.ipynb to save ground truth and inference tensors to disk then use poster_plots.ipynb to generate the comparison images and correlation plots.
For testing add this diff to pdearena to force using the train dataset for testing (they use 140 for training and the rest of the trajectory for test).
with h5py.File(path, "r") as f:
- data_h5 = f[self.mode]
+ data_h5 = f["train"] # f[self.mode]
data_key = [k for k in data_h5.keys() if k.startswith("pde_")][0]
data = {
"u": torch.tensor(data_h5[data_key][:].astype(self.dtype)),
@@ -140,7 +140,7 @@ def _valid_filter(fname):
def _test_filter(fname):
- return "_test_" in fname and "h5" in fname
+ return "_train_" in fname and "h5" in fname