HFM_PyTorch

An alternate version of HFM code with PyTorch. Right now only utilities.py and Cylinder2D_flower_systematic.py are presented.

Original Repository is Hidden Fluid Mechanics.
Original Paper
Original Data and Figures

Additional files

DataManager.py: to save training loss and errors during training.
test.py: to test the results.
plot.py: to plot the results.

Notes:

The results are in Results folder. Comparing to the original code, this error rate of this version is a bit higher.
However, further training can be done by using a smaller learning rate to achieve better error rate.
A learning rate of 1e-3 (used in the original code) is prone to overfitting and the results are bad.
A learning rate of 1e-4 (showed in results)can have similar results comparing to the original paper.

To train:

download data and place it into a Data folder.

python Cylinder2D_flower_systematic.py 201 15000 [cuda-device-num|optional] [using visdom|optional]

[cuda-device-num]: don't need if not using GPU.
[using visdom] : this was not tested, but was planned to use visdom to visualize training process.

To test:

You should have related files in Results folder.
python test.py v10 [cuda-device-num|optional]

To plot: