This is the python implementation of the Graph Neural Networks studied in arXiv:2409.02126 based on PyTorch and PyTorch Geometric.
- PyTorch 2.3.1
- PyTorch Geometric 2.5.3
- NetworkX 3.3
- Numpy 1.26.4
- Matplotlib 3.9.0
The program is composed of three main pieces:
utilities.py,torch_classe.py,Main_File.py.
The utilities.py file contains functions aimed to generate a Random Plumbing graph and perform Neumann moves on it to then generate homeomorphic or non-homeomorphic graphs pairs.
In torch_classe.py all the Graph Neural Nework architectures are implemented together with the PyTorch dataclass we convert our graph to.
Finally, the Main_File.py script contains the heart of the program: a database of graph pairs is generated and multiple network structures are trained on it. The resulting comparative accuracy on the test and validation sets plus the loss function evolution over the epoch are plotted and stored respectively in img_accuracy.pdf and img_loss.pdf.
The code can be run on cuda devices but not on mps ones, since as of now (August 2024) the torch_scatter routine is not supported.