GraphKAN: Graph Kolmogorov-Arnold Networks

https://github.com/willhua127/graphkan-graph-kolmogorov-arnold-networks

Admin

GraphKAN implements a Graph Neural Network version of Kolmogorov Arnold Networks (KAN) for...

Tokens:311

Snippets:3

Trust Score:8.2

Update:1 year ago

Show doc for...

Tokens:

Raw

### Configuring Device for PyTorch Training (GPU/CPU)

Source: https://github.com/willhua127/graphkan-graph-kolmogorov-arnold-networks/blob/main/README.md

This snippet demonstrates how to configure the device (GPU or CPU) for PyTorch operations based on CUDA availability. It shows the default setup for GPU preference and an alternative for forcing CPU usage, which is useful for debugging or environments without CUDA.

```Python
args.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
```

```Python
args.device = torch.device('cpu')
```

--------------------------------

### Incorrect Direct KANLayer Initialization for Input Features

Source: https://github.com/willhua127/graphkan-graph-kolmogorov-arnold-networks/blob/main/README.md

This snippet illustrates an approach that was found to be ineffective: directly applying a 'KANLayer' to the raw input features. Experimental observations indicate that this method prevents the KAN from receiving any training signal, highlighting the necessity of an initial linear projection into a latent space.

```Python
self.lin_in = KANLayer(in_feat, hidden_feat, grid_feat, addbias=use_bias)
```

--------------------------------

### Initializing Linear Layer for Latent Feature Projection in PyTorch

Source: https://github.com/willhua127/graphkan-graph-kolmogorov-arnold-networks/blob/main/README.md

This snippet initializes a standard 'nn.Linear' layer in PyTorch, used to project input features ('in_feat') into a lower-dimensional latent space ('hidden_feat'). This projection is crucial for the GraphKAN model's performance, as it provides a suitable feature representation before applying KAN layers.

```Python
self.lin_in = nn.Linear(in_feat, hidden_feat, bias=use_bias)
```

Tokens:

Raw