Documentation Discrepancy: Incompatibility of DIG with PyTorch Geometric 2.3.0

Hello,

I noticed an issue regarding the doc of DIG.

The doc states that DIG requires PyTorch Geometric (>=2.0.0). However, it seems that the DIG library is incompatible with PyTorch Geometric version 2.3.0.

In the 2.3.0 release of torch geometric, the API for the MessagePassing class was changed, as detailed in this commit. For example, the method __check_input__ was changed to _check_input`, which leads to incompatibility with certain lines in the DIG codebase, for instance: https://github.com/divelab/DIG/blob/97a726395f3ad1ed651658dd4d492a33e15929cd/dig/xgraph/models/models.py#L361-L362

https://github.com/divelab/DIG/blob/97a726395f3ad1ed651658dd4d492a33e15929cd/dig/xgraph/models/models.py#L367-L368 Also, the Dataset class was turned into an abstract class, requiring the implementation of two abstract methods len and get.

    @abstractmethod
    def len(self) -> int:
        r"""Returns the number of graphs stored in the dataset."""
        raise NotImplementedError

    @abstractmethod
    def get(self, idx: int) -> BaseData:
        r"""Gets the data object at index :obj:`idx`."""
        raise NotImplementedError

Consequently, classes such as MarginalSubgraphDataset need to implement these methods to function correctly: https://github.com/divelab/DIG/blob/97a726395f3ad1ed651658dd4d492a33e15929cd/dig/xgraph/method/shapley.py#L44-L64

This discrepancy might cause confusion for users attempting to install and use DIG with PyTorch Geometric 2.3.0.

divelab / DIG

Documentation Discrepancy: Incompatibility of DIG with PyTorch Geometric 2.3.0 #219