Sparse node features

[Rajjaa]

Hello,

I would like to know if there is a way to develop GNN models that accept sparse node features?

Thanks

[rusty1s]

Some GNNs support sparse node feature matrices directly as input such as GCNConv, but this is sadly not possible for all. For others, it might be a good idea to first project the sparse features into a low-dimensional (but dense) embedding space, and use that as input into the GNN:

lin = Linear(num_features, 32)
x = lin(sparse_x)

[Rajjaa]

Thank you for your reply but the two scenarios you mentioned throw a TypeError when using sparse node features from the NELL dataset:
TypeError: linear(): argument 'input' (position 1) must be Tensor, not SparseTensor
It seems that the Linear layer doesn't accept sparse tensors.

[rusty1s]

You will need to use the torch.sparse_coo_tensor from PyTorch for this, sorry about the confusion:

lin(sparse_x.to_torch_sparse_coo_tensor())

[Rajjaa]

Thank you it is working now. So why there is such a difference between sparse tensors from PyTorch and sparse tensors from torch-sparse?