nn.GCNConv adjacency matrix: SparseTensor vs COO format tensor

[dxyzx0]

I just find that in GCNConv, I can use the COO format to represent the adjacency matrix with edge_index and edge_weight or I can do it with sparse tensor edge_index and edge_weight=None. I run into RuntimeError: CUDA out of memory even with batch_size=1 since my graph has up to 300000 nodes. I wonder whether the latter input will consume less cuda memory than the former one?

[rusty1s]

Yes, the latter one will consume less memory on GCNConv. You can reduce the amount of memory further by performing adjacency normalization in a pre-processing step (e.g., via torch_geometric.transforms.GCNNorm), and initialize your GCN layers with normalize=False.

[dxyzx0]

But I must use torch_sparse.SparseTensor and CANNOT use torch.sparse_coo_tensor right?
However, If I use torch_sparse.SparseTensor, code won't be able to use torch.cuda.amp since SparseTensor won't be autocasted.

[rusty1s]

Let's move the discussion to here: #3036