Put Sparse Matrix on CUDA

[RX28666]

Hi, I am trying to compute the normalized adjacency matrix D(-1/2)(A+I)D(-1/2) on Cuda, this is what I am doing:

adj = torch_geometric.utils.to_scipy_sparse_matrix(data.edge_index, num_nodes = data.num_nodes)
M_test = calc_A_hat(adj_test)
A_test = sparse_matrix_to_torch((1 - args.alpha) * M_test)
A_test = A_test.to(device)

Where:

def calc_A_hat(adj_matrix: sp.spmatrix) -> sp.spmatrix:
       nnodes = adj_matrix.shape[0]
       A = adj_matrix + sp.eye(nnodes)
       D_vec = np.sum(A, axis=1).A1
       D_vec_invsqrt_corr = 1 / np.sqrt(D_vec)
       D_invsqrt_corr = sp.diags(D_vec_invsqrt_corr)
       return D_invsqrt_corr @ A @ D_invsqrt_corr

and:

def sparse_matrix_to_torch(X):
    coo = X.tocoo()
    indices = np.array([coo.row, coo.col])
    return torch.sparse.FloatTensor(torch.LongTensor(indices), torch.FloatTensor(coo.data), coo.shape)

It seems all calculations are done on CPU, I tried to put adj to Cuda first by adj = adj.to(device), but it returned to not found in sparse matrix, I am wondering how I can put the sparse matrix on CUDA so I can do this calculations on GPU, or there is another way to get the normalized adjacency matrix on CUDA, because the calculations are very slow when the graph is large. Any help would be appreciated.

[EdisonLeeeee]

You were using scipy sparse matrix that is not supported on CUDA. It's suggested to use torch_sparse.SparseTensor instead, see

pytorch_geometric/torch_geometric/nn/conv/gcn_conv.py

Lines 32 to 34 in 32c6b7a

	def gcn_norm(edge_index, edge_weight=None, num_nodes=None, improved=False,
	add_self_loops=True, flow="source_to_target", dtype=None):