🆕 dgl.sparse¶

dgl.sparse is a library for sparse operators that are commonly used in GNN models.

Sparse matrix class¶

class dgl.sparse.SparseMatrix[source]¶

A SparseMatrix can be created from Coordinate format indices using the spmatrix() constructor:

>>> indices = torch.tensor([[1, 1, 2],
>>>                         [2, 4, 3]])
>>> A = dglsp.spmatrix(indices)
SparseMatrix(indices=tensor([[1, 1, 2],
                             [2, 4, 3]]),
             values=tensor([1., 1., 1.]),
             shape=(3, 5), nnz=3)

Creation Ops¶

`spmatrix`(indices[, val, shape])	Creates a sparse matrix from Coordinate format indices.
`val_like`(mat, val)	Creates a sparse matrix from an existing sparse matrix using new values.
`from_coo`(row, col[, val, shape])	Creates a sparse matrix from a coordinate list (COO), which stores a list of (row, column, value) tuples.
`from_csr`(indptr, indices[, val, shape])	Creates a sparse matrix from compress sparse row (CSR) format.
`from_csc`(indptr, indices[, val, shape])	Creates a sparse matrix from compress sparse column (CSC) format.
`diag`(val[, shape])	Creates a sparse matrix based on the diagonal values.
`identity`(shape[, d, dtype, device])	Creates a sparse matrix with ones on the diagonal and zeros elsewhere.

Attributes and methods¶

`SparseMatrix.shape`	Returns the shape of the sparse matrix.
`SparseMatrix.nnz`	Returns the number of non-zero elements in the sparse matrix.
`SparseMatrix.dtype`	Returns the data type of the sparse matrix.
`SparseMatrix.device`	Returns the device the sparse matrix is on.
`SparseMatrix.val`	Returns the values of the non-zero elements.
`SparseMatrix.row`	Returns the row indices of the non-zero elements.
`SparseMatrix.col`	Returns the column indices of the non-zero elements.
`SparseMatrix.indices`()	Returns the coordinate list (COO) representation in one tensor with shape `(2, nnz)`.
`SparseMatrix.coo`()	Returns the coordinate list (COO) representation of the sparse matrix.
`SparseMatrix.csr`()	Returns the compressed sparse row (CSR) representation of the sparse matrix.
`SparseMatrix.csc`()	Returns the compressed sparse column (CSC) representation of the sparse matrix.
`SparseMatrix.coalesce`()	Returns a coalesced sparse matrix.
`SparseMatrix.has_duplicate`()	Returns `True` if the sparse matrix contains duplicate indices.
`SparseMatrix.to_dense`()	Returns a copy in dense matrix format of the sparse matrix.
`SparseMatrix.to`([device, dtype])	Performs matrix dtype and/or device conversion.
`SparseMatrix.cuda`()	Moves the matrix to GPU.
`SparseMatrix.cpu`()	Moves the matrix to CPU.
`SparseMatrix.float`()	Converts the matrix values to float32 data type.
`SparseMatrix.double`()	Converts the matrix values to double data type.
`SparseMatrix.int`()	Converts the matrix values to int32 data type.
`SparseMatrix.long`()	Converts the matrix values to long data type.
`SparseMatrix.transpose`()	Returns the transpose of this sparse matrix.
`SparseMatrix.t`()	Alias of `transpose()`
`SparseMatrix.T`	Alias of `transpose()`
`SparseMatrix.neg`()	Returns a new sparse matrix with the negation of the original nonzero values, equivalent to `-A`.
`SparseMatrix.reduce`([dim, rtype])	Computes the reduction of non-zero values of the `input` sparse matrix along the given dimension `dim`.
`SparseMatrix.sum`([dim])	Computes the sum of non-zero values of the `input` sparse matrix along the given dimension `dim`.
`SparseMatrix.smax`([dim])	Computes the maximum of non-zero values of the `input` sparse matrix along the given dimension `dim`.
`SparseMatrix.smin`([dim])	Computes the minimum of non-zero values of the `input` sparse matrix along the given dimension `dim`.
`SparseMatrix.smean`([dim])	Computes the mean of non-zero values of the `input` sparse matrix along the given dimension `dim`.
`SparseMatrix.softmax`([dim])	Applies softmax to the non-zero elements of the sparse matrix on the dimension :attr:`dim`.

Operators¶

Elementwise Operators¶

`add`(A, B)	Elementwise addition for `SparseMatrix`, equivalent to `A + B`.
`sub`(A, B)	Elementwise subtraction for `SparseMatrix`, equivalent to `A - B`.
`mul`(A, B)	Elementwise multiplication for `SparseMatrix`, equivalent to `A * B`.
`div`(A, B)	Elementwise division for `SparseMatrix`, equivalent to `A / B`.
`power`(A, scalar)	Elementwise exponentiation `SparseMatrix`, equivalent to `A ** scalar`.

Matrix Multiplication¶

`matmul`(A, B)	Multiplies two dense/sparse matrices, equivalent to `A @ B`.
`spmm`(A, X)	Multiplies a sparse matrix by a dense matrix, equivalent to `A @ X`.
`bspmm`(A, X)	Multiplies a sparse matrix by a dense matrix by batches, equivalent to `A @ X`.
`spspmm`(A, B)	Multiplies a sparse matrix by a sparse matrix, equivalent to `A @ B`.
`sddmm`(A, X1, X2)	Sampled-Dense-Dense Matrix Multiplication (SDDMM).
`bsddmm`(A, X1, X2)	Sampled-Dense-Dense Matrix Multiplication (SDDMM) by batches.

Non-linear activation functions¶

softmax(input[, dim])

Applies softmax to the non-zero elements of the sparse matrix on the dimension :attr:dim.

Broadcast operators¶

`sp_broadcast_v`(A, v, op)	Broadcast operator for sparse matrix and vector.
`sp_add_v`(A, v)	Broadcast addition for sparse matrix and vector.
`sp_sub_v`(A, v)	Broadcast substraction for sparse matrix and vector.
`sp_mul_v`(A, v)	Broadcast multiply for sparse matrix and vector.
`sp_div_v`(A, v)	Broadcast division for sparse matrix and vector.