MatrixChunk

MatrixChunk()

A block of the matrix, as compressed numpy arrays: consecutive rows (axis == "obs", CSR, every column) or consecutive genes (axis == "var", CSC, every row). The block is rows row_offset:row_offset + nrows by columns var_offset:var_offset + ncols.

The arrays take ownership of the buffers the reader decoded into, so there is no per-chunk copy. They are writable and outlive the chunk.

Attributes

Name Description
axis "obs" for a block of rows (CSR), "var" for a block of genes (CSC).
data (nnz,) values.
dtype NumPy dtype name of data, e.g. "float32".
indices (nnz,) uint32 column indices ("obs") or row indices ("var").
indptr uint64 pointers: (nrows + 1,) per row for "obs",
n_vars Number of columns (features) of the full matrix.
ncols Number of columns in this chunk.
nrows Number of rows in this chunk.
row_offset Index of the first row of this chunk in the full matrix.
var_offset Index of the first column of this chunk in the full matrix.

Methods

Name Description
to_scipy The chunk as a scipy.sparse.csr_array ("obs") or

to_scipy

MatrixChunk.to_scipy()

The chunk as a scipy.sparse.csr_array ("obs") or csc_array ("var") of shape (nrows, ncols) that shares data and indices with this chunk. scipy wants signed indices of one dtype: indices is viewed as int32 (same bits, never negative) and indptr is cast to int32, a copy of only its few entries. Given an int64 indptr, scipy would upcast, and so copy, indices.