scipy.sparse.

safely_cast_index_arrays

scipy.sparse.safely_cast_index_arrays(A, idx_dtype=<class 'numpy.int32'>, msg='')

Safely cast sparse array indices to idx_dtype.

Check the shape of A to determine if it is safe to cast its index arrays to dtype idx_dtype. If any dimension in shape is larger than fits in the dtype, casting is unsafe so raise ValueError. If safe, cast the index arrays to idx_dtype and return the result without changing the input A. The caller can assign results to A attributes if desired or use the recast index arrays directly.

Unless downcasting is needed, the original index arrays are returned. You can test e.g. A.indptr is new_indptr to see if downcasting occurred.

Added in version 1.15.0.

Parameters:

Asparse array or matrix

The array for which index arrays should be downcast.

idx_dtypedtype

Desired dtype. Should be an integer dtype (default: np.int32). Most of scipy.sparse uses either int64 or int32.

msgstring, optional

A string to be added to the end of the ValueError message if the array shape is too big to fit in idx_dtype. The error message is f"<index> values too large for {msg}" It should indicate why the downcasting is needed, e.g. “SuperLU”, and defaults to f”dtype {idx_dtype}”.

Returns:

idx_arraysndarray or tuple of ndarrays

Based on A.format, index arrays are returned after casting to idx_dtype. For CSC/CSR, returns (indices, indptr). For COO, returns coords. For DIA, returns offsets. For BSR, returns (indices, indptr).

Raises:

ValueError

If the array has shape that would not fit in the new dtype, or if the sparse format does not use index arrays.

Examples

Try it in your browser!

>>> import numpy as np
>>> from scipy import sparse
>>> data = [3]
>>> coords = (np.array([3]), np.array([1]))  # Note: int64 arrays
>>> A = sparse.coo_array((data, coords))
>>> A.coords[0].dtype
dtype('int64')

>>> # rescast after construction, raising exception if shape too big
>>> coords = sparse.safely_cast_index_arrays(A, np.int32)
>>> A.coords[0] is coords[0]  # False if casting is needed
False
>>> A.coords = coords  # set the index dtype of A
>>> A.coords[0].dtype
dtype('int32')

Go BackOpen In Tab