2024-10-02 22:15:59 +04:00

339 lines
12 KiB
Python

"""Test of 1D arithmetic operations"""
import pytest
import numpy as np
from numpy.testing import assert_equal, assert_allclose
from scipy.sparse import coo_array, csr_array
from scipy.sparse._sputils import isscalarlike
spcreators = [coo_array, csr_array]
math_dtypes = [np.int64, np.float64, np.complex128]
def toarray(a):
if isinstance(a, np.ndarray) or isscalarlike(a):
return a
return a.toarray()
@pytest.fixture
def dat1d():
return np.array([3, 0, 1, 0], 'd')
@pytest.fixture
def datsp_math_dtypes(dat1d):
dat_dtypes = {dtype: dat1d.astype(dtype) for dtype in math_dtypes}
return {
sp: [(dtype, dat, sp(dat)) for dtype, dat in dat_dtypes.items()]
for sp in spcreators
}
@pytest.mark.parametrize("spcreator", spcreators)
class TestArithmetic1D:
def test_empty_arithmetic(self, spcreator):
shape = (5,)
for mytype in [
np.dtype('int32'),
np.dtype('float32'),
np.dtype('float64'),
np.dtype('complex64'),
np.dtype('complex128'),
]:
a = spcreator(shape, dtype=mytype)
b = a + a
c = 2 * a
assert isinstance(a @ a.tocsr(), np.ndarray)
assert isinstance(a @ a.tocoo(), np.ndarray)
for m in [a, b, c]:
assert m @ m == a.toarray() @ a.toarray()
assert m.dtype == mytype
assert toarray(m).dtype == mytype
def test_abs(self, spcreator):
A = np.array([-1, 0, 17, 0, -5, 0, 1, -4, 0, 0, 0, 0], 'd')
assert_equal(abs(A), abs(spcreator(A)).toarray())
def test_round(self, spcreator):
A = np.array([-1.35, 0.56, 17.25, -5.98], 'd')
Asp = spcreator(A)
assert_equal(np.around(A, decimals=1), round(Asp, ndigits=1).toarray())
def test_elementwise_power(self, spcreator):
A = np.array([-4, -3, -2, -1, 0, 1, 2, 3, 4], 'd')
Asp = spcreator(A)
assert_equal(np.power(A, 2), Asp.power(2).toarray())
# element-wise power function needs a scalar power
with pytest.raises(NotImplementedError, match='input is not scalar'):
spcreator(A).power(A)
def test_real(self, spcreator):
D = np.array([1 + 3j, 2 - 4j])
A = spcreator(D)
assert_equal(A.real.toarray(), D.real)
def test_imag(self, spcreator):
D = np.array([1 + 3j, 2 - 4j])
A = spcreator(D)
assert_equal(A.imag.toarray(), D.imag)
def test_mul_scalar(self, spcreator, datsp_math_dtypes):
for dtype, dat, datsp in datsp_math_dtypes[spcreator]:
assert_equal(dat * 2, (datsp * 2).toarray())
assert_equal(dat * 17.3, (datsp * 17.3).toarray())
def test_rmul_scalar(self, spcreator, datsp_math_dtypes):
for dtype, dat, datsp in datsp_math_dtypes[spcreator]:
assert_equal(2 * dat, (2 * datsp).toarray())
assert_equal(17.3 * dat, (17.3 * datsp).toarray())
def test_sub(self, spcreator, datsp_math_dtypes):
for dtype, dat, datsp in datsp_math_dtypes[spcreator]:
if dtype == np.dtype('bool'):
# boolean array subtraction deprecated in 1.9.0
continue
assert_equal((datsp - datsp).toarray(), np.zeros(4))
assert_equal((datsp - 0).toarray(), dat)
A = spcreator([1, -4, 0, 2], dtype='d')
assert_equal((datsp - A).toarray(), dat - A.toarray())
assert_equal((A - datsp).toarray(), A.toarray() - dat)
# test broadcasting
assert_equal(datsp.toarray() - dat[0], dat - dat[0])
def test_add0(self, spcreator, datsp_math_dtypes):
for dtype, dat, datsp in datsp_math_dtypes[spcreator]:
# Adding 0 to a sparse matrix
assert_equal((datsp + 0).toarray(), dat)
# use sum (which takes 0 as a starting value)
sumS = sum([k * datsp for k in range(1, 3)])
sumD = sum([k * dat for k in range(1, 3)])
assert_allclose(sumS.toarray(), sumD)
def test_elementwise_multiply(self, spcreator):
# real/real
A = np.array([4, 0, 9])
B = np.array([0, 7, -1])
Asp = spcreator(A)
Bsp = spcreator(B)
assert_allclose(Asp.multiply(Bsp).toarray(), A * B) # sparse/sparse
assert_allclose(Asp.multiply(B).toarray(), A * B) # sparse/dense
# complex/complex
C = np.array([1 - 2j, 0 + 5j, -1 + 0j])
D = np.array([5 + 2j, 7 - 3j, -2 + 1j])
Csp = spcreator(C)
Dsp = spcreator(D)
assert_allclose(Csp.multiply(Dsp).toarray(), C * D) # sparse/sparse
assert_allclose(Csp.multiply(D).toarray(), C * D) # sparse/dense
# real/complex
assert_allclose(Asp.multiply(Dsp).toarray(), A * D) # sparse/sparse
assert_allclose(Asp.multiply(D).toarray(), A * D) # sparse/dense
def test_elementwise_multiply_broadcast(self, spcreator):
A = np.array([4])
B = np.array([[-9]])
C = np.array([1, -1, 0])
D = np.array([[7, 9, -9]])
E = np.array([[3], [2], [1]])
F = np.array([[8, 6, 3], [-4, 3, 2], [6, 6, 6]])
G = [1, 2, 3]
H = np.ones((3, 4))
J = H.T
K = np.array([[0]])
L = np.array([[[1, 2], [0, 1]]])
# Some arrays can't be cast as spmatrices (A, C, L) so leave
# them out.
Asp = spcreator(A)
Csp = spcreator(C)
Gsp = spcreator(G)
# 2d arrays
Bsp = spcreator(B)
Dsp = spcreator(D)
Esp = spcreator(E)
Fsp = spcreator(F)
Hsp = spcreator(H)
Hspp = spcreator(H[0, None])
Jsp = spcreator(J)
Jspp = spcreator(J[:, 0, None])
Ksp = spcreator(K)
matrices = [A, B, C, D, E, F, G, H, J, K, L]
spmatrices = [Asp, Bsp, Csp, Dsp, Esp, Fsp, Gsp, Hsp, Hspp, Jsp, Jspp, Ksp]
sp1dmatrices = [Asp, Csp, Gsp]
# sparse/sparse
for i in sp1dmatrices:
for j in spmatrices:
try:
dense_mult = i.toarray() * j.toarray()
except ValueError:
with pytest.raises(ValueError, match='inconsistent shapes'):
i.multiply(j)
continue
sp_mult = i.multiply(j)
assert_allclose(sp_mult.toarray(), dense_mult)
# sparse/dense
for i in sp1dmatrices:
for j in matrices:
try:
dense_mult = i.toarray() * j
except TypeError:
continue
except ValueError:
matchme = 'broadcast together|inconsistent shapes'
with pytest.raises(ValueError, match=matchme):
i.multiply(j)
continue
sp_mult = i.multiply(j)
assert_allclose(toarray(sp_mult), dense_mult)
def test_elementwise_divide(self, spcreator, dat1d):
datsp = spcreator(dat1d)
expected = np.array([1, np.nan, 1, np.nan])
actual = datsp / datsp
# need assert_array_equal to handle nan values
np.testing.assert_array_equal(actual, expected)
denom = spcreator([1, 0, 0, 4], dtype='d')
expected = [3, np.nan, np.inf, 0]
np.testing.assert_array_equal(datsp / denom, expected)
# complex
A = np.array([1 - 2j, 0 + 5j, -1 + 0j])
B = np.array([5 + 2j, 7 - 3j, -2 + 1j])
Asp = spcreator(A)
Bsp = spcreator(B)
assert_allclose(Asp / Bsp, A / B)
# integer
A = np.array([1, 2, 3])
B = np.array([0, 1, 2])
Asp = spcreator(A)
Bsp = spcreator(B)
with np.errstate(divide='ignore'):
assert_equal(Asp / Bsp, A / B)
# mismatching sparsity patterns
A = np.array([0, 1])
B = np.array([1, 0])
Asp = spcreator(A)
Bsp = spcreator(B)
with np.errstate(divide='ignore', invalid='ignore'):
assert_equal(Asp / Bsp, A / B)
def test_pow(self, spcreator):
A = np.array([1, 0, 2, 0])
B = spcreator(A)
# unusual exponents
with pytest.raises(ValueError, match='negative integer powers'):
B**-1
with pytest.raises(NotImplementedError, match='zero power'):
B**0
for exponent in [1, 2, 3, 2.2]:
ret_sp = B**exponent
ret_np = A**exponent
assert_equal(ret_sp.toarray(), ret_np)
assert_equal(ret_sp.dtype, ret_np.dtype)
def test_dot_scalar(self, spcreator, dat1d):
A = spcreator(dat1d)
scalar = 10
actual = A.dot(scalar)
expected = A * scalar
assert_allclose(actual.toarray(), expected.toarray())
def test_matmul(self, spcreator):
Msp = spcreator([2, 0, 3.0])
B = spcreator(np.array([[0, 1], [1, 0], [0, 2]], 'd'))
col = np.array([[1, 2, 3]]).T
# check sparse @ dense 2d column
assert_allclose(Msp @ col, Msp.toarray() @ col)
# check sparse1d @ sparse2d, sparse1d @ dense2d, dense1d @ sparse2d
assert_allclose((Msp @ B).toarray(), (Msp @ B).toarray())
assert_allclose(Msp.toarray() @ B, (Msp @ B).toarray())
assert_allclose(Msp @ B.toarray(), (Msp @ B).toarray())
# check sparse1d @ dense1d, sparse1d @ sparse1d
V = np.array([0, 0, 1])
assert_allclose(Msp @ V, Msp.toarray() @ V)
Vsp = spcreator(V)
Msp_Vsp = Msp @ Vsp
assert isinstance(Msp_Vsp, np.ndarray)
assert Msp_Vsp.shape == ()
# output is 0-dim ndarray
assert_allclose(np.array(3), Msp_Vsp)
assert_allclose(np.array(3), Msp.toarray() @ Vsp)
assert_allclose(np.array(3), Msp @ Vsp.toarray())
assert_allclose(np.array(3), Msp.toarray() @ Vsp.toarray())
# check error on matrix-scalar
with pytest.raises(ValueError, match='Scalar operands are not allowed'):
Msp @ 1
with pytest.raises(ValueError, match='Scalar operands are not allowed'):
1 @ Msp
def test_sub_dense(self, spcreator, datsp_math_dtypes):
# subtracting a dense matrix to/from a sparse matrix
for dtype, dat, datsp in datsp_math_dtypes[spcreator]:
if dtype == np.dtype('bool'):
# boolean array subtraction deprecated in 1.9.0
continue
# Manually add to avoid upcasting from scalar
# multiplication.
sum1 = (dat + dat + dat) - datsp
assert_equal(sum1, dat + dat)
sum2 = (datsp + datsp + datsp) - dat
assert_equal(sum2, dat + dat)
def test_size_zero_matrix_arithmetic(self, spcreator):
# Test basic matrix arithmetic with shapes like 0, (1, 0), (0, 3), etc.
mat = np.array([])
a = mat.reshape(0)
d = mat.reshape((1, 0))
f = np.ones([5, 5])
asp = spcreator(a)
dsp = spcreator(d)
# bad shape for addition
with pytest.raises(ValueError, match='inconsistent shapes'):
asp.__add__(dsp)
# matrix product.
assert_equal(asp.dot(asp), np.dot(a, a))
# bad matrix products
with pytest.raises(ValueError, match='dimension mismatch'):
asp.dot(f)
# elemente-wise multiplication
assert_equal(asp.multiply(asp).toarray(), np.multiply(a, a))
assert_equal(asp.multiply(a).toarray(), np.multiply(a, a))
assert_equal(asp.multiply(6).toarray(), np.multiply(a, 6))
# bad element-wise multiplication
with pytest.raises(ValueError, match='inconsistent shapes'):
asp.multiply(f)
# Addition
assert_equal(asp.__add__(asp).toarray(), a.__add__(a))