def test_mttkrp(self):
t = T.array([[[0, 1, 3, 4], [4, 0, 2, 1], [4, 2, 3, 4]],
[[2, 4, 2, 3], [3, 3, 2, 4], [2, 3, 0, 2]]])
dtensor = sktensor.dtensor(t)
factors = [None] * 3
for i in range(3):
factors[i] = T.random.rand(t.shape[i], 3)
# print(factors)
# print(dtensor.uttkrp(factors, 0))
# print(mttkrp(t, factors, 0))
# T.testing.assert_array_almost_equal(dtensor.uttkrp(factors, 0),
# mttkrp(t, factors, 0))
# 对展开的细节理解不一样，导致结果不一样，但是两种算法都没有问题
# T.testing.assert_array_almost_equal(dtensor.unfold(0), unfold(t, 0))
from sktensor.core import khatrirao
from .. import seq_kr
order = list(range(0)) + list(range(0 + 1, 3))
T.testing.assert_array_almost_equal(
seq_kr(factors, exclude=0, reverse=True),
khatrirao(tuple(factors[i] for i in order), reverse=True))
order = list(range(1)) + list(range(1 + 1, 3))
T.testing.assert_array_almost_equal(
seq_kr(factors, exclude=1, reverse=True),
khatrirao(tuple(factors[i] for i in order), reverse=True))
def test_khatrirao():
A = np.array([[1, 2, 3], [4, 5, 6], [7, 8, 9]])
B = np.array([[1, 4, 7], [2, 5, 8], [3, 6, 9]])
C = np.array([[1, 8, 21], [2, 10, 24], [3, 12, 27], [4, 20, 42],
[8, 25, 48], [12, 30, 54], [7, 32, 63], [14, 40, 72],
[21, 48, 81]])
D = khatrirao((A, B))
assert C.shape == D.shape
assert (C == D).all()
def toarray(self):
"""
Converts a ktensor into a dense multidimensional ndarray
Returns
-------
arr : np.ndarray
Fully computed multidimensional array whose shape matches
the original ktensor.
"""
A = dot(self.lmbda, khatrirao(tuple(self.U)).T)
return A.reshape(self.shape)
def toarray(self):
"""
Converts a ktensor into a dense multidimensional ndarray
Returns
-------
arr : np.ndarray
Fully computed multidimensional array whose shape matches
the original ktensor.
"""
A = dot(self.lmbda, khatrirao(tuple(self.U)).T)
return A.reshape(self.shape)
def uttkrp(self, U, n):
order = range(n) + range(n + 1, self.ndim)
Z = khatrirao(tuple(U[i] for i in order), reverse=True)
return self.unfold(n).dot(Z)
