def test_store_and_load_from_checkpoint(self, rank4_parafac2_tensor):
max_its = 20
checkpoint_frequency = 5
X = rank4_parafac2_tensor.construct_tensor()
with tempfile.TemporaryDirectory() as tempfolder:
checkpoint_path = f'{tempfolder}/checkpoint.h5'
parafac2_decomposer = parafac2.Parafac2_ALS(
4, max_its=max_its, convergence_tol=1e-20, checkpoint_frequency=checkpoint_frequency,
checkpoint_path=checkpoint_path, print_frequency=-5
)
decomposition = parafac2_decomposer.fit_transform(X)
assert Path(checkpoint_path).is_file()
with h5py.File(checkpoint_path) as h5:
for i in range(max_its):
if (i+1) % checkpoint_frequency == 0:
assert f'checkpoint_{i:05d}' in h5
parafac2_decomposer2 = parafac2.Parafac2_ALS(
4, max_its=100, convergence_tol=1e-20, init='from_checkpoint'
)
parafac2_decomposer2._init_fit(X, 100, checkpoint_path)
assert np.allclose(parafac2_decomposer.decomposition.A, parafac2_decomposer2.decomposition.A)
assert np.allclose(parafac2_decomposer.decomposition.C, parafac2_decomposer2.decomposition.C)
assert np.allclose(parafac2_decomposer.decomposition.blueprint_B, parafac2_decomposer2.decomposition.blueprint_B)
for P1, P2 in zip(parafac2_decomposer.decomposition.projection_matrices, parafac2_decomposer2.decomposition.projection_matrices):
assert np.allclose(P1, P2)
def test_rank4_decomposition(self, rank4_parafac2_tensor):
X = (rank4_parafac2_tensor.construct_tensor())
parafac2_als = parafac2.Parafac2_ALS(4, max_its=1000, convergence_tol=1e-10, print_frequency=1000)
estimated_pf2tensor = parafac2_als.fit_transform(X)
estimated_X = estimated_pf2tensor.construct_tensor()
assert np.allclose(X, estimated_X, rtol=1e-5, atol=1)
def test_rank4_non_negative_decomposition(self, rank4_parafac2_tensor):
X = (rank4_parafac2_tensor.construct_tensor())
parafac2_als = parafac2.Parafac2_ALS(4, non_negativity_constraints=[False, True, True], max_its=1000, convergence_tol=1e-10, print_frequency=1000)
for constraint in itertools.product([True, False], repeat=3):
parafac2_als.non_negativity_constraints = constraint
estimated_pf2tensor = parafac2_als.fit_transform(X)
estimated_X = estimated_pf2tensor.construct_tensor()
assert np.allclose(X, estimated_X, rtol=1e-5, atol=1)
def test_rank4_monotone_convergence(self, rank4_parafac2_tensor):
X = (rank4_parafac2_tensor.construct_tensor())
parafac2_als = parafac2.Parafac2_ALS(4, max_its=100, convergence_tol=1e-10, print_frequency=1000)
parafac2_als._update_parafac2_factors = ensure_monotonicity(
parafac2_als,
'_update_parafac2_factors',
'SSE',
atol=1e-8,
rtol=1e-4
)
parafac2_als.fit_transform(X)
def test_rank4_decomposition_correlation(self, rank4_parafac2_tensor):
X = (rank4_parafac2_tensor.construct_tensor())
parafac2_als = parafac2.Parafac2_ALS(4, max_its=1000, convergence_tol=1e-10, print_frequency=1000)
estimated_pf2tensor = parafac2_als.fit_transform(X)
self.assert_correlation(rank4_parafac2_tensor.A, estimated_pf2tensor.A)
for P1, P2 in zip(rank4_parafac2_tensor.projection_matrices, estimated_pf2tensor.projection_matrices):
Bk = P1@rank4_parafac2_tensor.blueprint_B
estimated_Bk = P2@estimated_pf2tensor.blueprint_B
self.assert_correlation(Bk, estimated_Bk)
self.assert_correlation(rank4_parafac2_tensor.C, estimated_pf2tensor.C)
def test_rank4_non_negative_decomposition_correlation(self, rank4_parafac2_tensor):
X = (rank4_parafac2_tensor.construct_tensor())
parafac2_als = parafac2.Parafac2_ALS(4, non_negativity_constraints=[False, False, True], max_its=1000, convergence_tol=1e-10, print_frequency=1000)
for constraint in itertools.product([True, False], repeat=3):
parafac2_als.non_negativity_constraints = constraint
estimated_pf2tensor = parafac2_als.fit_transform(X)
self.assert_correlation(rank4_parafac2_tensor.A, estimated_pf2tensor.A)
for P1, P2 in zip(rank4_parafac2_tensor.projection_matrices, estimated_pf2tensor.projection_matrices):
Bk = P1@rank4_parafac2_tensor.blueprint_B
estimated_Bk = P2@estimated_pf2tensor.blueprint_B
self.assert_correlation(Bk, estimated_Bk)
self.assert_correlation(rank4_parafac2_tensor.C, estimated_pf2tensor.C)