https://github.com/tensorly/tensorly
Tip revision: 14ca3224f65a437536bfc8c3988c85b16947076c authored by Jean Kossaifi on 08 May 2018, 21:04:53 UTC
FIX README + bump version
FIX README + bump version
Tip revision: 14ca322
test_candecomp_parafac.py
import numpy as np
import pytest
from ..candecomp_parafac import (
parafac, non_negative_parafac, normalize_factors, initialize_factors)
from ...kruskal_tensor import kruskal_to_tensor
from ...random import check_random_state
from ... import backend as T
def test_parafac():
"""Test for the CANDECOMP-PARAFAC decomposition
"""
rng = check_random_state(1234)
tol_norm_2 = 10e-2
tol_max_abs = 10e-2
tensor = T.tensor(rng.random_sample((3, 4, 2)))
factors_svd = parafac(tensor, rank=4, n_iter_max=200, init='svd', tol=10e-5)
factors_random = parafac(tensor, rank=4, n_iter_max=200, init='random', tol=10e-5, random_state=1234, verbose=0)
rec_svd = kruskal_to_tensor(factors_svd)
rec_random = kruskal_to_tensor(factors_random)
error = T.norm(rec_svd - tensor, 2)
error /= T.norm(tensor, 2)
T.assert_(error < tol_norm_2,
'norm 2 of reconstruction higher than tol')
# Test the max abs difference between the reconstruction and the tensor
T.assert_(T.max(T.abs(rec_svd - tensor)) < tol_max_abs,
'abs norm of reconstruction error higher than tol')
tol_norm_2 = 10e-1
tol_max_abs = 10e-1
error = T.norm(rec_svd - rec_random, 2)
error /= T.norm(rec_svd, 2)
T.assert_(error < tol_norm_2,
'norm 2 of difference between svd and random init too high')
T.assert_(T.max(T.abs(rec_svd - rec_random)) < tol_max_abs,
'abs norm of difference between svd and random init too high')
with np.testing.assert_raises(ValueError):
rank = 4
_ = initialize_factors(tensor, rank, init='bogus init type')
def test_non_negative_parafac():
"""Test for non-negative PARAFAC
TODO: more rigorous test
"""
tol_norm_2 = 10e-1
tol_max_abs = 1
rng = check_random_state(1234)
tensor = T.tensor(rng.random_sample((3, 3, 3))+1)
factors = parafac(tensor, rank=3, n_iter_max=120)
nn_factors = non_negative_parafac(tensor, rank=3, n_iter_max=100, tol=10e-4, init='svd', verbose=0)
# Make sure all components are positive
for factor in nn_factors:
T.assert_(T.all(factor >= 0))
reconstructed_tensor = kruskal_to_tensor(factors)
nn_reconstructed_tensor = kruskal_to_tensor(nn_factors)
error = T.norm(reconstructed_tensor - nn_reconstructed_tensor, 2)
error /= T.norm(reconstructed_tensor, 2)
T.assert_(error < tol_norm_2,
'norm 2 of reconstruction higher than tol')
# Test the max abs difference between the reconstruction and the tensor
T.assert_(T.max(T.abs(reconstructed_tensor - nn_reconstructed_tensor)) < tol_max_abs,
'abs norm of reconstruction error higher than tol')
factors_svd = non_negative_parafac(tensor, rank=3, n_iter_max=100, tol=10e-4,
init='svd')
factors_random = non_negative_parafac(tensor, rank=3, n_iter_max=100, tol=10e-4,
init='random', random_state=1234, verbose=0)
rec_svd = kruskal_to_tensor(factors_svd)
rec_random = kruskal_to_tensor(factors_random)
error = T.norm(rec_svd - rec_random, 2)
error /= T.norm(rec_svd, 2)
T.assert_(error < tol_norm_2,
'norm 2 of difference between svd and random init too high')
T.assert_(T.max(T.abs(rec_svd - rec_random)) < tol_max_abs,
'abs norm of difference between svd and random init too high')