def test_als(self):
""" Test accuracy of alternating least-squares NMF algorithm
against the MATLAB-computed version
"""
# set data and initializing constants
keys = [array([i + 1]) for i in range(4)]
data_local = array([[1.0, 2.0, 6.0], [1.0, 3.0, 0.0], [1.0, 4.0, 6.0],
[5.0, 1.0, 4.0]])
data = self.sc.parallelize(zip(keys, data_local))
mat = RowMatrix(data)
h0 = array([[0.09082617, 0.85490047, 0.57234593],
[0.82766740, 0.21301186, 0.90913979]])
# if the rows of h are not normalized on each iteration:
h_true = array([[0., 0.6010, 0.9163], [0.8970, 0.1556, 0.7423]])
w_true = array([[4.5885, 1.5348], [1.3651, 0.2184], [5.9349, 1.0030],
[0., 5.5147]])
# if the columns of h are normalized (as in the current implementation):
scale_mat = diag(norm(h_true, axis=1))
h_true = dot(LinAlg.inv(scale_mat), h_true)
w_true = dot(w_true, scale_mat)
# calculate NMF using the Thunder implementation
# (maxiter=9 corresponds with Matlab algorithm)
nmf_thunder = NMF(k=2, method="als", h0=h0, maxiter=9)
nmf_thunder.fit(mat)
h_thunder = nmf_thunder.h
w_thunder = array(nmf_thunder.w.values().collect())
tol = 1e-03 # allow small error
assert (allclose(w_thunder, w_true, atol=tol))
assert (allclose(h_thunder, h_true, atol=tol))
def test_init(self):
"""
test performance of whole function, including random initialization
"""
data_local = array([[1.0, 2.0, 6.0], [1.0, 3.0, 0.0], [1.0, 4.0, 6.0],
[5.0, 1.0, 4.0]])
data = self.sc.parallelize(
zip([array([i]) for i in range(data_local.shape[0])], data_local))
mat = RowMatrix(data)
nmf_thunder = NMF(k=2, recon_hist='final')
nmf_thunder.fit(mat)
# check to see if Thunder's solution achieves close-to-optimal reconstruction error
# scikit-learn's solution achieves 2.993952
# matlab's non-deterministic implementation usually achieves < 2.9950 (when it converges)
assert (nmf_thunder.recon_err < 2.9950)
def test_init(self):
"""
test performance of whole function, including random initialization
"""
dataLocal = array([
[1.0, 2.0, 6.0],
[1.0, 3.0, 0.0],
[1.0, 4.0, 6.0],
[5.0, 1.0, 4.0]])
data = self.sc.parallelize(zip([array([i]) for i in range(dataLocal.shape[0])], dataLocal))
mat = RowMatrix(data)
nmfThunder = NMF(k=2, reconHist='final')
nmfThunder.fit(mat)
# check to see if Thunder's solution achieves close-to-optimal reconstruction error
# scikit-learn's solution achieves 2.993952
# matlab's non-deterministic implementation usually achieves < 2.9950 (when it converges)
assert(nmfThunder.reconErr < 2.9950)
def test_als(self):
""" Test accuracy of alternating least-squares NMF algorithm
against the MATLAB-computed version
"""
# set data and initializing constants
keys = [array([i+1]) for i in range(4)]
dataLocal = array([
[1.0, 2.0, 6.0],
[1.0, 3.0, 0.0],
[1.0, 4.0, 6.0],
[5.0, 1.0, 4.0]])
data = self.sc.parallelize(zip(keys, dataLocal))
mat = RowMatrix(data)
h0 = array(
[[0.09082617, 0.85490047, 0.57234593],
[0.82766740, 0.21301186, 0.90913979]])
# if the rows of h are not normalized on each iteration:
hTrue = array(
[[0. , 0.6010, 0.9163],
[0.8970, 0.1556, 0.7423]])
wTrue = array(
[[4.5885, 1.5348],
[1.3651, 0.2184],
[5.9349, 1.0030],
[0. , 5.5147]])
# if the columns of h are normalized (as in the current implementation):
scaleMat = diag(norm(hTrue, axis=1))
hTrue = dot(LinAlg.inv(scaleMat), hTrue)
wTrue = dot(wTrue, scaleMat)
# calculate NMF using the Thunder implementation
# (maxiter=9 corresponds with Matlab algorithm)
nmfThunder = NMF(k=2, method="als", h0=h0, maxIter=9)
nmfThunder.fit(mat)
hThunder = nmfThunder.h
wThunder = array(nmfThunder.w.values().collect())
tol = 1e-03 # allow small error
assert(allclose(wThunder, wTrue, atol=tol))
assert(allclose(hThunder, hTrue, atol=tol))
