def build_poe(self, b_frames, tau_frames, win_mats, sdw=None):
"""
Computes natural parameters for a Gaussian product-of-experts model.
The natural parameters (b-value vector and precision matrix) are returned.
The returned precision matrix is stored as a BandMat.
Mathematically the b-value vector is given as:
b = \sum_d \transpose{W_d} \tilde{b}_d
and the precision matrix is given as:
P = \sum_d \transpose{W_d} \text{diag}(\tilde{tau}_d) W_d
where $W_d$ is the window matrix for window $d$ as specified by an element
of `win_mats`, $\tilde{b}_d$ is the sequence over time of b-value
parameters for window $d$ as given by a column of `b_frames`, and
$\tilde{\tau}_d$ is the sequence over time of precision parameters for
window $d$ as given by a column of `tau_frames`.
"""
if sdw is None:
sdw = max([win_mat.l + win_mat.u for win_mat in win_mats])
num_windows = len(win_mats)
frames = len(b_frames)
assert np.shape(b_frames) == (frames, num_windows)
assert np.shape(tau_frames) == (frames, num_windows)
assert all([win_mat.l + win_mat.u <= sdw for win_mat in win_mats])
b = np.zeros((frames, ))
prec = bm.zeros(sdw, sdw, frames)
for win_index, win_mat in enumerate(win_mats):
bm.dot_mv_plus_equals(win_mat.T, b_frames[:, win_index], target=b)
bm.dot_mm_plus_equals(win_mat.T,
win_mat,
target_bm=prec,
diag=float64(tau_frames[:, win_index]))
return b, prec
def test_dot_mv_plus_equals(self, its=100):
for it in range(its):
size = random.choice([0, 1, randint(0, 10), randint(0, 100)])
a_bm = gen_BandMat(size)
b = randn(size)
c = randn(size)
a_full = a_bm.full()
c_good = c.copy()
array_mem = get_array_mem(a_bm.data, b, c)
bm.dot_mv_plus_equals(a_bm, b, c)
c_good += np.dot(a_full, b)
assert_allclose(c, c_good)
assert get_array_mem(a_bm.data, b, c) == array_mem
def test_dot_mv_plus_equals(self, its=100):
for it in range(its):
size = random.choice([0, 1, randint(0, 10), randint(0, 100)])
a_bm = gen_BandMat(size)
b = randn(size)
c = randn(size)
a_full = a_bm.full()
c_good = c.copy()
array_mem = get_array_mem(a_bm.data, b, c)
bm.dot_mv_plus_equals(a_bm, b, c)
c_good += np.dot(a_full, b)
assert_allclose(c, c_good)
assert get_array_mem(a_bm.data, b, c) == array_mem
def build_poe(self, b_frames, tau_frames, win_mats, sdw=None):
if sdw is None:
sdw = max([win_mat.l + win_mat.u for win_mat in win_mats])
num_windows = len(win_mats)
frames = len(b_frames)
assert np.shape(b_frames) == (frames, num_windows)
assert np.shape(tau_frames) == (frames, num_windows)
assert all([win_mat.l + win_mat.u <= sdw for win_mat in win_mats])
b = np.zeros((frames,))
prec = bm.zeros(sdw, sdw, frames)
for win_index, win_mat in enumerate(win_mats):
bm.dot_mv_plus_equals(win_mat.T, b_frames[:, win_index], target=b)
bm.dot_mm_plus_equals(win_mat.T, win_mat, target_bm=prec, diag=float64(tau_frames[:, win_index]))
return b, prec
def build_poe(self, b_frames, tau_frames, win_mats, sdw=None):
# tau_frames.astype('float64')
if sdw is None:
sdw = max([ win_mat.l + win_mat.u for win_mat in win_mats ])
num_windows = len(win_mats)
frames = len(b_frames)
assert np.shape(b_frames) == (frames, num_windows)
assert np.shape(tau_frames) == (frames, num_windows)
assert all([ win_mat.l + win_mat.u <= sdw for win_mat in win_mats ])
b = np.zeros((frames,))
prec = bm.zeros(sdw, sdw, frames)
for win_index, win_mat in enumerate(win_mats):
bm.dot_mv_plus_equals(win_mat.T, b_frames[:, win_index], target=b)
bm.dot_mm_plus_equals(win_mat.T, win_mat, target_bm=prec,
diag=float64(tau_frames[:, win_index]))
return b, prec
def build_poe(b_frames, tau_frames, win_mats, sdw=None):
if sdw is None:
sdw = max([win_mat.l + win_mat.u for win_mat in win_mats])
num_windows = len(win_mats)
frames = len(b_frames)
b = numpy.zeros((frames, ))
prec = bandmat.zeros(sdw, sdw, frames)
for win_index, win_mat in enumerate(win_mats):
bandmat.dot_mv_plus_equals(win_mat.T, b_frames[:, win_index], target=b)
bandmat.dot_mm_plus_equals(win_mat.T,
win_mat,
target_bm=prec,
diag=tau_frames[:, win_index])
return b, prec
def build_poe(b_frames, tau_frames, win_mats, sdw=None):
r"""Computes natural parameters for a Gaussian product-of-experts model.
The natural parameters (b-value vector and precision matrix) are returned.
The returned precision matrix is stored as a BandMat.
Mathematically the b-value vector is given as:
b = \sum_d \transpose{W_d} \tilde{b}_d
and the precision matrix is given as:
P = \sum_d \transpose{W_d} \text{diag}(\tilde{tau}_d) W_d
where $W_d$ is the window matrix for window $d$ as specified by an element
of `win_mats`, $\tilde{b}_d$ is the sequence over time of b-value
parameters for window $d$ as given by a column of `b_frames`, and
$\tilde{\tau}_d$ is the sequence over time of precision parameters for
window $d$ as given by a column of `tau_frames`.
"""
if sdw is None:
sdw = max([ win_mat.l + win_mat.u for win_mat in win_mats ])
num_windows = len(win_mats)
frames = len(b_frames)
assert np.shape(b_frames) == (frames, num_windows)
assert np.shape(tau_frames) == (frames, num_windows)
assert all([ win_mat.l + win_mat.u <= sdw for win_mat in win_mats ])
b = np.zeros((frames,))
prec = bm.zeros(sdw, sdw, frames)
for win_index, win_mat in enumerate(win_mats):
bm.dot_mv_plus_equals(win_mat.T, b_frames[:, win_index], target=b)
bm.dot_mm_plus_equals(win_mat.T, win_mat, target_bm=prec,
diag=tau_frames[:, win_index])
return b, prec