def _update_dict_slow(self, subset, D_range):
"""Update dictionary from statistic
Parameters
----------
subset: ndarray (len_subset),
Mask used on X
"""
D_subset = self.D_[:, subset]
if self.projection == "full":
norm = enet_norm(self.D_, self.l1_ratio)
else:
norm = enet_norm(D_subset, self.l1_ratio)
R = self.B_[:, subset] - np.dot(D_subset.T, self.A_).T
ger, = linalg.get_blas_funcs(("ger",), (self.A_, D_subset))
for k in D_range:
ger(1.0, self.A_[k], D_subset[k], a=R, overwrite_a=True)
# R += np.dot(stat.A[:, j].reshape(n_components, 1),
D_subset[k] = R[k] / (self.A_[k, k])
if self.projection == "full":
self.D_[k][subset] = D_subset[k]
self.D_[k] = enet_projection(self.D_[k], norm[k], self.l1_ratio)
D_subset[k] = self.D_[k][subset]
else:
D_subset[k] = enet_projection(D_subset[k], norm[k], self.l1_ratio)
ger(-1.0, self.A_[k], D_subset[k], a=R, overwrite_a=True)
# R -= np.dot(stat.A[:, j].reshape(n_components, 1),
if self.projection == "partial":
self.D_[:, subset] = D_subset
def _update_dict_slow(self, subset, D_range):
"""Update dictionary from statistic
Parameters
----------
subset: ndarray (len_subset),
Mask used on X
"""
D_subset = self.D_[:, subset]
if self.projection == 'full':
norm = enet_norm(self.D_, self.l1_ratio)
else:
norm = enet_norm(D_subset, self.l1_ratio)
R = self.B_[:, subset] - np.dot(D_subset.T, self.A_).T
ger, = linalg.get_blas_funcs(('ger', ), (self.A_, D_subset))
for k in D_range:
ger(1.0, self.A_[k], D_subset[k], a=R, overwrite_a=True)
# R += np.dot(stat.A[:, j].reshape(n_components, 1),
D_subset[k] = R[k] / (self.A_[k, k])
if self.projection == 'full':
self.D_[k][subset] = D_subset[k]
self.D_[k] = enet_projection(self.D_[k], norm[k],
self.l1_ratio)
D_subset[k] = self.D_[k][subset]
else:
D_subset[k] = enet_projection(D_subset[k], norm[k],
self.l1_ratio)
ger(-1.0, self.A_[k], D_subset[k], a=R, overwrite_a=True)
# R -= np.dot(stat.A[:, j].reshape(n_components, 1),
if self.projection == 'partial':
self.D_[:, subset] = D_subset
def _update_dict_slow(self, subset, D_range):
"""Update dictionary from statistic
Parameters
----------
subset: ndarray (len_subset),
Mask used on X
Q: ndarray (n_components, n_features):
Dictionary to perform ridge regression
l1_ratio: float in [0, 1]:
Controls the sparsity of the dictionary
stat: DictMFStats,
Statistics kept by the algorithm, to be updated by the function
var_red: boolean,
Online update of the Gram matrix (Experimental)
random_state: int or RandomState
Pseudo number generator state used for random sampling.
"""
D_subset = self.D_[:, subset]
if self.projection == 'full':
norm = enet_norm(self.D_, self.l1_ratio)
else:
if self.var_red != 'weight_based':
self.G_ -= D_subset.dot(D_subset.T)
norm = enet_norm(D_subset, self.l1_ratio)
R = self.B_[:, subset] - np.dot(D_subset.T, self.A_).T
ger, = linalg.get_blas_funcs(('ger',), (self.A_, D_subset))
for k in D_range:
ger(1.0, self.A_[k], D_subset[k], a=R, overwrite_a=True)
# R += np.dot(stat.A[:, j].reshape(n_components, 1),
D_subset[k] = R[k] / (self.A_[k, k])
if self.projection == 'full':
self.D_[k][subset] = D_subset[k]
self.D_[k] = enet_projection(self.D_[k],
norm[k],
self.l1_ratio)
D_subset[k] = self.D_[k][subset]
else:
D_subset[k] = enet_projection(D_subset[k], norm[k],
self.l1_ratio)
ger(-1.0, self.A_[k], D_subset[k], a=R, overwrite_a=True)
# R -= np.dot(stat.A[:, j].reshape(n_components, 1),
if self.projection == 'partial':
self.D_[:, subset] = D_subset
if self.var_red != 'weight_based':
self.G_ += D_subset.dot(D_subset.T)
elif self.var_red != 'weight_based':
self.G_ = self.D_.dot(self.D_.T).T
def test_fast_enet_l2_ball():
random_state = check_random_state(0)
norms = np.zeros(10)
for i in range(10):
a = random_state.randn(100)
c = np.zeros(100)
c[:] = enet_projection(a, 2, 0.0)
norms[i] = np.sqrt(np.sum(c ** 2))
assert_array_almost_equal(norms, np.ones(10) * sqrt(2))
for i in range(10):
a = random_state.randn(100)
a /= np.sqrt(np.sum(a ** 2)) * 10
c = enet_projection(a, 2, 0.0)
assert_array_almost_equal(a, c)
def test_fast_enet_l2_ball():
random_state = check_random_state(0)
norms = np.zeros(10)
for i in range(10):
a = random_state.randn(100)
c = np.zeros(100)
c[:] = enet_projection(a, 2, 0.0)
norms[i] = np.sqrt(np.sum(c**2))
assert_array_almost_equal(norms, np.ones(10) * sqrt(2))
for i in range(10):
a = random_state.randn(100)
a /= np.sqrt(np.sum(a**2)) * 10
c = enet_projection(a, 2, 0.0)
assert_array_almost_equal(a, c)
def _update_dict_slow(self, subset, D_range):
"""Update dictionary from statistic
Parameters
----------
subset: ndarray (len_subset),
Mask used on X
Q: ndarray (n_components, n_features):
Dictionary to perform ridge regression
l1_ratio: float in [0, 1]:
Controls the sparsity of the dictionary
stat: DictMFStats,
Statistics kept by the algorithm, to be updated by the function
var_red: boolean,
Online update of the Gram matrix (Experimental)
random_state: int or RandomState
Pseudo number generator state used for random sampling.
"""
D_subset = self.D_[:, subset]
if self.projection == 'full':
norm = enet_norm(self.D_, self.l1_ratio)
else:
if self.var_red != 'weight_based':
self.G_ -= D_subset.dot(D_subset.T)
norm = enet_norm(D_subset, self.l1_ratio)
R = self.B_[:, subset] - np.dot(D_subset.T, self.A_).T
ger, = linalg.get_blas_funcs(('ger', ), (self.A_, D_subset))
for k in D_range:
ger(1.0, self.A_[k], D_subset[k], a=R, overwrite_a=True)
# R += np.dot(stat.A[:, j].reshape(n_components, 1),
D_subset[k] = R[k] / (self.A_[k, k])
if self.projection == 'full':
self.D_[k][subset] = D_subset[k]
self.D_[k] = enet_projection(self.D_[k], norm[k],
self.l1_ratio)
D_subset[k] = self.D_[k][subset]
else:
D_subset[k] = enet_projection(D_subset[k], norm[k],
self.l1_ratio)
ger(-1.0, self.A_[k], D_subset[k], a=R, overwrite_a=True)
# R -= np.dot(stat.A[:, j].reshape(n_components, 1),
if self.projection == 'partial':
self.D_[:, subset] = D_subset
if self.var_red != 'weight_based':
self.G_ += D_subset.dot(D_subset.T)
elif self.var_red != 'weight_based':
self.G_ = self.D_.dot(self.D_.T).T
def test_fast_enet_l1_ball():
random_state = check_random_state(0)
norms = np.zeros(10)
for i in range(10):
a = random_state.randn(100)
b = np.zeros(100)
b[:] = enet_projection(a, 1, 1.0)
norms[i] = np.sum(np.abs(b))
assert_array_almost_equal(norms, np.ones(10))
def test_fast_enet_projection():
c = np.empty((10, 100))
b = np.empty((10, 100))
random_state = check_random_state(0)
for i in range(10):
a = random_state.randn(100)
b[i, :] = enet_projection_slow(a, radius=1, l1_ratio=0.1)
c[i] = enet_projection(a, radius=1, l1_ratio=0.1)
assert_array_almost_equal(c, b, 4)
def test_fast_enet_l1_ball():
random_state = check_random_state(0)
norms = np.zeros(10)
for i in range(10):
a = random_state.randn(100)
b = np.zeros(100)
b[:] = enet_projection(a, 1, 1.0)
norms[i] = np.sum(np.abs(b))
assert_array_almost_equal(norms, np.ones(10))
def test_fast_enet_projection():
c = np.empty((10, 100))
b = np.empty((10, 100))
random_state = check_random_state(0)
for i in range(10):
a = random_state.randn(100)
b[i, :] = enet_projection_slow(a, radius=1, l1_ratio=0.1)
c[i] = enet_projection(a, radius=1, l1_ratio=0.1)
assert_array_almost_equal(c, b, 4)
def test_fast_enet_projection_norm():
random_state = check_random_state(0)
norms = np.zeros(10)
for i in range(10):
a = random_state.randn(20000)
a /= np.sqrt(np.sum(a ** 2))
c = np.zeros(20000)
c[:] = enet_projection(a, 1, 0.15)
norms[i] = enet_norm(c, l1_ratio=0.15)
assert_array_almost_equal(norms, np.ones(10))
def test_fast_enet_projection_norm():
random_state = check_random_state(0)
norms = np.zeros(10)
for i in range(10):
a = random_state.randn(20000)
a /= np.sqrt(np.sum(a**2))
c = np.zeros(20000)
c[:] = enet_projection(a, 1, 0.15)
norms[i] = enet_norm(c, l1_ratio=0.15)
assert_array_almost_equal(norms, np.ones(10))
