def maxvol(A):
""" Find the rxr submatrix of maximal volume in A(nxr), n>=r
We want to decompose matrix A as
A = A[:,J] * (A[I,J])^-1 * A[I,:]
This algorithm helps us find this submatrix A[I,J] from A, which has the largest determinant.
We greedily find vector of max norm, and subtract its projection from the rest of rows.
Parameters
----------
A: matrix
The matrix to find maximal volume
Returns
-------
row_idx: list of int
is the list or rows of A forming the matrix with maximal volume,
A_inv: matrix
is the inverse of the matrix with maximal volume.
References
----------
S. A. Goreinov, I. V. Oseledets, D. V. Savostyanov, E. E. Tyrtyshnikov, N. L. Zamarashkin.
How to find a good submatrix.Goreinov, S. A., et al.
Matrix Methods: Theory, Algorithms and Applications: Dedicated to the Memory of Gene Golub. 2010. 247-256.
Ali Çivril, Malik Magdon-Ismail
On selecting a maximum volume sub-matrix of a matrix and related problems
Theoretical Computer Science. Volume 410, Issues 47–49, 6 November 2009, Pages 4801-4811
"""
(n, r) = tl.shape(A)
# The index of row of the submatrix
row_idx = tl.zeros(r)
# Rest of rows / unselected rows
rest_of_rows = tl.tensor(list(range(n)),dtype= tl.int64)
# Find r rows iteratively
i = 0
A_new = A
while i < r:
mask = list(range(tl.shape(A_new)[0]))
# Compute the square of norm of each row
rows_norms = tl.sum(A_new ** 2, axis=1)
# If there is only one row of A left, let's just return it. MxNet is not robust about this case.
if tl.shape(rows_norms) == ():
row_idx[i] = rest_of_rows
break
# If a row is 0, we delete it.
if any(rows_norms == 0):
zero_idx = tl.argmin(rows_norms,axis=0)
mask.pop(zero_idx)
rest_of_rows = rest_of_rows[mask]
A_new = A_new[mask,:]
continue
# Find the row of max norm
max_row_idx = tl.argmax(rows_norms, axis=0)
max_row = A[rest_of_rows[max_row_idx], :]
# Compute the projection of max_row to other rows
# projection a to b is computed as: <a,b> / sqrt(|a|*|b|)
projection = tl.dot(A_new, tl.transpose(max_row))
normalization = tl.sqrt(rows_norms[max_row_idx] * rows_norms)
# make sure normalization vector is of the same shape of projection (causing bugs for MxNet)
normalization = tl.reshape(normalization, tl.shape(projection))
projection = projection/normalization
# Subtract the projection from A_new: b <- b - a * projection
A_new = A_new - A_new * tl.reshape(projection, (tl.shape(A_new)[0], 1))
# Delete the selected row
mask.pop(max_row_idx)
A_new = A_new[mask,:]
# update the row_idx and rest_of_rows
row_idx[i] = rest_of_rows[max_row_idx]
rest_of_rows = rest_of_rows[mask]
i = i + 1
row_idx = tl.tensor(row_idx, dtype=tl.int64)
inverse = tl.solve(A[row_idx,:],
tl.eye(tl.shape(A[row_idx,:])[0], **tl.context(A)))
row_idx = tl.to_numpy(row_idx)
return row_idx, inverse
def active_set_nnls(Utm, UtU, x=None, n_iter_max=100, tol=10e-8):
"""
Active set algorithm for non-negative least square solution.
Computes an approximate non-negative solution for Ux=m linear system.
Parameters
----------
Utm : vectorized ndarray
Pre-computed product of the transposed of U and m
UtU : ndarray
Pre-computed Kronecker product of the transposed of U and U
x : init
Default: None
n_iter_max : int
Maximum number of iteration
Default: 100
tol : float
Early stopping criterion
Returns
-------
x : ndarray
Notes
-----
This function solves following problem:
.. math::
\\begin{equation}
\\min_{x} ||Ux - m||^2
\\end{equation}
According to [1], non-negativity-constrained least square estimation problem becomes:
.. math::
\\begin{equation}
x' = (Utm) - (UTU)\\times x
\\end{equation}
Reference
----------
[1] : Bro, R., & De Jong, S. (1997). A fast non‐negativity‐constrained
least squares algorithm. Journal of Chemometrics: A Journal of
the Chemometrics Society, 11(5), 393-401.
"""
if tl.get_backend() == 'tensorflow':
raise ValueError(
"Active set is not supported with the tensorflow backend. Consider using fista method with tensorflow."
)
if x is None:
x_vec = tl.zeros(tl.shape(UtU)[1], **tl.context(UtU))
else:
x_vec = tl.base.tensor_to_vec(x)
x_gradient = Utm - tl.dot(UtU, x_vec)
passive_set = x_vec > 0
active_set = x_vec <= 0
support_vec = tl.zeros(tl.shape(x_vec), **tl.context(x_vec))
for iteration in range(n_iter_max):
if iteration > 0 or tl.all(x_vec == 0):
indice = tl.argmax(x_gradient)
passive_set = tl.index_update(passive_set, tl.index[indice], True)
active_set = tl.index_update(active_set, tl.index[indice], False)
# To avoid singularity error when initial x exists
try:
passive_solution = tl.solve(UtU[passive_set, :][:, passive_set],
Utm[passive_set])
indice_list = []
for i in range(tl.shape(support_vec)[0]):
if passive_set[i]:
indice_list.append(i)
support_vec = tl.index_update(
support_vec, tl.index[int(i)],
passive_solution[len(indice_list) - 1])
else:
support_vec = tl.index_update(support_vec,
tl.index[int(i)], 0)
# Start from zeros if solve is not achieved
except:
x_vec = tl.zeros(tl.shape(UtU)[1])
support_vec = tl.zeros(tl.shape(x_vec), **tl.context(x_vec))
passive_set = x_vec > 0
active_set = x_vec <= 0
if tl.any(active_set):
indice = tl.argmax(x_gradient)
passive_set = tl.index_update(passive_set, tl.index[indice],
True)
active_set = tl.index_update(active_set, tl.index[indice],
False)
passive_solution = tl.solve(UtU[passive_set, :][:, passive_set],
Utm[passive_set])
indice_list = []
for i in range(tl.shape(support_vec)[0]):
if passive_set[i]:
indice_list.append(i)
support_vec = tl.index_update(
support_vec, tl.index[int(i)],
passive_solution[len(indice_list) - 1])
else:
support_vec = tl.index_update(support_vec,
tl.index[int(i)], 0)
# update support vector if it is necessary
if tl.min(support_vec[passive_set]) <= 0:
for i in range(len(passive_set)):
alpha = tl.min(
x_vec[passive_set][support_vec[passive_set] <= 0] /
(x_vec[passive_set][support_vec[passive_set] <= 0] -
support_vec[passive_set][support_vec[passive_set] <= 0]))
update = alpha * (support_vec - x_vec)
x_vec = x_vec + update
passive_set = x_vec > 0
active_set = x_vec <= 0
passive_solution = tl.solve(
UtU[passive_set, :][:, passive_set], Utm[passive_set])
indice_list = []
for i in range(tl.shape(support_vec)[0]):
if passive_set[i]:
indice_list.append(i)
support_vec = tl.index_update(
support_vec, tl.index[int(i)],
passive_solution[len(indice_list) - 1])
else:
support_vec = tl.index_update(support_vec,
tl.index[int(i)], 0)
if tl.any(passive_set) != True or tl.min(
support_vec[passive_set]) > 0:
break
# set x to s
x_vec = tl.clip(support_vec, 0, tl.max(support_vec))
# gradient update
x_gradient = Utm - tl.dot(UtU, x_vec)
if tl.any(active_set) != True or tl.max(x_gradient[active_set]) <= tol:
break
return x_vec