You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

736 lines
20 KiB
Python

6 years ago
"""Functions used by least-squares algorithms."""
from __future__ import division, print_function, absolute_import
from math import copysign
import numpy as np
from numpy.linalg import norm
from scipy.linalg import cho_factor, cho_solve, LinAlgError
from scipy.sparse import issparse
from scipy.sparse.linalg import LinearOperator, aslinearoperator
EPS = np.finfo(float).eps
# Functions related to a trust-region problem.
def intersect_trust_region(x, s, Delta):
"""Find the intersection of a line with the boundary of a trust region.
This function solves the quadratic equation with respect to t
||(x + s*t)||**2 = Delta**2.
Returns
-------
t_neg, t_pos : tuple of float
Negative and positive roots.
Raises
------
ValueError
If `s` is zero or `x` is not within the trust region.
"""
a = np.dot(s, s)
if a == 0:
raise ValueError("`s` is zero.")
b = np.dot(x, s)
c = np.dot(x, x) - Delta**2
if c > 0:
raise ValueError("`x` is not within the trust region.")
d = np.sqrt(b*b - a*c) # Root from one fourth of the discriminant.
# Computations below avoid loss of significance, see "Numerical Recipes".
q = -(b + copysign(d, b))
t1 = q / a
t2 = c / q
if t1 < t2:
return t1, t2
else:
return t2, t1
def solve_lsq_trust_region(n, m, uf, s, V, Delta, initial_alpha=None,
rtol=0.01, max_iter=10):
"""Solve a trust-region problem arising in least-squares minimization.
This function implements a method described by J. J. More [1]_ and used
in MINPACK, but it relies on a single SVD of Jacobian instead of series
of Cholesky decompositions. Before running this function, compute:
``U, s, VT = svd(J, full_matrices=False)``.
Parameters
----------
n : int
Number of variables.
m : int
Number of residuals.
uf : ndarray
Computed as U.T.dot(f).
s : ndarray
Singular values of J.
V : ndarray
Transpose of VT.
Delta : float
Radius of a trust region.
initial_alpha : float, optional
Initial guess for alpha, which might be available from a previous
iteration. If None, determined automatically.
rtol : float, optional
Stopping tolerance for the root-finding procedure. Namely, the
solution ``p`` will satisfy ``abs(norm(p) - Delta) < rtol * Delta``.
max_iter : int, optional
Maximum allowed number of iterations for the root-finding procedure.
Returns
-------
p : ndarray, shape (n,)
Found solution of a trust-region problem.
alpha : float
Positive value such that (J.T*J + alpha*I)*p = -J.T*f.
Sometimes called Levenberg-Marquardt parameter.
n_iter : int
Number of iterations made by root-finding procedure. Zero means
that Gauss-Newton step was selected as the solution.
References
----------
.. [1] More, J. J., "The Levenberg-Marquardt Algorithm: Implementation
and Theory," Numerical Analysis, ed. G. A. Watson, Lecture Notes
in Mathematics 630, Springer Verlag, pp. 105-116, 1977.
"""
def phi_and_derivative(alpha, suf, s, Delta):
"""Function of which to find zero.
It is defined as "norm of regularized (by alpha) least-squares
solution minus `Delta`". Refer to [1]_.
"""
denom = s**2 + alpha
p_norm = norm(suf / denom)
phi = p_norm - Delta
phi_prime = -np.sum(suf ** 2 / denom**3) / p_norm
return phi, phi_prime
suf = s * uf
# Check if J has full rank and try Gauss-Newton step.
if m >= n:
threshold = EPS * m * s[0]
full_rank = s[-1] > threshold
else:
full_rank = False
if full_rank:
p = -V.dot(uf / s)
if norm(p) <= Delta:
return p, 0.0, 0
alpha_upper = norm(suf) / Delta
if full_rank:
phi, phi_prime = phi_and_derivative(0.0, suf, s, Delta)
alpha_lower = -phi / phi_prime
else:
alpha_lower = 0.0
if initial_alpha is None or not full_rank and initial_alpha == 0:
alpha = max(0.001 * alpha_upper, (alpha_lower * alpha_upper)**0.5)
else:
alpha = initial_alpha
for it in range(max_iter):
if alpha < alpha_lower or alpha > alpha_upper:
alpha = max(0.001 * alpha_upper, (alpha_lower * alpha_upper)**0.5)
phi, phi_prime = phi_and_derivative(alpha, suf, s, Delta)
if phi < 0:
alpha_upper = alpha
ratio = phi / phi_prime
alpha_lower = max(alpha_lower, alpha - ratio)
alpha -= (phi + Delta) * ratio / Delta
if np.abs(phi) < rtol * Delta:
break
p = -V.dot(suf / (s**2 + alpha))
# Make the norm of p equal to Delta, p is changed only slightly during
# this. It is done to prevent p lie outside the trust region (which can
# cause problems later).
p *= Delta / norm(p)
return p, alpha, it + 1
def solve_trust_region_2d(B, g, Delta):
"""Solve a general trust-region problem in 2 dimensions.
The problem is reformulated as a 4-th order algebraic equation,
the solution of which is found by numpy.roots.
Parameters
----------
B : ndarray, shape (2, 2)
Symmetric matrix, defines a quadratic term of the function.
g : ndarray, shape (2,)
Defines a linear term of the function.
Delta : float
Radius of a trust region.
Returns
-------
p : ndarray, shape (2,)
Found solution.
newton_step : bool
Whether the returned solution is the Newton step which lies within
the trust region.
"""
try:
R, lower = cho_factor(B)
p = -cho_solve((R, lower), g)
if np.dot(p, p) <= Delta**2:
return p, True
except LinAlgError:
pass
a = B[0, 0] * Delta**2
b = B[0, 1] * Delta**2
c = B[1, 1] * Delta**2
d = g[0] * Delta
f = g[1] * Delta
coeffs = np.array(
[-b + d, 2 * (a - c + f), 6 * b, 2 * (-a + c + f), -b - d])
t = np.roots(coeffs) # Can handle leading zeros.
t = np.real(t[np.isreal(t)])
p = Delta * np.vstack((2 * t / (1 + t**2), (1 - t**2) / (1 + t**2)))
value = 0.5 * np.sum(p * B.dot(p), axis=0) + np.dot(g, p)
i = np.argmin(value)
p = p[:, i]
return p, False
def update_tr_radius(Delta, actual_reduction, predicted_reduction,
step_norm, bound_hit):
"""Update the radius of a trust region based on the cost reduction.
Returns
-------
Delta : float
New radius.
ratio : float
Ratio between actual and predicted reductions. Zero if predicted
reduction is zero.
"""
if predicted_reduction > 0:
ratio = actual_reduction / predicted_reduction
else:
ratio = 0
if ratio < 0.25:
Delta = 0.25 * step_norm
elif ratio > 0.75 and bound_hit:
Delta *= 2.0
return Delta, ratio
# Construction and minimization of quadratic functions.
def build_quadratic_1d(J, g, s, diag=None, s0=None):
"""Parameterize a multivariate quadratic function along a line.
The resulting univariate quadratic function is given as follows:
::
f(t) = 0.5 * (s0 + s*t).T * (J.T*J + diag) * (s0 + s*t) +
g.T * (s0 + s*t)
Parameters
----------
J : ndarray, sparse matrix or LinearOperator shape (m, n)
Jacobian matrix, affects the quadratic term.
g : ndarray, shape (n,)
Gradient, defines the linear term.
s : ndarray, shape (n,)
Direction vector of a line.
diag : None or ndarray with shape (n,), optional
Addition diagonal part, affects the quadratic term.
If None, assumed to be 0.
s0 : None or ndarray with shape (n,), optional
Initial point. If None, assumed to be 0.
Returns
-------
a : float
Coefficient for t**2.
b : float
Coefficient for t.
c : float
Free term. Returned only if `s0` is provided.
"""
v = J.dot(s)
a = np.dot(v, v)
if diag is not None:
a += np.dot(s * diag, s)
a *= 0.5
b = np.dot(g, s)
if s0 is not None:
u = J.dot(s0)
b += np.dot(u, v)
c = 0.5 * np.dot(u, u) + np.dot(g, s0)
if diag is not None:
b += np.dot(s0 * diag, s)
c += 0.5 * np.dot(s0 * diag, s0)
return a, b, c
else:
return a, b
def minimize_quadratic_1d(a, b, lb, ub, c=0):
"""Minimize a 1-d quadratic function subject to bounds.
The free term `c` is 0 by default. Bounds must be finite.
Returns
-------
t : float
Minimum point.
y : float
Minimum value.
"""
t = [lb, ub]
if a != 0:
extremum = -0.5 * b / a
if lb < extremum < ub:
t.append(extremum)
t = np.asarray(t)
y = a * t**2 + b * t + c
min_index = np.argmin(y)
return t[min_index], y[min_index]
def evaluate_quadratic(J, g, s, diag=None):
"""Compute values of a quadratic function arising in least squares.
The function is 0.5 * s.T * (J.T * J + diag) * s + g.T * s.
Parameters
----------
J : ndarray, sparse matrix or LinearOperator, shape (m, n)
Jacobian matrix, affects the quadratic term.
g : ndarray, shape (n,)
Gradient, defines the linear term.
s : ndarray, shape (k, n) or (n,)
Array containing steps as rows.
diag : ndarray, shape (n,), optional
Addition diagonal part, affects the quadratic term.
If None, assumed to be 0.
Returns
-------
values : ndarray with shape (k,) or float
Values of the function. If `s` was 2-dimensional then ndarray is
returned, otherwise float is returned.
"""
if s.ndim == 1:
Js = J.dot(s)
q = np.dot(Js, Js)
if diag is not None:
q += np.dot(s * diag, s)
else:
Js = J.dot(s.T)
q = np.sum(Js**2, axis=0)
if diag is not None:
q += np.sum(diag * s**2, axis=1)
l = np.dot(s, g)
return 0.5 * q + l
# Utility functions to work with bound constraints.
def in_bounds(x, lb, ub):
"""Check if a point lies within bounds."""
return np.all((x >= lb) & (x <= ub))
def step_size_to_bound(x, s, lb, ub):
"""Compute a min_step size required to reach a bound.
The function computes a positive scalar t, such that x + s * t is on
the bound.
Returns
-------
step : float
Computed step. Non-negative value.
hits : ndarray of int with shape of x
Each element indicates whether a corresponding variable reaches the
bound:
* 0 - the bound was not hit.
* -1 - the lower bound was hit.
* 1 - the upper bound was hit.
"""
non_zero = np.nonzero(s)
s_non_zero = s[non_zero]
steps = np.empty_like(x)
steps.fill(np.inf)
with np.errstate(over='ignore'):
steps[non_zero] = np.maximum((lb - x)[non_zero] / s_non_zero,
(ub - x)[non_zero] / s_non_zero)
min_step = np.min(steps)
return min_step, np.equal(steps, min_step) * np.sign(s).astype(int)
def find_active_constraints(x, lb, ub, rtol=1e-10):
"""Determine which constraints are active in a given point.
The threshold is computed using `rtol` and the absolute value of the
closest bound.
Returns
-------
active : ndarray of int with shape of x
Each component shows whether the corresponding constraint is active:
* 0 - a constraint is not active.
* -1 - a lower bound is active.
* 1 - a upper bound is active.
"""
active = np.zeros_like(x, dtype=int)
if rtol == 0:
active[x <= lb] = -1
active[x >= ub] = 1
return active
lower_dist = x - lb
upper_dist = ub - x
lower_threshold = rtol * np.maximum(1, np.abs(lb))
upper_threshold = rtol * np.maximum(1, np.abs(ub))
lower_active = (np.isfinite(lb) &
(lower_dist <= np.minimum(upper_dist, lower_threshold)))
active[lower_active] = -1
upper_active = (np.isfinite(ub) &
(upper_dist <= np.minimum(lower_dist, upper_threshold)))
active[upper_active] = 1
return active
def make_strictly_feasible(x, lb, ub, rstep=1e-10):
"""Shift a point to the interior of a feasible region.
Each element of the returned vector is at least at a relative distance
`rstep` from the closest bound. If ``rstep=0`` then `np.nextafter` is used.
"""
x_new = x.copy()
active = find_active_constraints(x, lb, ub, rstep)
lower_mask = np.equal(active, -1)
upper_mask = np.equal(active, 1)
if rstep == 0:
x_new[lower_mask] = np.nextafter(lb[lower_mask], ub[lower_mask])
x_new[upper_mask] = np.nextafter(ub[upper_mask], lb[upper_mask])
else:
x_new[lower_mask] = (lb[lower_mask] +
rstep * np.maximum(1, np.abs(lb[lower_mask])))
x_new[upper_mask] = (ub[upper_mask] -
rstep * np.maximum(1, np.abs(ub[upper_mask])))
tight_bounds = (x_new < lb) | (x_new > ub)
x_new[tight_bounds] = 0.5 * (lb[tight_bounds] + ub[tight_bounds])
return x_new
def CL_scaling_vector(x, g, lb, ub):
"""Compute Coleman-Li scaling vector and its derivatives.
Components of a vector v are defined as follows:
::
| ub[i] - x[i], if g[i] < 0 and ub[i] < np.inf
v[i] = | x[i] - lb[i], if g[i] > 0 and lb[i] > -np.inf
| 1, otherwise
According to this definition v[i] >= 0 for all i. It differs from the
definition in paper [1]_ (eq. (2.2)), where the absolute value of v is
used. Both definitions are equivalent down the line.
Derivatives of v with respect to x take value 1, -1 or 0 depending on a
case.
Returns
-------
v : ndarray with shape of x
Scaling vector.
dv : ndarray with shape of x
Derivatives of v[i] with respect to x[i], diagonal elements of v's
Jacobian.
References
----------
.. [1] M.A. Branch, T.F. Coleman, and Y. Li, "A Subspace, Interior,
and Conjugate Gradient Method for Large-Scale Bound-Constrained
Minimization Problems," SIAM Journal on Scientific Computing,
Vol. 21, Number 1, pp 1-23, 1999.
"""
v = np.ones_like(x)
dv = np.zeros_like(x)
mask = (g < 0) & np.isfinite(ub)
v[mask] = ub[mask] - x[mask]
dv[mask] = -1
mask = (g > 0) & np.isfinite(lb)
v[mask] = x[mask] - lb[mask]
dv[mask] = 1
return v, dv
def reflective_transformation(y, lb, ub):
"""Compute reflective transformation and its gradient."""
if in_bounds(y, lb, ub):
return y, np.ones_like(y)
lb_finite = np.isfinite(lb)
ub_finite = np.isfinite(ub)
x = y.copy()
g_negative = np.zeros_like(y, dtype=bool)
mask = lb_finite & ~ub_finite
x[mask] = np.maximum(y[mask], 2 * lb[mask] - y[mask])
g_negative[mask] = y[mask] < lb[mask]
mask = ~lb_finite & ub_finite
x[mask] = np.minimum(y[mask], 2 * ub[mask] - y[mask])
g_negative[mask] = y[mask] > ub[mask]
mask = lb_finite & ub_finite
d = ub - lb
t = np.remainder(y[mask] - lb[mask], 2 * d[mask])
x[mask] = lb[mask] + np.minimum(t, 2 * d[mask] - t)
g_negative[mask] = t > d[mask]
g = np.ones_like(y)
g[g_negative] = -1
return x, g
# Functions to display algorithm's progress.
def print_header_nonlinear():
print("{0:^15}{1:^15}{2:^15}{3:^15}{4:^15}{5:^15}"
.format("Iteration", "Total nfev", "Cost", "Cost reduction",
"Step norm", "Optimality"))
def print_iteration_nonlinear(iteration, nfev, cost, cost_reduction,
step_norm, optimality):
if cost_reduction is None:
cost_reduction = " " * 15
else:
cost_reduction = "{0:^15.2e}".format(cost_reduction)
if step_norm is None:
step_norm = " " * 15
else:
step_norm = "{0:^15.2e}".format(step_norm)
print("{0:^15}{1:^15}{2:^15.4e}{3}{4}{5:^15.2e}"
.format(iteration, nfev, cost, cost_reduction,
step_norm, optimality))
def print_header_linear():
print("{0:^15}{1:^15}{2:^15}{3:^15}{4:^15}"
.format("Iteration", "Cost", "Cost reduction", "Step norm",
"Optimality"))
def print_iteration_linear(iteration, cost, cost_reduction, step_norm,
optimality):
if cost_reduction is None:
cost_reduction = " " * 15
else:
cost_reduction = "{0:^15.2e}".format(cost_reduction)
if step_norm is None:
step_norm = " " * 15
else:
step_norm = "{0:^15.2e}".format(step_norm)
print("{0:^15}{1:^15.4e}{2}{3}{4:^15.2e}".format(
iteration, cost, cost_reduction, step_norm, optimality))
# Simple helper functions.
def compute_grad(J, f):
"""Compute gradient of the least-squares cost function."""
if isinstance(J, LinearOperator):
return J.rmatvec(f)
else:
return J.T.dot(f)
def compute_jac_scale(J, scale_inv_old=None):
"""Compute variables scale based on the Jacobian matrix."""
if issparse(J):
scale_inv = np.asarray(J.power(2).sum(axis=0)).ravel()**0.5
else:
scale_inv = np.sum(J**2, axis=0)**0.5
if scale_inv_old is None:
scale_inv[scale_inv == 0] = 1
else:
scale_inv = np.maximum(scale_inv, scale_inv_old)
return 1 / scale_inv, scale_inv
def left_multiplied_operator(J, d):
"""Return diag(d) J as LinearOperator."""
J = aslinearoperator(J)
def matvec(x):
return d * J.matvec(x)
def matmat(X):
return d[:, np.newaxis] * J.matmat(X)
def rmatvec(x):
return J.rmatvec(x.ravel() * d)
return LinearOperator(J.shape, matvec=matvec, matmat=matmat,
rmatvec=rmatvec)
def right_multiplied_operator(J, d):
"""Return J diag(d) as LinearOperator."""
J = aslinearoperator(J)
def matvec(x):
return J.matvec(np.ravel(x) * d)
def matmat(X):
return J.matmat(X * d[:, np.newaxis])
def rmatvec(x):
return d * J.rmatvec(x)
return LinearOperator(J.shape, matvec=matvec, matmat=matmat,
rmatvec=rmatvec)
def regularized_lsq_operator(J, diag):
"""Return a matrix arising in regularized least squares as LinearOperator.
The matrix is
[ J ]
[ D ]
where D is diagonal matrix with elements from `diag`.
"""
J = aslinearoperator(J)
m, n = J.shape
def matvec(x):
return np.hstack((J.matvec(x), diag * x))
def rmatvec(x):
x1 = x[:m]
x2 = x[m:]
return J.rmatvec(x1) + diag * x2
return LinearOperator((m + n, n), matvec=matvec, rmatvec=rmatvec)
def right_multiply(J, d, copy=True):
"""Compute J diag(d).
If `copy` is False, `J` is modified in place (unless being LinearOperator).
"""
if copy and not isinstance(J, LinearOperator):
J = J.copy()
if issparse(J):
J.data *= d.take(J.indices, mode='clip') # scikit-learn recipe.
elif isinstance(J, LinearOperator):
J = right_multiplied_operator(J, d)
else:
J *= d
return J
def left_multiply(J, d, copy=True):
"""Compute diag(d) J.
If `copy` is False, `J` is modified in place (unless being LinearOperator).
"""
if copy and not isinstance(J, LinearOperator):
J = J.copy()
if issparse(J):
J.data *= np.repeat(d, np.diff(J.indptr)) # scikit-learn recipe.
elif isinstance(J, LinearOperator):
J = left_multiplied_operator(J, d)
else:
J *= d[:, np.newaxis]
return J
def check_termination(dF, F, dx_norm, x_norm, ratio, ftol, xtol):
"""Check termination condition for nonlinear least squares."""
ftol_satisfied = dF < ftol * F and ratio > 0.25
xtol_satisfied = dx_norm < xtol * (xtol + x_norm)
if ftol_satisfied and xtol_satisfied:
return 4
elif ftol_satisfied:
return 2
elif xtol_satisfied:
return 3
else:
return None
def scale_for_robust_loss_function(J, f, rho):
"""Scale Jacobian and residuals for a robust loss function.
Arrays are modified in place.
"""
J_scale = rho[1] + 2 * rho[2] * f**2
J_scale[J_scale < EPS] = EPS
J_scale **= 0.5
f *= rho[1] / J_scale
return left_multiply(J, J_scale, copy=False), f