def Lines(grid: np.ndarray, prior: Tuple[np.ndarray] = None) -> Formula:
"""Collection of lines
"""
lines = listmap(lambda c: lambda t: (c - t))(grid)
prior = ((np.zeros(len(grid)),
1e-6 * np.identity(len(grid))) if prior is None else prior)
return Formula(terms=[lines], prior=prior)
def ReLU(grid: np.ndarray, prior: Tuple[np.ndarray] = None) -> Formula:
"""Rectified linear unit shaped basis
"""
relus = listmap(lambda c: lambda t: (t > c) * (c - t))(grid[1:-1])
prior = ((np.zeros(len(grid) - 2),
1e-6 * np.identity(len(grid) - 2)) if prior is None else prior)
return Formula(terms=[relus], prior=prior)
def FlippedReLU(grid: np.ndarray, prior: Tuple[np.ndarray] = None) -> Formula:
"""Mirrored ReLU basis
"""
relus = listmap(lambda c: lambda t: (t < c) * (c - t))(grid[1:-1])
prior = ((np.zeros(len(grid) - 2),
1e-6 * np.identity(len(grid) - 2)) if prior is None else prior)
return Formula(terms=[relus], prior=prior)
def Kron(a, b) -> Formula:
"""Tensor product of two Formula terms
Non-commutative!
Let ``u, v`` be eigenvectors of matrices ``A, B``, respectively. Then
``u ⊗ v`` is an eigenvector of ``A ⊗ B`` and ``λμ`` is the corresponding
eigenvalue.
Parameters
----------
a : Formula
Left input
b : Formula
Right input
"""
# NOTE: This is somewhat experimental. The terms must correspond to
# "zero-mean" r.v.. Then Kronecker product of covariances
# corresponds to the product r.v. of independent r.v.'s.
# Check the formula of variance of product of independent r.v.'s.
# TODO / FIXME: Don't flatten a and b.
gen = (
# Careful! Must be same order as in a Kronecker product.
(f, g) for f in sum(a.terms, []) for g in sum(b.terms, []))
# Outer product of terms
basis = listmap(lambda funcs: lambda t: funcs[0](t) * funcs[1](t))(gen)
# Kronecker product of prior means and covariances
return Formula(
terms=[basis],
prior=(
np.kron(a.prior[0], b.prior[0]),
# Although we kron-multiply precision matrices here (inverse
# of covariance), the order of inputs doesn't flip because
# (A ⊗ B) ^ -1 = (A ^ -1) ⊗ (B ^ -1)
np.kron(a.prior[1], b.prior[1])))
def BSpline1d(grid,
order=3,
extrapolate=True,
prior=None,
mu_basis=None,
mu_hyper=None) -> Formula:
"""B-spline basis on a fixed one-dimensional grid
Number of spline basis functions is always ``N = len(grid) + order - 2``
TODO: Verify that this doesn't break when scaling the grid
(extrapolation + damping)
Parameters
----------
grid : np.ndarray
Discretization grid
order : int
Order of the spline function. Polynomial degree is ``order - 1``
extrapolate : bool
Extrapolate outside of the grid using basis functions "touching" the
endpoints
prior : Tuple[np.ndarray]
Prior mean and precision matrix
mu_basis : List[Callable]
Basis for estimating the mean hyperparameter
mu_hyper : Tuple[np.ndarray]
Hyperprior mean and precision matrix
"""
mu_basis = [] if mu_basis is None else mu_basis
grid_ext = utils.extend_spline_grid(grid, order)
def build_basis_element(spline_arg):
(knots, extrapolate, loc) = spline_arg
def right_damp(t):
return t > knots[-1]
def left_damp(t):
return knots[0] > t
def element(t):
sp_element = interpolate.BSpline.basis_element(
knots, extrapolate if loc in (-1, 1) else False)
return sp_element(t) if loc == 0 else (
sp_element(t) * right_damp(t) if loc == -1 else sp_element(t) *
left_damp(t))
return utils.compose2(np.nan_to_num, element)
basis = listmap(build_basis_element)(utils.gen_spline_args_from_grid_ext(
grid_ext, order, extrapolate))
# Default prior is white noise
prior = ((np.zeros(len(basis)),
1e-6 * np.identity(len(basis))) if prior is None else prior)
return Formula(terms=[mu_basis + basis],
prior=prior if mu_hyper is None else utils.concat_gaussians(
[mu_hyper, prior]))
