Source code for morphogen.optimization.filters

"""Regularization of density designs: filter, projection and symmetry.

* Density filter (Bruns and Tortorelli 2001, Bourdin 2001): weighted mean of the design
  over a ball of radius ``r`` with weights ``r - distance`` (linear hat), periodic. It
  sets a minimal length scale and removes checkerboards.
* Heaviside projection (Wang, Lazarov and Sigmund 2011) of the filtered densities:
  ``(tanh(beta eta) + tanh(beta (x - eta))) / (tanh(beta eta) + tanh(beta (1 - eta)))``,
  sharper as ``beta`` grows (continuation).
* Symmetry: projection on the designs invariant by the mirror planes of the cell
  (orthotropic), and by the permutations of the axes (cubic; square in 2D).

The three operators are linear (filter, symmetry) or diagonal (projection), with
their transposes applied to sensitivities.
"""

from __future__ import annotations

import itertools

import numpy as np

from morphogen.automata.lattice import kernel_sum, offsets

SYMMETRIES = ("none", "orthotropic", "cubic")


[docs] class DensityFilter: """Periodic linear-hat filter of radius ``radius`` voxels (planar for nz = 1).""" def __init__(self, shape, radius: float): self.shape = tuple(int(n) for n in shape) self.radius = float(radius) dims = 2 if self.shape[2] == 1 else 3 if self.radius < 1.0: self.offsets = np.zeros((1, 3), dtype=np.int64) self.weights = np.ones(1) else: self.offsets = offsets("euclidean", self.radius, dims=dims, include_center=True) dist = np.linalg.norm(self.offsets, axis=1) self.weights = np.maximum(0.0, self.radius - dist) self.weights = self.weights / self.weights.sum() def __call__(self, x: np.ndarray) -> np.ndarray: if len(self.weights) == 1: return np.asarray(x, dtype=float).copy() return kernel_sum( np.asarray(x, dtype=float).reshape(self.shape), self.offsets, self.weights )
[docs] def transpose(self, g: np.ndarray) -> np.ndarray: """Transpose of the filter (the kernel mirrored: the same kernel).""" if len(self.weights) == 1: return np.asarray(g, dtype=float).copy() return kernel_sum( np.asarray(g, dtype=float).reshape(self.shape), -self.offsets, self.weights )
[docs] def heaviside(x, beta: float, eta: float = 0.5): """Smoothed Heaviside projection (identity-like for small beta).""" x = np.asarray(x, dtype=float) if beta <= 0: return x.copy() den = np.tanh(beta * eta) + np.tanh(beta * (1.0 - eta)) return (np.tanh(beta * eta) + np.tanh(beta * (x - eta))) / den
[docs] def heaviside_derivative(x, beta: float, eta: float = 0.5): """The derivative of the smoothed Heaviside projection (:func:`heaviside`) with respect to the density, of sharpness ``beta`` and threshold ``eta``; 1 without projection (``beta`` ≤ 0).""" x = np.asarray(x, dtype=float) if beta <= 0: return np.ones_like(x) den = np.tanh(beta * eta) + np.tanh(beta * (1.0 - eta)) return beta * (1.0 - np.tanh(beta * (x - eta)) ** 2) / den
[docs] class Symmetry: """Projection on the designs invariant by a symmetry group of the cell.""" def __init__(self, shape, kind: str = "none"): if kind not in SYMMETRIES: raise ValueError(f"symmetry: one of {SYMMETRIES}") self.shape = tuple(int(n) for n in shape) self.kind = kind planar = self.shape[2] == 1 axes = (0, 1) if planar else (0, 1, 2) if kind == "cubic": n = {self.shape[a] for a in axes} if len(n) != 1: raise ValueError("cubic symmetry needs a cubic (square in 2D) grid") self.ops = [] if kind == "none": return perms = list(itertools.permutations(axes)) if kind == "cubic" else [axes] for perm in perms: for flips in itertools.product((False, True), repeat=len(axes)): self.ops.append((perm, flips)) def __call__(self, x: np.ndarray) -> np.ndarray: x = np.asarray(x, dtype=float).reshape(self.shape) if not self.ops: return x.copy() acc = np.zeros(self.shape) planar = self.shape[2] == 1 for perm, flips in self.ops: full = tuple(perm) + ((2,) if planar else ()) y = np.transpose(x, full) for a, f in enumerate(flips): if f: y = np.flip(y, axis=a) acc += y return acc / len(self.ops) transpose = __call__ # an orthogonal projection is self-adjoint