Introduce a (barely) working InflationGrader

Added Distributor classes that calculate best cell distribution and an Approximator class that decides which Chops to choose to meet them;
rewrote SmoothGrader base the whole InflationGrader on those.
Fixed a bug in grading relations (border case with count=1)

To do:
- handle cases where both ends of a Wire are on walls;
- write tests!
- refactor: organize autograding source and its hierarchy and their respective tests
- get rid of repeated code (LayerStack in Params and Distributor)
- add direct imports
- test/use graders with some serious examples

Author: FranzBangar

examples/advanced/inflation_grader.py

@@ -12,13 +12,20 @@
 shape = cb.ExtrudedShape(base, 1)
 
 # turn one block around to test grader's skillz
-shape.grid[1][0].rotate(np.pi, [0, 0, 1])
+shape.grid[0][1].rotate(np.pi, [0, 0, 1])
 
 mesh.add(shape)
 
 mesh.set_default_patch("boundary", "patch")
-for i in (0, 1, 2):
+for i in (0, 2):
     shape.operations[i].set_patch("front", "walls")
+shape.operations[1].set_patch("back", "walls")
+
+for i in (3, 4, 5):
+    shape.operations[i].set_patch("back", "walls")
+
+for i in (0, 3):
+    shape.operations[i].set_patch("left", "walls")
 
 mesh.modify_patch("walls", "wall")
 
@@ -32,7 +39,7 @@
     vertex = list(finder.find_in_sphere(point))[0]
     vertex.translate([0, 0.8, 0])
 
-grader = InflationGrader(mesh, 1e-2, 0.1)
+grader = InflationGrader(mesh, 5e-3, 0.1)
 grader.grade(take="max")
 
 mesh.write(os.path.join("..", "case", "system", "blockMeshDict"), debug_path="debug.vtk")

src/classy_blocks/grading/autograding/params/approximator.py

@@ -0,0 +1,145 @@
+import dataclasses
+import itertools
+from typing import List, Sequence
+
+import numpy as np
+
+from classy_blocks.grading.chop import Chop
+from classy_blocks.types import FloatListType
+from classy_blocks.util.constants import TOL
+
+
+@dataclasses.dataclass
+class Piece:
+    coords: FloatListType
+
+    @property
+    def point_count(self) -> int:
+        return len(self.coords)
+
+    @property
+    def cell_count(self) -> int:
+        return self.point_count - 1
+
+    @property
+    def start_size(self) -> float:
+        return abs(self.coords[1] - self.coords[0])
+
+    @property
+    def end_size(self) -> float:
+        return abs(self.coords[-1] - self.coords[-2])
+
+    @property
+    def length(self) -> float:
+        return abs(self.coords[-1] - self.coords[0])
+
+    @property
+    def total_expansion(self) -> float:
+        return self.end_size / self.start_size
+
+    @property
+    def is_simple(self) -> bool:
+        """If all sizes are equal, this is a simple grading case"""
+        return abs(self.start_size - self.end_size) < TOL
+
+    def get_chop_coords(self) -> FloatListType:
+        """Calculates coordinates as will be produced by a Chop"""
+        if self.is_simple:
+            return np.linspace(self.coords[0], self.coords[-1], num=self.point_count)
+
+        sizes = np.geomspace(self.start_size, self.end_size, num=self.cell_count)
+        coords = np.ones(self.point_count) * self.coords[0]
+
+        add = np.cumsum(sizes)
+
+        if self.coords[-1] < self.coords[0]:
+            add *= -1
+
+        coords[1:] += add
+
+        return coords
+
+    def get_fitness(self) -> float:
+        differences = (self.coords - self.get_chop_coords()) ** 2
+        return float(np.sum(differences) / self.cell_count)
+
+    def get_chop(self) -> Chop:
+        return Chop(
+            length_ratio=self.length,
+            count=self.cell_count,
+            total_expansion=self.total_expansion,
+        )
+
+
+class Approximator:
+    """Takes a list of arbitrary cell sizes and creates Chop objects
+    that blockMesh will understand; Tries to minimize differences
+    between the rudimentary total expansion + count and actual
+    desired (optimized) cell sizes.
+
+    In a possible future scenario where a custom mesher would be employed,
+    actual cell sizes could be used without intervention of this object."""
+
+    def __init__(self, coords: FloatListType):
+        self.coords = coords
+        self.sizes = np.diff(self.coords)
+        self.ratios = np.roll(self.sizes, -1)[:-1] / self.sizes[:-1]
+        self.count = len(coords) - 1
+
+    @property
+    def length(self) -> float:
+        return abs(self.coords[-1] - self.coords[0])
+
+    @property
+    def is_simple(self) -> bool:
+        """Returns True if this is a simple one-chop equal-size scenario"""
+        return max(self.sizes) - min(self.sizes) < TOL
+
+    def get_pieces(self, indexes: List[int]) -> List[Piece]:
+        """Creates Piece objects between given indexes (adds 0 and -1 automatically);
+        does not check if count is smaller than length of indexes"""
+        indexes = [0, *indexes]
+
+        pieces = [Piece(self.coords[indexes[i] : indexes[i + 1] + 1]) for i in range(len(indexes) - 1)]
+        pieces.append(Piece(self.coords[indexes[-1] :]))
+
+        return pieces
+
+    def get_chops(self, number: int) -> List[Chop]:
+        if self.is_simple:
+            # don't bother
+            return [Chop(count=self.count)]
+
+        # limit number of chops so there's no zero-cell chops
+        number = min(number + 1, self.count - 1)
+
+        # brute force: choose the best combination of chops
+        # but don't overdo it; just try a couple of scenarios
+        refinement = 4
+        indexes = np.linspace(0, self.count + 1, num=number * refinement, dtype="int")[1:-1]
+
+        best_fit = 1e12
+        best_scenario: Sequence[int] = [0] * (number - 1)
+
+        combinations = itertools.combinations(indexes, r=number - 2)
+        for scenario in combinations:
+            try:
+                pieces = self.get_pieces(scenario)  # type:ignore
+                fit = sum(piece.get_fitness() for piece in pieces)
+                if fit < best_fit:
+                    best_fit = fit
+                    best_scenario = scenario
+            except (IndexError, ValueError):
+                # obviously not the best scenario, eh?
+                continue
+
+        pieces = self.get_pieces(best_scenario)  # type:ignore
+
+        chops = [piece.get_chop() for piece in pieces]
+
+        # normalize length ratios
+        length = self.length
+        for chop in chops:
+            chop.length_ratio = chop.length_ratio / length
+
+        return chops

src/classy_blocks/grading/autograding/params/distributor.py

@@ -6,7 +6,7 @@
 import scipy.interpolate
 import scipy.optimize
 
-from classy_blocks.grading.autograding.params.base import sum_length
+from classy_blocks.grading.autograding.params.approximator import Approximator
 from classy_blocks.grading.autograding.params.layers import InflationLayer
 from classy_blocks.grading.chop import Chop
 from classy_blocks.types import FloatListType
@@ -17,7 +17,7 @@
 class DistributorBase(abc.ABC):
     """Algorithm that creates chops from given count and sizes;
     first distributes cells using a predefined 'ideal' sizes and ratios,
-    then optimizes their individual size to get as close to that ideal as possible.
+    then optimizes their *individual* size to get as close to that ideal as possible.
 
     Then, when cells are placed, Chops are created so that they produce as similar
     cells to the calculated as possible. Since blockMesh only supports equal
@@ -32,10 +32,10 @@ class DistributorBase(abc.ABC):
 
     @staticmethod
     def get_actual_ratios(coords: FloatListType) -> FloatListType:
+        # TODO: repeated code within Approximator! un-repeat
         lengths = np.diff(coords)
-        return (lengths[:-1] / np.roll(lengths, -1)[:-1]) ** -1
+        return np.roll(lengths, -1)[:-1] / lengths[:-1]
 
-    @abc.abstractmethod
     def get_ideal_ratios(self) -> FloatListType:
         """Returns desired cell-to-cell ratios"""
         return np.ones(self.count + 1)
@@ -45,7 +45,9 @@ def get_ratio_weights(self) -> FloatListType:
         """Returns weights of cell ratios"""
 
     def get_raw_coords(self) -> FloatListType:
-        # 'count' denotes number of 'intervals' so there must be another point
+        # 'count' denotes number of 'intervals' so add one more point to get points;
+        # first and last cells are added artificially ('ghost cells') to calculate
+        # proper expansion ratios and sizes
         return np.concatenate(
             ([-self.size_before], np.linspace(0, self.length, num=self.count + 1), [self.length + self.size_after])
         )
@@ -60,69 +62,31 @@ def get_smooth_coords(self) -> FloatListType:
         def ratios(inner_coords):
             coords[2:-2] = inner_coords
 
-            difference = self.get_actual_ratios(coords) - self.get_ideal_ratios()
-            return difference * self.get_ratio_weights()
+            # to prevent 'flipping' over and producing zero or negative length, scale with e^-ratio
+            difference = -(self.get_ratio_weights() * (self.get_actual_ratios(coords) - self.get_ideal_ratios()))
+            return np.exp(difference) - 1
 
-        scale = min(self.size_before, self.size_after, self.length / self.count) / 10
-        _ = scipy.optimize.least_squares(ratios, coords[2:-2], method="lm", ftol=scale / 100, x_scale=scale)
+        scale = min(self.size_before, self.size_after, self.length / self.count) / 100
+        _ = scipy.optimize.least_squares(ratios, coords[2:-2], ftol=scale / 10, x_scale=scale)
 
-        return coords
+        # omit the 'ghost' cells
+        return coords[1:-1]
 
     @property
     def is_simple(self) -> bool:
         # Don't overdo basic, simple-graded blocks
         base_size = self.length / self.count
 
         # TODO: use a more relaxed criterion?
-        return base_size - self.size_before < TOL and base_size - self.size_after < TOL
+        return base_size - self.size_before < TOL and base_size - self.size_after < 10 * TOL
 
     def get_chops(self, pieces: int) -> List[Chop]:
-        if self.is_simple:
-            return [Chop(count=self.count)]
+        approximator = Approximator(self.get_smooth_coords())
+        return approximator.get_chops(pieces)
 
+    def get_last_size(self) -> float:
         coords = self.get_smooth_coords()
-        sizes = np.diff(coords)
-        ratios = self.get_actual_ratios(coords)
-
-        count = len(ratios) - 1
-
-        # create a piecewise linear function from a number of chosen indexes and their respective c2c ratios
-        # then optimize indexes to obtain best fit
-        # fratios = scipy.interpolate.interp1d(range(len(ratios)), ratios)
-
-        # def get_piecewise(indexes: List[int]) -> Callable:
-        #     values = np.take(ratios, indexes)
-        #     return scipy.interpolate.interp1d(indexes, values)
-
-        # def get_fitness(indexes: List[int]) -> float:
-        #     fitted = get_piecewise(indexes)(range(len(ratios)))
-
-        #     ss_tot = np.sum((ratios - np.mean(ratios)) ** 2)
-        #     ss_res = np.sum((ratios - fitted) ** 2)
-
-        #     return 1 - (ss_res / ss_tot)
-
-        # print(get_fitness([0, 9, 10, count]))
-        # print(get_fitness(np.linspace(0, count, num=pieces + 1, dtype=int)))
-
-        chops: List[Chop] = []
-        indexes = np.linspace(0, count, num=pieces + 1, dtype=int)
-
-        for i, index in enumerate(indexes[:-1]):
-            chop_ratios = ratios[index : indexes[i + 1]]
-            ratio = np.prod(chop_ratios)
-            chop_count = indexes[i + 1] - indexes[i]
-            avg_ratio = ratio ** (1 / chop_count)
-            length = sum_length(sizes[index], chop_count, avg_ratio)
-            chop = Chop(length_ratio=length, total_expansion=ratio, count=chop_count)
-            chops.append(chop)
-
-        # normalize length ratios
-        sum_ratio = sum([chop.length_ratio for chop in chops])
-        for chop in chops:
-            chop.length_ratio = chop.length_ratio / sum_ratio
-
-        return chops
+        return coords[-2] - coords[-3]
 
 
 @dataclasses.dataclass
@@ -132,8 +96,9 @@ def get_ideal_ratios(self):
         return super().get_ideal_ratios()
 
     def get_ratio_weights(self):
-        # Enforce stricter policy on size_before and size_after
         weights = np.ones(self.count + 1)
+        # Enforce stricter policy on the first few cells
+        # to match size_before and size_after
         for i in (0, 1, 2, 3):
             w = 2 ** (3 - i)
             weights[i] = w
@@ -146,26 +111,34 @@ def get_ratio_weights(self):
 class InflationDistributor(SmoothDistributor):
     c2c_expansion: float
     bl_thickness_factor: int
+    buffer_expansion: float
+    bulk_size: float
 
     @property
     def is_simple(self) -> bool:
         return False
 
     def get_ideal_ratios(self):
+        # TODO: combine this logic and LayerStack;
+        # possibly package all parameters into a separate dataclass
+        ratios = super().get_ideal_ratios()
+
         # Ideal growth ratio in boundary layer is user-specified c2c_expansion;
         inflation_layer = InflationLayer(self.size_before, self.c2c_expansion, self.bl_thickness_factor, 1e12)
         inflation_count = inflation_layer.count
-        print(f"Inflation count: {inflation_count}")
-
-        ratios = super().get_ideal_ratios()
 
         ratios[:inflation_count] = self.c2c_expansion
-        print(ratios)
+
+        # add a buffer layer if needed
+        last_inflation_size = inflation_layer.end_size
+        if self.bulk_size > self.buffer_expansion * last_inflation_size:
+            buffer_count = int(np.log(self.bulk_size / last_inflation_size) / np.log(self.buffer_expansion)) + 1
+            ratios[inflation_count : inflation_count + buffer_count] = self.buffer_expansion
 
         return ratios
 
     def get_ratio_weights(self):
-        return super().get_ratio_weights()
-
-    def _get_ratio_weights(self):
+        # using the same weights as in SmoothDistributor
+        # can trigger overflow warnings but doesn't produce
+        # better chops; thus, keep it simple
         return np.ones(self.count + 1)