Merge branch '114--learnerND-convex-hull' into 'master'

Resolve "(LearnerND) allow any convex hull as domain"

Closes #114

See merge request qt/adaptive!127

Author: jbweston

adaptive/learner/learnerND.py

@@ -107,9 +107,10 @@ class LearnerND(BaseLearner):
     func: callable
         The function to learn. Must take a tuple of N real
         parameters and return a real number or an arraylike of length M.
-    bounds : list of 2-tuples
+    bounds : list of 2-tuples or `scipy.spatial.ConvexHull`
         A list ``[(a_1, b_1), (a_2, b_2), ..., (a_n, b_n)]`` containing bounds,
         one pair per dimension.
+        Or a ConvexHull that defines the boundary of the domain.
     loss_per_simplex : callable, optional
         A function that returns the loss for a simplex.
         If not provided, then a default is used, which uses
@@ -150,14 +151,21 @@ class LearnerND(BaseLearner):
     """
 
     def __init__(self, func, bounds, loss_per_simplex=None):
-        self.ndim = len(bounds)
         self._vdim = None
         self.loss_per_simplex = loss_per_simplex or default_loss
-        self.bounds = tuple(tuple(map(float, b)) for b in bounds)
         self.data = OrderedDict()
         self.pending_points = set()
 
-        self._bounds_points = list(map(tuple, itertools.product(*bounds)))
+        if isinstance(bounds, scipy.spatial.ConvexHull):
+            hull_points = bounds.points[bounds.vertices]
+            self._bounds_points = sorted(list(map(tuple, hull_points)))
+            self._bbox = tuple(zip(hull_points.min(axis=0), hull_points.max(axis=0)))
+            self._interior = scipy.spatial.Delaunay(self._bounds_points)
+        else:
+            self._bounds_points = sorted(list(map(tuple, itertools.product(*bounds))))
+            self._bbox = tuple(tuple(map(float, b)) for b in bounds)
+
+        self.ndim = len(self._bbox)
 
         self.function = func
         self._tri = None
@@ -169,7 +177,7 @@ def __init__(self, func, bounds, loss_per_simplex=None):
         self._subtriangulations = dict()  # simplex → triangulation
 
         # scale to unit
-        self._transform = np.linalg.inv(np.diag(np.diff(bounds).flat))
+        self._transform = np.linalg.inv(np.diag(np.diff(self._bbox).flat))
 
         # create a private random number generator with fixed seed
         self._random = random.Random(1)
@@ -275,7 +283,12 @@ def _simplex_exists(self, simplex):
 
     def inside_bounds(self, point):
         """Check whether a point is inside the bounds."""
-        return all(mn <= p <= mx for p, (mn, mx) in zip(point, self.bounds))
+        if hasattr(self, '_interior'):
+            return self._interior.find_simplex(point, tol=1e-8) >= 0
+        else:
+            eps = 1e-8
+            return all((mn - eps) <= p <= (mx + eps) for p, (mn, mx)
+                       in zip(point, self._bbox))
 
     def tell_pending(self, point, *, simplex=None):
         point = tuple(point)
@@ -349,11 +362,13 @@ def _ask_point_without_known_simplices(self):
         assert not self._bounds_available
         # pick a random point inside the bounds
         # XXX: change this into picking a point based on volume loss
-        a = np.diff(self.bounds).flat
-        b = np.array(self.bounds)[:, 0]
-        r = np.array([self._random.random() for _ in range(self.ndim)])
-        p = r * a + b
-        p = tuple(p)
+        a = np.diff(self._bbox).flat
+        b = np.array(self._bbox)[:, 0]
+        p = None
+        while p is None or not self.inside_bounds(p):
+            r = np.array([self._random.random() for _ in range(self.ndim)])
+            p = r * a + b
+            p = tuple(p)
 
         self.tell_pending(p)
         return p, np.inf
@@ -489,10 +504,10 @@ def plot(self, n=None, tri_alpha=0):
         if self.vdim > 1:
             raise NotImplementedError('holoviews currently does not support',
                                       '3D surface plots in bokeh.')
-        if len(self.bounds) != 2:
+        if len(self.ndim) != 2:
             raise NotImplementedError("Only 2D plots are implemented: You can "
                                       "plot a 2D slice with 'plot_slice'.")
-        x, y = self.bounds
+        x, y = self._bbox
         lbrt = x[0], y[0], x[1], y[1]
 
         if len(self.data) >= 4:
@@ -549,7 +564,7 @@ def plot_slice(self, cut_mapping, n=None):
                 raise NotImplementedError('multidimensional output not yet'
                                           ' supported by `plot_slice`')
             n = n or 201
-            values = [cut_mapping.get(i, np.linspace(*self.bounds[i], n))
+            values = [cut_mapping.get(i, np.linspace(*self._bbox[i], n))
                       for i in range(self.ndim)]
             ind = next(i for i in range(self.ndim) if i not in cut_mapping)
             x = values[ind]
@@ -574,9 +589,9 @@ def plot_slice(self, cut_mapping, n=None):
             xys = [xs[:, None], ys[None, :]]
             values = [cut_mapping[i] if i in cut_mapping
                       else xys.pop(0) * (b[1] - b[0]) + b[0]
-                      for i, b in enumerate(self.bounds)]
+                      for i, b in enumerate(self._bbox)]
 
-            lbrt = [b for i, b in enumerate(self.bounds)
+            lbrt = [b for i, b in enumerate(self._bbox)
                     if i not in cut_mapping]
             lbrt = np.reshape(lbrt, (2, 2)).T.flatten().tolist()
 

adaptive/tests/test_learnernd.py

@@ -1,8 +1,9 @@
 # -*- coding: utf-8 -*-
 
 from ..learner import LearnerND
-from ..runner import replay_log
-
+from ..runner import replay_log, simple
+from .test_learners import ring_of_fire, generate_random_parametrization
+import scipy.spatial
 
 def test_faiure_case_LearnerND():
     log = [
@@ -21,3 +22,15 @@ def test_faiure_case_LearnerND():
     ]
     learner = LearnerND(lambda *x: x, bounds=[(-1, 1), (-1, 1), (-1, 1)])
     replay_log(learner, log)
+
+def test_interior_vs_bbox_gives_same_result():
+    f = generate_random_parametrization(ring_of_fire)
+    
+    control = LearnerND(f, bounds=[(-1, 1), (-1, 1)])
+    hull = scipy.spatial.ConvexHull(control._bounds_points)
+    learner = LearnerND(f, bounds=hull)
+
+    simple(control, goal=lambda l: l.loss() < 0.1)
+    simple(learner, goal=lambda l: l.loss() < 0.1)
+
+    assert learner.data == control.data

docs/source/tutorial/tutorial.LearnerND.rst

@@ -90,3 +90,37 @@ is a result of the fact that the learner chooses points in 3 dimensions
 and the simplices are not in the same face as we try to interpolate our
 lines. However, as always, when you sample more points the graph will
 become gradually smoother.
+
+Using any convex shape as domain
+--------------------------------
+
+Suppose you do not simply want to sample your function on a square (in 2D) or in
+a cube (in 3D). The LearnerND supports using a `scipy.spatial.ConvexHull` as
+your domain. This is best illustrated in the following example.
+
+Suppose you would like to sample you function in a cube split in half diagonally.
+You could use the following code as an example:
+
+.. jupyter-execute::
+
+    import scipy
+
+    def f(xyz):
+        x, y, z = xyz
+        return x**4 + y**4 + z**4 - (x**2+y**2+z**2)**2
+
+    # set the bound points, you can change this to be any shape
+    b = [(-1, -1, -1),
+         (-1,  1, -1),
+         (-1, -1,  1),
+         (-1,  1,  1),
+         ( 1,  1, -1),
+         ( 1, -1, -1)]
+
+    # you have to convert the points into a scipy.spatial.ConvexHull
+    hull = scipy.spatial.ConvexHull(b)
+
+    learner = adaptive.LearnerND(f, hull)
+    adaptive.BlockingRunner(learner, goal=lambda l: l.npoints > 2000)
+
+    learner.plot_isosurface(-0.5)