Resolve "(Learner1D) add possibility to use the direct neighbors in the loss"

Closes #119 (closed)

Currently works for one R^1 \to R^1

Still have to make it work for R^1 \to R^N

Also performance is actually quite good. As in: the learner slows down about 1.5 times. Going (on my laptop) from 3 seconds per 1000 points to 4.5 seconds per 1000 points.

Which, I believe, will be more than compensated by the fact that the chosen points are generally better

adaptive/learner/learner1D.py

@@ -15,7 +15,66 @@ from ..notebook_integration import ensure_holoviews
 from ..utils import cache_latest
 
 
-def uniform_loss(interval, scale, function_values, neighbors):
+def use_nn_neighbors(n):
+    """Decorator to specify how many neighboring intervals the loss function uses.
+    
+    This decorator can wrap around a loss function to let `~adaptive.Learner1D` 
+    know that you would like to look at the N-nearest neighboring intervals.
+    
+    The loss function is then guaranteed to receive the data of at least the 
+    nn-neighbors and a dict that tells you what the neighboring points of these 
+    are. And the `~Learner1D` will then make sure that the loss is updated
+    whenever one of the nn-neighbours changes.
+
+    Examples
+    --------
+
+    This is a part of the curvature loss function
+
+    >>> @use_nn_neighbors(1)
+    ... def triangle_loss(interval, scale, data, neighbors):
+    ...     x_left, x_right = interval
+    ...     xs = [neighbors[x_left][0], x_left, x_right, neighbors[x_right][1]]
+    ...     # at the boundary, neighbours[<left boundary x>] is (None, <some other x>)
+    ...     xs = [x for x in xs if x is not None] 
+    ...     if len(xs) <= 2:
+    ...         return (x_right - x_left) / scale[0]
+    ...     
+    ...     y_scale = scale[1] or 1
+    ...     ys_scaled = [data[x] / y_scale for x in xs]
+    ...     xs_scaled = [x / scale[0] for x in xs]
+    ...     N = len(xs) - 2
+    ...     pts = [(x, y) for x, y in zip(xs_scaled, ys_scaled)]
+    ...     return sum(volume(pts[i:i+3]) for i in range(N)) / N
+
+    Or you may define a loss that favours the (local) minima of a function.
+
+    >>> @use_nn_neighbors(1)
+    ... def loss(interval, scale, data, neighbors):
+    ...     x_left, x_right = interval
+    ...     n_left = neighbors[x_left][0]
+    ...     n_right = neighbors[x_right][1]
+    ...     is_min = True
+    ...
+    ...     if n_left is not None and data[x_left] > data[n_left]:
+    ...         is_min = False
+    ...     if n_right is not None and data[x_right] > data[n_right]:
+    ...         is_min = False
+    ...     
+    ...     loss = (x_right - x_left) / scale[0]
+    ...     
+    ...     if is_min: 
+    ...         return loss * 100
+    ...     return loss
+    """
+    def _wrapped(loss_per_interval):
+        loss_per_interval.nn_neighbors = n
+        return loss_per_interval
+    return _wrapped
+
+
+@use_nn_neighbors(0)
+def uniform_loss(interval, scale, data, neighbors):
     """Loss function that samples the domain uniformly.
 
     Works with `~adaptive.Learner1D` only.
@@ -36,7 +95,8 @@ def uniform_loss(interval, scale, function_values, neighbors):
     return dx
 
 
-def default_loss(interval, scale, function_values, neighbors):
+@use_nn_neighbors(0)
+def default_loss(interval, scale, data, neighbors):
     """Calculate loss on a single interval.
 
     Currently returns the rescaled length of the interval. If one of the
@@ -44,7 +104,7 @@ def default_loss(interval, scale, function_values, neighbors):
     never touched. This behavior should be improved later.
     """
     x_left, x_right = interval
-    y_right, y_left = function_values[x_right], function_values[x_left]
+    y_right, y_left = data[x_right], data[x_left]
     x_scale, y_scale = scale
     dx = (x_right - x_left) / x_scale
     if y_scale == 0:
@@ -70,27 +130,26 @@ def _loss_of_multi_interval(xs, ys):
     return sum(vol(pts[i:i+3]) for i in range(N)) / N
 
 
-def triangle_loss(interval, scale, function_values, neighbors):
+@use_nn_neighbors(1)
+def triangle_loss(interval, scale, data, neighbors):
     x_left, x_right = interval
-    xs = [x_left, x_right]
-    if x_left in neighbors:
-        xs.insert(0, neighbors[x_left][1])
-    if x_right in neighbors:
-        xs.append(neighbors[x_right][0])
+    xs = [neighbors[x_left][0], x_left, x_right, neighbors[x_right][1]]
+    xs = [x for x in xs if x is not None]
 
     if len(xs) <= 2:
         return (x_right - x_left) / scale[0]
     else:
         y_scale = scale[1] or 1
-        ys_scaled = [function_values[x] / y_scale for x in xs]
+        ys_scaled = [data[x] / y_scale for x in xs]
         xs_scaled = [x / scale[0] for x in xs]
         return _loss_of_multi_interval(xs_scaled, ys_scaled)
 
 
 def get_curvature_loss(area_factor=1, euclid_factor=0.02, horizontal_factor=0.02):
-    def curvature_loss(interval, scale, function_values, neighbors):
-        triangle_loss_ = triangle_loss(interval, scale, function_values, neighbors)
-        default_loss_ = default_loss(interval, scale, function_values, neighbors)
+    @use_nn_neighbors(1)
+    def curvature_loss(interval, scale, data, neighbors):
+        triangle_loss_ = triangle_loss(interval, scale, data, neighbors)
+        default_loss_ = default_loss(interval, scale, data, neighbors)
         dx = (interval[1] - interval[0]) / scale[0]
         return (area_factor * (triangle_loss_**0.5)
                 + euclid_factor * default_loss_
@@ -121,11 +180,13 @@ def _get_neighbors_from_list(xs):
     return sortedcontainers.SortedDict(neighbors)
 
 
-def _get_interval(x, neighbors, nn_neighbors):
-    n = nn_neighbors
-    index = neighbors.bisect_left(x)
-    return [neighbors.iloc[i] for i in range(index - n - 1, index + 2 + n)
-            if 0 <= i < len(neighbors)]
+def _get_intervals(x, neighbors, nn_neighbors):
+    nn = nn_neighbors
+    i = neighbors.index(x)
+    start = max(0, i - nn - 1)
+    end = min(len(neighbors), i + nn + 2)
+    points = neighbors.keys()[start:end]
+    return list(zip(points, points[1:]))
 
 
 class Learner1D(BaseLearner):
@@ -142,10 +203,12 @@ class Learner1D(BaseLearner):
         A function that returns the loss for a single interval of the domain.
         If not provided, then a default is used, which uses the scaled distance
         in the x-y plane as the loss. See the notes for more details.
-    nn_neighbors : int, default: 0
+    nn_neighbors : int, optional, default: None
         The number of neighboring intervals that the loss function
         takes into account. If ``loss_per_interval`` doesn't use the neighbors
-        at all, then it should be 0.
+        at all, then it should be 0. By default we try to access the
+        ``loss_per_interval.nn_neighbors`` attribute which is set for all
+        implemented loss functions.
 
     Attributes
     ----------
@@ -164,21 +227,22 @@ class Learner1D(BaseLearner):
     scale : (float, float)
         The x and y scale over all the intervals, useful for rescaling the
         interval loss.
-    function_values : dict(float → float)
+    data : dict(float → float)
         A map containing evaluated function values. It is guaranteed
         to have values for both of the points in 'interval'.
     neighbors : dict(float → (float, float))
         A map containing points as keys to its neighbors as a tuple.
     """
 
-    def __init__(self, function, bounds, loss_per_interval=None, nn_neighbors=0):
+    def __init__(self, function, bounds, loss_per_interval=None):
         self.function = function
-        self.nn_neighbors = nn_neighbors
 
-        if nn_neighbors == 0:
-            self.loss_per_interval = loss_per_interval or default_loss
+        if hasattr(loss_per_interval, 'nn_neighbors'):
+            self.nn_neighbors = loss_per_interval.nn_neighbors
         else:
-            self.loss_per_interval = loss_per_interval or get_curvature_loss()
+            self.nn_neighbors = 0
+
+        self.loss_per_interval = loss_per_interval or default_loss
 
         # A dict storing the loss function for each interval x_n.
         self.losses = {}
@@ -243,9 +307,8 @@ class Learner1D(BaseLearner):
             return 0
 
         # we need to compute the loss for this interval
-        interval = (x_left, x_right)
         return self.loss_per_interval(
-            interval, self._scale, self.data, self.neighbors)
+            (x_left, x_right), self._scale, self.data, self.neighbors)
 
 
     def _update_interpolated_loss_in_interval(self, x_left, x_right):
@@ -281,8 +344,7 @@ class Learner1D(BaseLearner):
             # (x_left, x), (x, x_right) and the nn_neighbors nearest
             # neighboring intervals. Since the addition of the
             # point 'x' could change their loss.
-            points = _get_interval(x, self.neighbors, self.nn_neighbors)
-            for ival in zip(points, points[1:]):
+            for ival in _get_intervals(x, self.neighbors, self.nn_neighbors):
                 self._update_interpolated_loss_in_interval(*ival)
 
             # Since 'x' is in between (x_left, x_right),
@@ -429,10 +491,8 @@ class Learner1D(BaseLearner):
 
         # The the losses for the "real" intervals.
         self.losses = {}
-        for x_left, x_right in intervals:
-            self.losses[x_left, x_right] = (
-                self._get_loss_in_interval(x_left, x_right)
-                if x_right - x_left >= self._dx_eps else 0)
+        for ival in intervals:
+            self.losses[ival] = self._get_loss_in_interval(*ival)
 
         # List with "real" intervals that have interpolated intervals inside
         to_interpolate = []