update docs; clean propagate labels;

Author: JelmerBot

doc/detecting_branches.ipynb

@@ -195,7 +195,7 @@
    "metadata": {},
    "source": [
     "Using the branch detection functionality is fairly straightforward. Using\n",
-    "`fast_hdbscan` one can simply configure the ``BranchDetector`` class and fit is\n",
+    "`fast_hdbscan` one can simply configure the ``BranchDetector`` class and fit it\n",
     "with the `fast_hdbscan.HDBSCAN` object. By default `BranchDetector` uses the\n",
     "values of the given `HDBSCAN` object for the parameters they share.\n",
     "\n",
@@ -236,10 +236,9 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "By default, the centers of clusters get a non-noise label different from the\n",
-    "branches in the cluster. This behavior can be changed by setting the\n",
-    "`propagate_labels=True` parameter or by  calling `propagated_labels()` after\n",
-    "fitting."
+    "The centers of clusters get a non-noise label different from the branches in the\n",
+    "cluster. This behavior can be changed by setting the `propagate_labels=True`\n",
+    "parameter or by  calling `propagated_labels()` after fitting."
    ]
   },
   {
@@ -306,7 +305,12 @@
     "  default a cluster needs to have one bifurcation (Y-shape) before the detected\n",
     "  branches are represented in the final labelling.\n",
     "\n",
-    "Unlike `hdbscan.branches.BranchDetector`, the `fast_hdbscan` version does not support the `branch_detection_method` parameter. This implementation will always use a `\"core\"` graph to determine which points are connected within a cluster."
+    "Unlike the `hdbscan` version, `fast_hdbscan`'s `BranchDetector` does not support\n",
+    "the `branch_detection_method` parameter. This implementation will always use a\n",
+    "`\"core\"` graph to determine which points are connected within a cluster. A\n",
+    "`\"core\"` graph combines nearest neighbors and the minimum spanning tree of a\n",
+    "cluster. It contains all connectivity within the points' core distances and\n",
+    "forms a single connected component per cluster."
    ]
   },
   {
@@ -315,8 +319,8 @@
    "source": [
     "## Useful attributes\n",
     "\n",
-    "Like the HDBSCAN class, the BranchDetector class contains several useful\n",
-    "attributes for exploring datasets.\n",
+    "The `BranchDetector` class contains several useful attributes for exploring\n",
+    "datasets.\n",
     "\n",
     "### Branch hierarchy\n",
     "\n",

doc/for_developers.ipynb

@@ -77,6 +77,7 @@ User Guide
    benchmarks
    comparable_clusterings
    detecting_branches
+   for_developers
 
 
 ----------

doc/index.rst

@@ -142,12 +142,10 @@ def propagate_labels_per_cluster(graph, sub_labels):
                 undirected[idx][neighbor] = 1 / graph.weights[i]
 
     # repeat density-weighted majority votes on noise points until all are assigned
-    prev = 0
     while True:
         noise_idx = np.nonzero(sub_labels == -1)[0]
-        if noise_idx.shape[0] == prev:
+        if noise_idx.shape[0] == 0:
             break
-        prev = noise_idx.shape[0]
         for idx in noise_idx:
             candidates = {np.int64(0): np.float64(0.0) for _ in range(0)}
             for neighbor_idx, weight in undirected[idx].items():
@@ -165,7 +163,7 @@ def propagate_labels_per_cluster(graph, sub_labels):
                     max_weight = weight
                     max_candidate = candidate
             sub_labels[idx] = max_candidate
-    return sub_labels, prev
+    return sub_labels
 
 
 def propagate_sub_cluster_labels(labels, sub_labels, graph_list, points_list):
@@ -177,11 +175,9 @@ def propagate_sub_cluster_labels(labels, sub_labels, graph_list, points_list):
         unique_sub_labels = np.unique(sub_labels[points])
         has_noise = unique_sub_labels[0] == -1 and len(unique_sub_labels) > 1
         if has_noise:
-            sub_labels[points], remaining = propagate_labels_per_cluster(
+            sub_labels[points] = propagate_labels_per_cluster(
                 core_graph, sub_labels[points]
             )
-            if remaining > 0:
-                raise RuntimeError('Failed to propagate all labels in sub-cluster')
         labels[points] = sub_labels[points] + running_id
         running_id += len(unique_sub_labels) - int(has_noise)
 
@@ -488,7 +484,6 @@ def fit(
             self.sub_cluster_probabilities_,
             self._approximation_graphs,
             self._condensed_trees,
-
             self._linkage_trees,
             self._spanning_trees,
             self.lens_values_,