[Experimental] IVF index support

.gitignore

@@ -73,6 +73,8 @@ cluster-mol-ids*.json
 config*.json
 timings*.json
 max-rss.txt
+*.smi
+!chembl-33-natural-products-subset.smi
 
 # Pickle
 *.pkl

bblean/bitbirch.py

@@ -1482,9 +1482,12 @@ def global_clustering(
         if not self.is_init:
             raise ValueError("The model has not been fitted yet.")
         centroids = np.vstack(self.get_centroids(packed=False))
-        labels = self._centrals_global_clustering(
+        predictor = self._global_clustering_predictor(
             centroids, n_clusters, method=method, input_is_packed=False, **method_kwargs
         )
+        # Add 1 to start labels from 1 instead of 0, so 0 can be used as sentinel
+        # value
+        labels = predictor.fit_predict(centroids) + 1
         num_centroids = len(centroids)
         self._n_global_clusters = (
             n_clusters if num_centroids > n_clusters else num_centroids
@@ -1493,7 +1496,7 @@ def global_clustering(
         return self
 
     @staticmethod
-    def _centrals_global_clustering(
+    def _global_clustering_predictor(
         centrals: NDArray[np.uint8],
         n_clusters: int,
         *,
@@ -1502,7 +1505,7 @@ def _centrals_global_clustering(
         n_features: int | None = None,
         # TODO: Type correctly
         **method_kwargs: tp.Any,
-    ) -> NDArray[np.int64]:
+    ) -> tp.Any:
         r""":meta private:"""
         if method not in {"agglomerative", "kmeans", "kmeans-normalized"}:
             raise ValueError(f"Unknown method {method}")
@@ -1532,11 +1535,7 @@ def _centrals_global_clustering(
         else:
             raise ValueError("method must be one of 'kmeans' or 'agglomerative'")
 
-        # Add 1 to start labels from 1 instead of 0, so 0 can be used as sentinel
-        # value
-        # This is the bottleneck for building this index
-        # K-means is feasible, agglomerative is extremely expensive
-        return predictor.fit_predict(centrals) + 1
+        return predictor
 
 
 # There are 4 cases here: