Merge pull request #4113 from samuelgarcia/isosplit

Implement isosplit6 into spikeinterface in pure numpy/numba

Author: yger

src/spikeinterface/extractors/phykilosortextractors.py

@@ -40,7 +40,7 @@ class BasePhyKilosortSortingExtractor(BaseSorting):
     The cluster_id column is used as the merge key to combine properties from multiple files.
     All loaded properties are added to the sorting extractor as unit properties, with some
     renamed for SpikeInterface conventions: 'group' becomes 'quality', 'cluster_id'
-    becomes 'original_cluster_id'. These properties can be accessed via ``sorting.get_property()`` 
+    becomes 'original_cluster_id'. These properties can be accessed via ``sorting.get_property()``
     function.
     """
 

src/spikeinterface/sorters/internal/tridesclous2.py

@@ -57,14 +57,22 @@ class Tridesclous2Sorter(ComponentsBasedSorter):
             #     "allow_single_cluster": True,
             #     "cluster_selection_method": "eom",
             # },
-            "clusterer": "isosplit6",
-            "clusterer_kwargs": {},
+            # "clusterer": "isosplit6",
+            # "clusterer_kwargs": {},
+            "clusterer": "isosplit",
+            "clusterer_kwargs": {
+                "n_init": 50,
+                "min_cluster_size": 10,
+                "max_iterations_per_pass": 500,
+                "isocut_threshold": 2.0,
+            },
             "do_merge": True,
             "merge_kwargs": {
                 "similarity_metric": "l1",
                 "num_shifts": 4,
                 "similarity_thresh": 0.75,
             },
+            "min_size_split": 25,
         },
         "templates": {
             "ms_before": 2.0,
@@ -98,7 +106,7 @@ class Tridesclous2Sorter(ComponentsBasedSorter):
 
     @classmethod
     def get_sorter_version(cls):
-        return "2025.07"
+        return "2025.08"
 
     @classmethod
     def _run_from_folder(cls, sorter_output_folder, params, verbose):
@@ -194,20 +202,11 @@ def _run_from_folder(cls, sorter_output_folder, params, verbose):
         clustering_kwargs["clustering"] = params["clustering"].copy()
 
         if clustering_kwargs["clustering"]["clusterer"] == "isosplit6":
-            # this is a patch to make the github CI because isosplit is installable on python > 3.11
             have_sisosplit6 = importlib.util.find_spec("isosplit6") is not None
             if not have_sisosplit6:
-                if verbose:
-                    print(
-                        "By default tridesclous2 need isosplit6 package for better reults please install it, automatically switch to hdbscan instead"
-                    )
-                clustering_kwargs["clustering"]["clusterer"] = "hdbscan"
-                clustering_kwargs["clustering"]["clusterer_kwargs"] = {
-                    "min_cluster_size": 10,
-                    "min_samples": 1,
-                    "allow_single_cluster": True,
-                    "cluster_selection_method": "eom",
-                }
+                raise ValueError(
+                    "You want to run tridesclous2 with the isosplit6 (the C++) implementation, but this is not installed, please `pip install isosplit6`"
+                )
 
         unit_ids, clustering_label, more_outs = find_cluster_from_peaks(
             recording, peaks, method="tdc-clustering", method_kwargs=clustering_kwargs, extra_outputs=True, **job_kwargs

src/spikeinterface/sortingcomponents/clustering/circus.py

@@ -31,7 +31,7 @@ class CircusClustering:
     """
 
     _default_params = {
-        "clusterer": "hdbscan",  # 'isosplit6', 'hdbscan', 'isocut5'
+        "clusterer": "hdbscan",  # 'isosplit6', 'hdbscan', 'isosplit'
         "clusterer_kwargs": {
             "min_cluster_size": 20,
             "cluster_selection_epsilon": 0.5,
@@ -77,8 +77,8 @@ def main_function(cls, recording, peaks, params, job_kwargs=dict()):
         assert clusterer in [
             "isosplit6",
             "hdbscan",
-            "isocut5",
-        ], "Circus clustering only supports isosplit6, isocut5 or hdbscan"
+            "isosplit",
+        ], "Circus clustering only supports isosplit6, isosplit or hdbscan"
         if clusterer in ["isosplit6", "hdbscan"]:
             have_dep = importlib.util.find_spec(clusterer) is not None
             if not have_dep:

src/spikeinterface/sortingcomponents/clustering/clustering_tools.py

@@ -682,10 +682,10 @@ def remove_duplicates_via_matching(
     return labels, new_labels
 
 
-def remove_duplicates_via_dip(wfs_arrays, peak_labels, dip_threshold=1, cosine_threshold=None):
+def remove_duplicates_via_dip(wfs_arrays, peak_labels, dip_threshold=2.0, cosine_threshold=None):
     import sklearn
 
-    from spikeinterface.sortingcomponents.clustering.isocut5 import isocut5
+    from spikeinterface.sortingcomponents.clustering.isosplit_isocut import isocut
 
     new_labels = peak_labels.copy()
 
@@ -756,7 +756,7 @@ def remove_duplicates_via_dip(wfs_arrays, peak_labels, dip_threshold=1, cosine_t
                                 v = t_i - t_j
                                 pr_i = np.dot(data_i, v)
                                 pr_j = np.dot(data_j, v)
-                                diptest, _ = isocut5(np.concatenate((pr_i, pr_j)))
+                                diptest, _ = isocut(np.concatenate((pr_i, pr_j)))
                                 diptests[i][j] = diptest
 
                             if diptest < min_dip: