DeepLabCut · natsukium · Jul 20, 2026 · Jul 20, 2026 · Jul 20, 2026 · Jul 20, 2026
diff --git a/deeplabcut/core/crossvalutils.py b/deeplabcut/core/crossvalutils.py
@@ -30,6 +30,8 @@
 )
 from deeplabcut.utils import auxfun_multianimal, auxiliaryfunctions
 
+_trapz = getattr(np, "trapezoid", np.trapz)  # NumPy 2.0+ compat; drop once NumPy 1 unsupported
+
 
 def _set_up_evaluation(data):
     params = dict()
@@ -84,7 +86,13 @@ def find_closest_neighbors(query: np.ndarray, ref: np.ndarray, k: int = 3) -> np
     return neighbors
 
 
-def _calc_separability(vals_left, vals_right, n_bins=101, metric="jeffries", max_sensitivity=False):
+def calc_separability(
+    vals_left: np.ndarray,
+    vals_right: np.ndarray,
+    n_bins: int = 101,
+    metric: str = "jeffries",
+    max_sensitivity: bool = False,
+) -> tuple[float, float]:
     if metric not in ("jeffries", "auc"):
         raise ValueError("`metric` should be either 'jeffries' or 'auc'.")
 
@@ -97,7 +105,7 @@ def _calc_separability(vals_left, vals_right, n_bins=101, metric="jeffries", max
     if metric == "jeffries":
         sep = np.sqrt(2 * (1 - np.sum(np.sqrt(hist_left * hist_right))))  # Jeffries-Matusita distance
     else:
-        sep = np.trapz(np.cumsum(hist_left), tpr)
+        sep = _trapz(np.cumsum(hist_left), tpr)
     if max_sensitivity:
         threshold = bins[max(1, np.argmax(tpr > 0))]
     else:
@@ -344,7 +352,7 @@ def _get_n_best_paf_graphs(
         return ([existing_edges], dict(zip(existing_edges, [0] * len(existing_edges), strict=False)))
 
     scores, _ = zip(
-        *[_calc_separability(between_train[n], within_train[n], metric=metric) for n in existing_edges], strict=False
+        *[calc_separability(between_train[n], within_train[n], metric=metric) for n in existing_edges], strict=False
     )
 
     # Find minimal skeleton

diff --git a/deeplabcut/core/inferenceutils.py b/deeplabcut/core/inferenceutils.py
@@ -1085,7 +1085,7 @@ def find_outlier_assemblies(dict_of_assemblies, criterion="area", qs=(5, 95)):
             tuples.append((frame_ind, getattr(assembly, criterion)))
     frame_inds, vals = zip(*tuples, strict=False)
     vals = np.asarray(vals)
-    lo, up = np.percentile(vals, qs, interpolation="nearest")
+    lo, up = np.percentile(vals, qs, method="nearest")
     inds = np.flatnonzero((vals < lo) | (vals > up)).tolist()
     return list(set(frame_inds[i] for i in inds))
 

diff --git a/deeplabcut/core/trackingutils.py b/deeplabcut/core/trackingutils.py
@@ -539,7 +539,7 @@ def object_keypoint_similarity(x, y):
         xx = x[mask]
         yy = y[mask]
         dist = np.linalg.norm(xx - yy, axis=1)
-        scale = np.sqrt(np.product(np.ptp(yy, axis=0)))  # square root of bounding box area
+        scale = np.sqrt(np.prod(np.ptp(yy, axis=0)))  # square root of bounding box area
         oks = np.exp(-0.5 * (dist / (0.05 * scale)) ** 2)
         return np.mean(oks)
 

diff --git a/deeplabcut/pose_estimation_pytorch/apis/prune_paf_graph.py b/deeplabcut/pose_estimation_pytorch/apis/prune_paf_graph.py
@@ -23,7 +23,7 @@
 import deeplabcut.pose_estimation_pytorch.data as data
 import deeplabcut.pose_estimation_pytorch.models.predictors as predictors
 import deeplabcut.utils.auxiliaryfunctions as auxiliaryfunctions
-from deeplabcut.core.crossvalutils import find_closest_neighbors
+from deeplabcut.core.crossvalutils import calc_separability, find_closest_neighbors
 from deeplabcut.pose_estimation_pytorch.models import PoseModel
 from deeplabcut.pose_estimation_pytorch.models.predictors.paf_predictor import Graph
 
@@ -162,33 +162,6 @@ def benchmark_paf_graphs(
     return results
 
 
-def _calc_separability(
-    vals_left: np.ndarray,
-    vals_right: np.ndarray,
-    n_bins: int = 101,
-    metric: str = "jeffries",
-    max_sensitivity: bool = False,
-) -> tuple[float, float]:
-    if metric not in ("jeffries", "auc"):
-        raise ValueError("`metric` should be either 'jeffries' or 'auc'.")
-
-    bins = np.linspace(0, 1, n_bins)
-    hist_left = np.histogram(vals_left, bins=bins)[0]
-    hist_left = hist_left / hist_left.sum()
-    hist_right = np.histogram(vals_right, bins=bins)[0]
-    hist_right = hist_right / hist_right.sum()
-    tpr = np.cumsum(hist_right)
-    if metric == "jeffries":
-        sep = np.sqrt(2 * (1 - np.sum(np.sqrt(hist_left * hist_right))))  # Jeffries-Matusita distance
-    else:
-        sep = np.trapz(np.cumsum(hist_left), tpr)
-    if max_sensitivity:
-        threshold = bins[max(1, np.argmax(tpr > 0))]
-    else:
-        threshold = bins[np.argmin(1 - np.cumsum(hist_left) + tpr)]
-    return sep, threshold
-
-
 @torch.no_grad()
 def compute_within_between_paf_costs(
     model: PoseModel,
@@ -263,7 +236,7 @@ def get_n_best_paf_graphs(
     existing_edges = list(set(k for k, v in within_train.items() if v))
 
     scores, _ = zip(
-        *[_calc_separability(between_train[n], within_train[n], metric=metric) for n in existing_edges], strict=False
+        *[calc_separability(between_train[n], within_train[n], metric=metric) for n in existing_edges], strict=False
     )
 
     # Find minimal skeleton

diff --git a/deeplabcut/pose_estimation_tensorflow/core/predict_multianimal.py b/deeplabcut/pose_estimation_tensorflow/core/predict_multianimal.py
@@ -15,6 +15,8 @@
 from scipy.ndimage import measurements
 from skimage.feature import peak_local_max
 
+_trapz = getattr(np, "trapezoid", np.trapz)  # NumPy 2.0+ compat; drop once NumPy 1 unsupported
+
 
 def extract_cnn_output(outputs_np, cfg):
     """Extract locref, scmap and partaffinityfield from network."""
@@ -120,7 +122,7 @@ def compute_edge_costs(
         xy[..., 1],
         edge_inds.reshape((-1, 1)),
     ]
-    integ = np.trapz(y, xy[..., ::-1], axis=1)
+    integ = _trapz(y, xy[..., ::-1], axis=1)
     affinities = np.linalg.norm(integ, axis=1).astype(np.float32)
     # unit_vecs = vecs / lengths[:, np.newaxis]
     # affinities = np.squeeze(y @ np.expand_dims(unit_vecs, axis=2)).sum(axis=1)

diff --git a/deeplabcut/refine_training_dataset/stitch.py b/deeplabcut/refine_training_dataset/stitch.py
@@ -22,7 +22,6 @@
 import networkx as nx
 import numpy as np
 import pandas as pd
-import scipy.linalg.interpolative as sli
 from networkx.algorithms.flow import preflow_push
 from scipy.linalg import hankel
 from scipy.spatial.distance import directed_hausdorff
@@ -384,11 +383,11 @@ def estimate_rank(self, tol):
         4/sqrt(3)
         """
         mat = self.to_hankelet()
-        if np.any(mat):  # check that the matrix contains non-zero entries
+        if np.any(mat):
             # nrows, ncols = mat.shape
             # beta = nrows / ncols
             # omega = 0.56 * beta ** 3 - 0.95 * beta ** 2 + 1.82 * beta + 1.43
-            _, s, _ = sli.svd(mat, min(10, min(mat.shape)))
+            s = np.linalg.svd(mat, compute_uv=False)[:10]
         else:
             s = np.zeros(min(10, min(mat.shape)))
 

diff --git a/deeplabcut/utils/plotting.py b/deeplabcut/utils/plotting.py
@@ -464,7 +464,7 @@ def plot_edge_affinity_distributions(
         i1, i2 = graph[ind]
         w_tr = w_train[ind]
         b_tr = b_train[ind]
-        sep, _ = crossvalutils._calc_separability(b_tr, w_tr, metric="auc")
+        sep, _ = crossvalutils.calc_separability(b_tr, w_tr, metric="auc")
         axes[n].text(
             0.5,
             0.8,

diff --git a/pyproject.toml b/pyproject.toml
@@ -41,7 +41,7 @@ dependencies = [
   "matplotlib>=3.3,<3.9,!=3.7,!=3.7.1",
   "networkx>=2.6",
   "numba>=0.54",
-  "numpy>=1.18.5,<2",
+  "numpy>=1.22.4,<2",
   "packaging>=26",
   # Migration to pandas 3.0 is tracked in https://github.com/DeepLabCut/DeepLabCut/issues/3362.
   "pandas[hdf5,performance]>=2.2,<3",

diff --git a/tests/test_inferenceutils.py b/tests/test_inferenceutils.py
@@ -215,7 +215,15 @@ def test_assembler_with_unique_bodypart(real_assemblies_montblanc, test_data_dir
     assemblies_gt = np.concatenate(
         [ass.xy for assemblies in real_assemblies_montblanc[0].values() for ass in assemblies]
     )
-    np.testing.assert_equal(assemblies, assemblies_gt)
+
+    # Individuals are interchangeable in multi-animal assembly, so the order of
+    # equally-scored assemblies within a frame is arbitrary and depends on sort
+    # tie-breaking, which changed in NumPy 2.0. Compare the two point sets in a
+    # canonical, order-independent order (lexsort keeps NaN rows aligned).
+    def _canonical(a):
+        return a[np.lexsort((a[:, 1], a[:, 0]))]
+
+    np.testing.assert_equal(_canonical(assemblies), _canonical(assemblies_gt))
 
 
 def test_assembler_with_identity(tmpdir_factory, real_assemblies, test_data_dir):

diff --git a/tests/test_stitcher.py b/tests/test_stitcher.py
@@ -261,3 +261,29 @@ def weight_func(t1, t2):
     assert all(0.998 <= track.likelihood <= 1 for track in stitcher.tracks)
     tracks = sorted(stitcher.tracks, key=lambda t: t.identity)
     assert all(track.identity == i for i, track in enumerate(tracks))
+
+
+_n_frames = 20
+_t = np.linspace(0, 4 * np.pi, _n_frames)
+
+
+@pytest.mark.parametrize(
+    "signal, expected",
+    [
+        (np.sin(_t), 2),
+        (np.sin(_t) + np.sin(2 * _t), 4),
+    ],
+)
+def test_estimate_rank(signal, expected):
+    """A Hankelet of a pure frequency signal has rank 2 per complex exponential
+    (two real dimensions); a second distinct frequency doubles it to rank 4.
+    """
+    n_frames = len(signal)
+    data = np.zeros((n_frames, 2, 3))
+    data[:, 0, 0] = signal
+    data[:, 0, 1] = signal
+    data[:, 1, 0] = signal
+    data[:, 1, 1] = signal
+    data[:, :, 2] = 1.0
+    tracklet = Tracklet(data, np.arange(n_frames))
+    assert tracklet.estimate_rank(tol=0.01) == expected