Fix test fail with numpy 1.25 (#3354)

Author: wjsi

mars/learn/cluster/_k_means_init.py

@@ -12,6 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import inspect
+
 import numpy as np
 
 from ... import opcodes
@@ -212,12 +214,18 @@ def _kmeans_plusplus(*args, **kwargs):
         )
 
         with device(device_id):
+            arg_spec = inspect.getfullargspec(_kmeans_plusplus)
+            kw = {}
+            if "sample_weight" in arg_spec.args:
+                kw["sample_weight"] = np.ones(x.shape[0])
+
             ctx[op.outputs[0].key] = _kmeans_plusplus(
                 x,
                 op.n_clusters,
                 x_squared_norms=x_squared_norms,
                 random_state=op.state,
                 n_local_trials=op.n_local_trials,
+                **kw
             )[0]
 
 

mars/learn/neighbors/base.py

@@ -31,9 +31,15 @@
 from ._kneighbors_graph import KNeighborsGraph
 
 
+def _call_if_callable(metrics):
+    if callable(metrics):
+        return metrics()
+    return metrics
+
+
 VALID_METRICS = dict(
-    ball_tree=SklearnBallTree.valid_metrics,
-    kd_tree=SklearnKDTree.valid_metrics,
+    ball_tree=_call_if_callable(SklearnBallTree.valid_metrics),
+    kd_tree=_call_if_callable(SklearnKDTree.valid_metrics),
     # The following list comes from the
     # sklearn.metrics.pairwise doc string
     brute=(

mars/lib/sparse/tests/test_sparse.py

@@ -150,21 +150,21 @@ def test_sparse_divide():
     s1 = SparseNDArray(s1_data)
     s2 = SparseNDArray(s2_data)
 
-    assert_array_equal(s1 / s2, s1 / s2)
-    assert_array_equal(s1 / d1, s1 / d1)
-    assert_array_equal(d1 / s1, d1 / s1.toarray())
-    assert_array_equal(s1 / 2, s1 / 2)
-    assert_array_equal(2 / s1, 2 / s1.toarray())
+    assert_array_equal(s1 / s2, s1 / s2, almost=True)
+    assert_array_equal(s1 / d1, s1 / d1, almost=True)
+    assert_array_equal(d1 / s1, d1 / s1.toarray(), almost=True)
+    assert_array_equal(s1 / 2, s1 / 2, almost=True)
+    assert_array_equal(2 / s1, 2 / s1.toarray(), almost=True)
 
     # test sparse vector
     v = SparseNDArray(v1, shape=(3,))
-    assert_array_equal(v / v, v1_data / v1_data)
-    assert_array_equal(v / d1, v1_data / d1)
-    assert_array_equal(d1 / v, d1 / v1_data)
+    assert_array_equal(v / v, v1_data / v1_data, almost=True)
+    assert_array_equal(v / d1, v1_data / d1, almost=True)
+    assert_array_equal(d1 / v, d1 / v1_data, almost=True)
     r = sps.csr_matrix(((v1.data / 1), v1.indices, v1.indptr), v1.shape)
-    assert_array_equal(v / 1, r.toarray().reshape(3))
+    assert_array_equal(v / 1, r.toarray().reshape(3), almost=True)
     r = sps.csr_matrix(((1 / v1.data), v1.indices, v1.indptr), v1.shape)
-    assert_array_equal(1 / v, r.toarray().reshape(3))
+    assert_array_equal(1 / v, r.toarray().reshape(3), almost=True)
 
 
 def test_sparse_floor_divide():

mars/tensor/reduction/core.py

@@ -328,11 +328,17 @@ def execute_agg(cls, ctx, op):
         func_name = getattr(cls, "_func_name", None)
         reduce_func = getattr(xp, func_name)
         out = op.outputs[0]
+
+        def get_reduce_args(func):
+            if hasattr(func, "_implementation"):
+                func = func._implementation
+            return inspect.getfullargspec(func).args
+
         with device(device_id):
             if input_chunk.size == 0 and op.keepdims:
                 # input chunk is empty, when keepdims is True, return itself
                 ret = input_chunk
-            elif "dtype" in inspect.getfullargspec(reduce_func).args:
+            elif "dtype" in get_reduce_args(reduce_func):
                 ret = reduce_func(
                     input_chunk, axis=axis, dtype=op.dtype, keepdims=bool(op.keepdims)
                 )

mars/tensor/statistics/tests/test_statistics.py

@@ -16,10 +16,13 @@
 import pytest
 
 from ....core import tile
+from ....lib.version import Version
 from ...datasource import tensor, array
 from .. import digitize, histogram_bin_edges, quantile, percentile
 from ..quantile import INTERPOLATION_TYPES
 
+np_version = Version(np.__version__)
+
 
 def test_digitize():
     x = tensor(np.array([0.2, 6.4, 3.0, 1.6]), chunk_size=2)
@@ -70,7 +73,11 @@ def test_quantile():
     raw = np.random.rand(100)
     q = np.random.rand(10)
 
-    for dtype in [np.float32, np.int64, np.complex128]:
+    quantile_dtypes = [np.float32, np.int64]
+    if np_version < Version("1.25"):
+        quantile_dtypes.extend([np.complex128])
+
+    for dtype in quantile_dtypes:
         raw2 = raw.astype(dtype)
         a = tensor(raw2, chunk_size=100)
 
@@ -84,7 +91,7 @@ def test_quantile():
     raw = np.random.rand(20, 10)
     q = np.random.rand(10)
 
-    for dtype in [np.float32, np.int64, np.complex128]:
+    for dtype in quantile_dtypes:
         for axis in (None, 0, 1, [0, 1]):
             for interpolation in INTERPOLATION_TYPES:
                 for keepdims in [True, False]: