utilities: enhance method to recursively convert items to np.ndarray,…

… handling hard cases for np.array

psyneulink/globals/utilities.py

@@ -1063,4 +1063,15 @@ def convert_all_elements_to_np_array(arr):
         else:
             return arr
 
-    return np.asarray([convert_all_elements_to_np_array(x) for x in arr])
+    subarr = [convert_all_elements_to_np_array(x) for x in arr]
+    try:
+        return np.array(subarr)
+    except ValueError:
+        # numpy cannot easily create arrays with subarrays of certain dimensions, workaround here
+        # https://stackoverflow.com/q/26885508/3131666
+        len_subarr = len(subarr)
+        elementwise_subarr = np.empty(len_subarr, dtype=np.ndarray)
+        for i in range(len_subarr):
+            elementwise_subarr[i] = subarr[i]
+
+        return elementwise_subarr

tests/misc/test_utilities.py

@@ -0,0 +1,35 @@
+import collections
+import numpy as np
+import pytest
+
+from psyneulink.globals.utilities import convert_all_elements_to_np_array
+
+
+@pytest.mark.parametrize(
+    'arr, expected',
+    [
+        ([[0], [0, 0]], np.array([np.array([0]), np.array([0, 0])])),
+        # should test these but numpy cannot easily create an array from them
+        # [np.ones((2,2)), np.zeros((2,1))]
+        # [np.array([[0]]), np.array([[[ 1.,  1.,  1.], [ 1.,  1.,  1.]]])]
+
+    ]
+)
+def test_convert_all_elements_to_np_array(arr, expected):
+    converted = convert_all_elements_to_np_array(arr)
+
+    # no current numpy methods can test this
+    def check_equality_recursive(arr, expected):
+        if (
+            not isinstance(arr, collections.Iterable)
+            or (isinstance(arr, np.ndarray) and arr.ndim == 0)
+        ):
+            assert arr == expected
+        else:
+            assert isinstance(expected, type(arr))
+            assert len(arr) == len(expected)
+
+            for i in range(len(arr)):
+                check_equality_recursive(arr[i], expected[i])
+
+    check_equality_recursive(converted, expected)