Fixed #328, by using the backported as_strided()

bridge/npbackend/bohrium/array_manipulation.py

@@ -7,6 +7,7 @@
 from . import bhary
 from . import _util
 from .bhary import fix_biclass_wrapper
+from . import numpy_backport
 
 
 @fix_biclass_wrapper
@@ -154,7 +155,7 @@ def diagonal(ary, offset=0, axis1=0, axis2=1):
     ary = ary[..., :diag_size, offset:(offset + diag_size)]
 
     ret_strides = ary.strides[:-2] + (ary.strides[-1] + ary.strides[-2],)
-    return numpy.lib.stride_tricks.as_strided(ary, shape=ret_shape, strides=ret_strides)
+    return numpy_backport.as_strided(ary, shape=ret_shape, strides=ret_strides)
 
 
 @fix_biclass_wrapper

bridge/npbackend/bohrium/bhary.pyx

@@ -28,6 +28,7 @@ If not, see <http://www.gnu.org/licenses/>.
 import sys
 from ._util import dtype_equal, dtype_support, dtype_in
 from . import target
+from . import numpy_backport
 import operator
 import functools
 import numpy_force as numpy
@@ -234,7 +235,7 @@ def get_bhc(ary):
             # All this is simply a hack to reinterpret 'ary' as a complex view of the 'base'
             offset = (get_cdata(ary) - get_cdata(base)) // base.itemsize
             cary = numpy.frombuffer(base, dtype=base.dtype, offset=offset * base.itemsize)
-            cary = numpy.lib.stride_tricks.as_strided(cary, ary.shape, ary.strides, subok=True)
+            cary = numpy_backport.as_strided(cary, ary.shape, ary.strides)
 
             # if the view/base offset is aligned with the complex dtype, we know that the
             # 'ary' is a view of the real part of 'base'

bridge/npbackend/bohrium/reorganization.py

@@ -4,14 +4,14 @@
 """
 import warnings
 import numpy_force as numpy
-from numpy.lib.stride_tricks import as_strided
 from . import bhary
 from ._util import is_scalar
 from .bhary import fix_biclass_wrapper, get_bhc
 from . import target
 from . import array_create
 from . import array_manipulation
 from . import ufuncs
+from . import numpy_backport
 
 
 @fix_biclass_wrapper
@@ -311,11 +311,11 @@ def put(a, ind, v, mode='raise'):
     if indexes.size > values.size:
         if values.size == 1:
             # When 'values' is a scalar, we can broadcast it to match 'indexes'
-            values = as_strided(values, shape=indexes.shape, strides=(0,))
+            values = numpy_backport.as_strided(values, shape=indexes.shape, strides=(0,))
         else:  # else we repeat 'values' enough times to be larger than 'indexes'
-            values = as_strided(values,
-                                shape=(indexes.size // values.size + 2, values.size),
-                                strides=(0, values.itemsize))
+            values = numpy_backport.as_strided(values,
+                                               shape=(indexes.size // values.size + 2, values.size),
+                                               strides=(0, values.itemsize))
             values = array_manipulation.flatten(values, always_copy=False)
 
     # When 'values' is too large, we simple cut the end off

bridge/npbackend/bohrium/signal.py

@@ -6,13 +6,13 @@
 
 """
 import numpy_force as numpy
-from numpy_force.lib.stride_tricks import as_strided
 from . import array_create
 from . import bhary
 from . import ufuncs
 from . import linalg
 from . import summations
 from . import _util
+from . import numpy_backport
 
 
 # 1d
@@ -34,8 +34,8 @@ def _correlate_and_convolve_body(vector, filter, d, mode):
     padded[0:filter.size - 1] = 0
     padded[filter.size - 1:vector.size + filter.size - 1] = vector
     padded[vector.size + filter.size - 1:] = 0
-    s = as_strided(padded, shape=(padded.shape[0] - filter.size + 1, filter.size),
-                   strides=(padded.strides[0], padded.strides[0]))
+    s = numpy_backport.as_strided(padded, shape=(padded.shape[0] - filter.size + 1, filter.size),
+                                  strides=(padded.strides[0], padded.strides[0]))
     result = linalg.dot(s, filter)
     if mode == 'same':
         return result[d:vector.size + d]

bridge/npbackend/bohrium/target/target_numpy.py

@@ -3,6 +3,7 @@
 """
 from .. import bhc
 from .._util import dtype_name
+from .. import numpy_backport
 import numpy as np
 import mmap
 import time
@@ -53,7 +54,7 @@ class View(interface.View):
     def __init__(self, ndim, start, shape, strides, base):
         super(View, self).__init__(ndim, start, shape, strides, base)
         buf = np.frombuffer(self.base.mmap, dtype=self.dtype, offset=self.start)
-        self.ndarray = np.lib.stride_tricks.as_strided(buf, shape, self.strides)
+        self.ndarray = numpy_backport.as_strided(buf, shape, self.strides)
 
 
 def views2numpy(views):