Rely on __cuda_array_interface__ for tensor conversions

chainer_pytorch_migration/tensor.py

@@ -27,20 +27,7 @@ def asarray(tensor):
                 return cupy.ndarray(
                     tuple(tensor.shape),
                     dtype=to_numpy_dtype(tensor.dtype))
-            itemsize = tensor.element_size()
-            storage = tensor.storage()
-            memptr = cupy.cuda.MemoryPointer(
-                cupy.cuda.UnownedMemory(
-                    storage.data_ptr(), storage.size() * itemsize, tensor,
-                ),
-                tensor.storage_offset() * itemsize,
-            )
-            return cupy.ndarray(
-                tuple(tensor.shape),
-                dtype=to_numpy_dtype(tensor.dtype),
-                memptr=memptr,
-                strides=tuple(s * itemsize for s in tensor.stride()),
-            )
+            return cupy.array(tensor.detach(), copy=False)
     if dev_type == 'cpu':
         return tensor.detach().numpy()
     raise ValueError('tensor on device "{}" is not supported', dev_type)
@@ -59,7 +46,7 @@ def astensor(array):
         view are gone.
 
     Note:
-        If the array has negative strides, a copy is made
+        aaIf the array has negative strides, a copy is made
     """
     if array is None:
         raise TypeError('array cannot be None')
@@ -78,28 +65,14 @@ def astensor(array):
                 device=array.device.id
             )
         return torch.as_tensor(
-            _ArrayWithCudaArrayInterfaceHavingStrides(array),
+            array,
             device=array.device.id,
         )
     if isinstance(array, numpy.ndarray):
         return torch.from_numpy(array)
     raise TypeError('array of type {} is not supported'.format(type(array)))
 
 
-# Workaround to avoid a bug in converting cupy.ndarray to torch.Tensor via
-# __cuda_array_interface__. See: https://github.com/pytorch/pytorch/pull/24947
-class _ArrayWithCudaArrayInterfaceHavingStrides:
-
-    def __init__(self, array):
-        self._array = array
-
-    @property
-    def __cuda_array_interface__(self):
-        d = self._array.__cuda_array_interface__
-        d['strides'] = self._array.strides
-        return d
-
-
 def to_numpy_dtype(torch_dtype):
     """Convert PyTorch dtype to NumPy dtype.