Merge pull request 'fix_load_master (#955)' (#38) from OpenI/MSAdapter:master into master

Reviewed-on: #38

mindtorch/torch/_utils.py

@@ -56,7 +56,7 @@ def _rebuild_tensor(storage, storage_offset, size, stride):
     unsupported_attr(stride)
     from mindtorch.torch.tensor import tensor # pylint: disable=R0401, C0415
     t = tensor([], dtype=storage.dtype, device=storage._untyped().device)
-    return t.set_(storage._untyped(), storage_offset, size)
+    return t._set_storage(storage._untyped(), storage_offset, size)
 
 
 def _rebuild_tensor_v2(storage, storage_offset, size, stride, requires_grad, backward_hooks, metadata=None):
@@ -82,7 +82,7 @@ def _rebuild_parameter(data, requires_grad, backward_hooks):
     unsupported_attr(backward_hooks)
     from mindtorch.torch.nn import Parameter # pylint: disable=R0401, C0415
     param = Parameter(data, requires_grad)
-    param.set_(data.storage()._untyped(), 0, data.size())
+    param._set_storage(data.storage()._untyped(), 0, data.size())
     return param
 
 def _rebuild_parameter_with_state(data, requires_grad, backward_hooks, state):

mindtorch/torch/storage.py

@@ -72,10 +72,12 @@ class _StorageBase():
         self._update_referenced_tensor()
         return self
 
-    def _update_referenced_tensor(self, strict=True, size=None):
+    def _update_referenced_tensor(self, strict=True, size=None, storage_offset=0):
         if self.referenced_tensor is not None:
             np_data = np.frombuffer(self.inner_data,
                                     _TypeDict.get(self.referenced_tensor.dtype))
+            if storage_offset:
+                np_data = np_data[storage_offset:]
             if size is not None:
                 np_data = np_data.reshape(size)
             if strict:
@@ -177,10 +179,12 @@ class _StorageBase():
         return False
 
     def _set_from_file(self, f, offset, is_real_file, element_size):
-        if not is_real_file:
-            raise RuntimeError("Currently, in `storage._set_from_file` only is_real_file==True supported.")
         nbytes = np.frombuffer(f.read(8), np.int64).item() * element_size
-        array = np.fromfile(f, dtype=np.uint8, count=nbytes, offset=offset)
+        if is_real_file:
+            array = np.fromfile(f, dtype=np.uint8, count=nbytes, offset=offset)
+        else:
+            f.seek(offset)
+            array = np.frombuffer(f.read(nbytes), np.uint8)
         self.inner_data[:] = array
         self._update_referenced_tensor()
         return self

mindtorch/torch/tensor.py

@@ -1659,7 +1659,6 @@ class Tensor(StubTensor, metaclass=_TensorMeta):
         if graph_mode_condition():
             warning('`Tensor.set_` is an in-place operation and "x.set_()" is not supported to use '
                                'in MindSpore static graph mode.')
-
         if isinstance(source, Tensor):
             if source.dtype != self.dtype:
                 raise RuntimeError("In `tensor.set_`, sourse.dtype must equal to self.dtype.")
@@ -1669,19 +1668,24 @@ class Tensor(StubTensor, metaclass=_TensorMeta):
             self.assign_value(source)
             return self
 
+        return self._set_storage(source=source, storage_offset=storage_offset, size=size, stride=stride)
+
+    def _set_storage(self,source=None, storage_offset=0, size=None, stride=None):
+        unsupported_attr(stride)
         if isinstance(source, _TypedStorage):
             # handle source is a _TypedStorage
             if source.dtype != self.dtype:
                 raise RuntimeError("In `tensor.set_`, _TypedStorage.dtype must equal to self.dtype.")
             source._storage.referenced_tensor = self
-            source._storage._update_referenced_tensor(strict=False, size=size)
+            source._storage._update_referenced_tensor(strict=False, size=size, storage_offset=storage_offset)
             return self
 
         # handle source is a _UntypedStorage
         source.referenced_tensor = self
-        source._update_referenced_tensor(strict=False, size=size)
+        source._update_referenced_tensor(strict=False, size=size,storage_offset=storage_offset)
         return self
 
+
     def to(self, *args, **kwargs):
         # TODO:
         # Note that this API requires the user to ensure the correctness of the input currently,