optimized transformation form tensor to numpy (#26447)

* optimized transformation form tensor to numpy, test=develop

* optimized transformation form tensor to numpy, pass pre-commit, test=develop

* modify fetchophandle zerocopy to deepcopy in PE&CUP, test=develop

* modify py:array construct, test=develop

* fix _fetch_var to use deep copy, test=develop

paddle/fluid/framework/details/fetch_op_handle.cc

@@ -117,7 +117,7 @@ static void TransData(const framework::LoDTensor &src_item,
       TensorCopy(src_item, platform::CPUPlace(), dst_item);
 #endif
     } else {
-      dst_item->ShareDataWith(src_item);
+      TensorCopy(src_item, platform::CPUPlace(), dst_item);
     }
   } else {
     dst_item->clear();

paddle/fluid/pybind/tensor_py.h

@@ -19,6 +19,7 @@ limitations under the License. */
 #include <memory>
 #include <string>
 #include <tuple>
+#include <utility>
 #include <vector>
 #include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/framework/lod_tensor.h"
@@ -564,9 +565,9 @@ inline py::array TensorToPyArray(const framework::Tensor &tensor,
 
   if (!is_gpu_tensor && !is_xpu_tensor) {
     if (!need_deep_copy) {
-      return py::array(py::buffer_info(
-          const_cast<void *>(tensor_buf_ptr), sizeof_dtype, py_dtype_str,
-          static_cast<size_t>(tensor.dims().size()), py_dims, py_strides));
+      auto base = py::cast(std::move(tensor));
+      return py::array(py::dtype(py_dtype_str.c_str()), py_dims, py_strides,
+                       const_cast<void *>(tensor_buf_ptr), base);
     } else {
       py::array py_arr(py::dtype(py_dtype_str.c_str()), py_dims, py_strides);
       PADDLE_ENFORCE_EQ(

python/paddle/fluid/executor.py

@@ -110,7 +110,7 @@ def scope_guard(scope):
         _switch_scope(ex)
 
 
-def as_numpy(tensor):
+def as_numpy(tensor, copy=False):
     """
     Convert a Tensor to a numpy.ndarray, its only support Tensor without LoD information.
     For higher dimensional sequence data, please use LoDTensor directly.
@@ -129,6 +129,7 @@ def as_numpy(tensor):
 
     Args:
        tensor(Variable): a instance of Tensor
+       copy(bool, optional): Whether to use deep copy.
 
     Returns:
         numpy.ndarray
@@ -145,7 +146,10 @@ def as_numpy(tensor):
             Please set the parameter 'return_numpy' as 'False' to \
             return LoDTensor itself directly.")
     if tensor._is_initialized():
-        return np.array(tensor)
+        if copy:
+            return np.array(tensor)
+        else:
+            return np.asarray(tensor)
     else:
         return None
 
@@ -350,7 +354,7 @@ def _fetch_var(name, scope=None, return_numpy=True):
         " program.")
     tensor = var.get_tensor()
     if return_numpy:
-        tensor = as_numpy(tensor)
+        tensor = as_numpy(tensor, copy=True)
     return tensor