[Zero-Dim] change Tensor.numpy() usage to other equivalent usage, avoid hack (#52197)

paddle/fluid/pybind/eager_method.cc

@@ -141,10 +141,11 @@ static PyObject* tensor_method_numpy(TensorObject* self,
              "order to avoid this problem, "
              "0D Tensor will be changed to 1D numpy currently, but it's not "
              "correct and will be "
-             "removed in future. Please modify "
-             " 'Tensor.numpy()[0]' to 'float(Tensor)' as soon as "
+             "removed in future. For Tensor contain only one element, Please "
+             "modify "
+             " 'Tensor.numpy()[0]' to 'Tensor.item()' as soon as "
              "possible, "
-             "otherwise 'Tensor.numpy()[0]' will raise error";
+             "otherwise 'Tensor.numpy()[0]' will raise error in future.";
       py_rank = 1;
       py_dims[0] = 1;
       py_strides[0] = sizeof_dtype * numel;

python/paddle/distributed/fleet/meta_parallel/pp_utils/p2p_communication.py

@@ -83,7 +83,7 @@ class SendRecvMeta:
         # recv stop_gradient
         stop_grad = paddle.to_tensor([0])
         paddle.distributed.recv(stop_grad, src=src_rank, group=group)
-        return shape.numpy().tolist(), dtype.item(), stop_grad.item()
+        return shape.tolist(), dtype.item(), stop_grad.item()
 
     def recv_meta(self, group):
         tensor_type = paddle.to_tensor([0])

python/paddle/fluid/dataloader/dataset.py

@@ -514,7 +514,7 @@ def random_split(dataset, lengths, generator=None):
         )
     # TODO(@Joejiong): support Variable or Tensor type with .tolist class member function.
     # For example var.item() and var.tolist()
-    indices = paddle.randperm(sum(lengths)).numpy().tolist()
+    indices = paddle.randperm(sum(lengths)).tolist()
     return [
         Subset(dataset, indices[offset - length : offset])
         for offset, length in zip(_accumulate(lengths), lengths)

python/paddle/fluid/dygraph/math_op_patch.py

@@ -140,21 +140,21 @@ def monkey_patch_math_varbase():
         ), "only one element variable can be converted to float."
         tensor = var.value().get_tensor()
         assert tensor._is_initialized(), "variable's tensor is not initialized"
-        return float(var.numpy().flatten()[0])
+        return float(var.item())
 
     def _long_(var):
         numel = np.prod(var.shape)
         assert numel == 1, "only one element variable can be converted to long."
         tensor = var.value().get_tensor()
         assert tensor._is_initialized(), "variable's tensor is not initialized"
-        return int(var.numpy().flatten()[0])
+        return int(var.item())
 
     def _int_(var):
         numel = np.prod(var.shape)
         assert numel == 1, "only one element variable can be converted to int."
         tensor = var.value().get_tensor()
         assert tensor._is_initialized(), "variable's tensor is not initialized"
-        return int(var.numpy().flatten()[0])
+        return int(var.item())
 
     def _len_(var):
         assert var.ndim > 0, "len() of a 0D tensor is wrong"
@@ -172,7 +172,7 @@ def monkey_patch_math_varbase():
         ), "only one element variable can be converted to python index."
         tensor = var.value().get_tensor()
         assert tensor._is_initialized(), "variable's tensor is not initialized"
-        return int(var.numpy().flatten()[0])
+        return int(var.item())
 
     @property
     def _ndim_(var):

python/paddle/fluid/dygraph/varbase_patch_methods.py

@@ -379,8 +379,8 @@ def monkey_patch_varbase():
             if self.grad is None:
                 return None
             if self.grad.is_selected_rows():
-                return (np.array(self.grad.numpy()), np.array(self.grad.rows()))
-            return self.grad.numpy()
+                return (np.array(self.grad), np.array(self.grad.rows()))
+            return np.array(self.grad)
         else:
             if self._grad_ivar() is None:
                 return None
@@ -735,11 +735,11 @@ def monkey_patch_varbase():
         ), "When Variable is used as the condition of if/while , Variable can only contain one element."
         if framework.global_var._in_eager_mode_:
             assert self._is_initialized(), "tensor not initialized"
-            return bool(np.all(self.numpy() > 0))
+            return bool(self.item() > 0)
         else:
             tensor = self.value().get_tensor()
             assert tensor._is_initialized(), "tensor not initialized"
-            return bool(np.all(tensor.__array__() > 0))
+            return bool(self.item() > 0)
 
     def __bool__(self):
         return self.__nonzero__()

python/paddle/fluid/layers/control_flow.py

@@ -1150,7 +1150,7 @@ def while_loop(cond, body, loop_vars, is_test=False, name=None):
         )
 
     if in_dygraph_mode():
-        now_cond = pre_cond.numpy().item()
+        now_cond = pre_cond.item()
         while now_cond:
             output_vars = body(*loop_vars)
             if not isinstance(output_vars, (list, tuple)):
@@ -1160,7 +1160,7 @@ def while_loop(cond, body, loop_vars, is_test=False, name=None):
                     "body in while_loop should return the same arity "
                     "(length and structure) and types as loop_vars"
                 )
-            now_cond = cond(*output_vars).numpy().item()
+            now_cond = cond(*output_vars).item()
             map_structure(assign_skip_lod_tensor_array, output_vars, loop_vars)
         return loop_vars
     else:

python/paddle/framework/io.py

@@ -63,7 +63,7 @@ def _build_saved_state_dict(state_dict):
                     raise ValueError(
                         "The saved tensor is not initialized. If you used group sharded, please use save_group_sharded_model."
                     )
-                save_dict[key] = value.numpy()
+                save_dict[key] = np.array(value)
             name_table[key] = value.name
         else:
             save_dict[key] = value
@@ -92,7 +92,7 @@ def _load_state_dict_from_save_inference_model(model_path, config):
         # 3. construct state_dict
         load_param_dict = {}
         for var_name in persistable_var_dict:
-            load_param_dict[var_name] = persistable_var_dict[var_name].numpy()
+            load_param_dict[var_name] = np.array(persistable_var_dict[var_name])
 
         # if *.info exists, we can recover structured_name
         var_info_filename = str(config.params_filename) + ".info"
@@ -146,7 +146,7 @@ def _load_state_dict_from_save_params(model_path):
     # 3. construct state_dict
     load_param_dict = {}
     for var in load_var_list:
-        load_param_dict[var.name] = var.numpy()
+        load_param_dict[var.name] = np.array(var)
 
     return load_param_dict
 
@@ -291,7 +291,7 @@ def _pickle_save(obj, f, protocol):
         )
 
     def reduce_varbase(self):
-        data = self.numpy()
+        data = np.array(self)
         name = self.name
 
         return (tuple, ((name, data),))

python/paddle/framework/io_utils.py

@@ -180,9 +180,7 @@ def _load_program_scope(main=None, startup=None, scope=None):
 @static_only
 def _legacy_static_save(param_dict, model_path, protocol=2):
     def get_tensor(var):
-        if isinstance(var, (core.VarBase, core.eager.Tensor)):
-            return var.numpy()
-        elif isinstance(var, core.LoDTensor):
+        if isinstance(var, (core.VarBase, core.eager.Tensor, core.LoDTensor)):
             return np.array(var)
         return var
 

python/paddle/hapi/model.py

@@ -61,7 +61,7 @@ def to_numpy(var):
         var, (Variable, fluid.core.VarBase, fluid.core.eager.Tensor)
     ), "not a variable"
     if isinstance(var, (fluid.core.VarBase, fluid.core.eager.Tensor)):
-        return var.numpy()
+        return np.array(var)
     t = global_scope().find_var(var.name).get_tensor()
     return np.array(t)
 

python/paddle/metric/metrics.py

@@ -292,7 +292,7 @@ class Accuracy(Metric):
             Tensor: the accuracy of current step.
         """
         if isinstance(correct, (paddle.Tensor, paddle.fluid.core.eager.Tensor)):
-            correct = correct.numpy()
+            correct = np.array(correct)
         num_samples = np.prod(np.array(correct.shape[:-1]))
         accs = []
         for i, k in enumerate(self.topk):
@@ -420,12 +420,12 @@ class Precision(Metric):
                 The data type is 'int32' or 'int64'.
         """
         if isinstance(preds, (paddle.Tensor, paddle.fluid.core.eager.Tensor)):
-            preds = preds.numpy()
+            preds = np.array(preds)
         elif not _is_numpy_(preds):
             raise ValueError("The 'preds' must be a numpy ndarray or Tensor.")
 
         if isinstance(labels, (paddle.Tensor, paddle.fluid.core.eager.Tensor)):
-            labels = labels.numpy()
+            labels = np.array(labels)
         elif not _is_numpy_(labels):
             raise ValueError("The 'labels' must be a numpy ndarray or Tensor.")
 
@@ -553,12 +553,12 @@ class Recall(Metric):
                 Shape: [batch_size, 1], Dtype: 'int32' or 'int64'.
         """
         if isinstance(preds, (paddle.Tensor, paddle.fluid.core.eager.Tensor)):
-            preds = preds.numpy()
+            preds = np.array(preds)
         elif not _is_numpy_(preds):
             raise ValueError("The 'preds' must be a numpy ndarray or Tensor.")
 
         if isinstance(labels, (paddle.Tensor, paddle.fluid.core.eager.Tensor)):
-            labels = labels.numpy()
+            labels = np.array(labels)
         elif not _is_numpy_(labels):
             raise ValueError("The 'labels' must be a numpy ndarray or Tensor.")
 
@@ -705,12 +705,12 @@ class Auc(Metric):
                 representing the label of the instance i.
         """
         if isinstance(labels, (paddle.Tensor, paddle.fluid.core.eager.Tensor)):
-            labels = labels.numpy()
+            labels = np.array(labels)
         elif not _is_numpy_(labels):
             raise ValueError("The 'labels' must be a numpy ndarray or Tensor.")
 
         if isinstance(preds, (paddle.Tensor, paddle.fluid.core.eager.Tensor)):
-            preds = preds.numpy()
+            preds = np.array(preds)
         elif not _is_numpy_(preds):
             raise ValueError("The 'preds' must be a numpy ndarray or Tensor.")
 

python/paddle/nn/decode.py

@@ -712,7 +712,7 @@ def _dynamic_decode_imperative(
 
     step_idx = 0
     step_idx_tensor = paddle.full(shape=[1], fill_value=step_idx, dtype="int64")
-    while cond.numpy():
+    while cond.item():
         (step_outputs, next_states, next_inputs, next_finished) = decoder.step(
             step_idx_tensor, inputs, states, **kwargs
         )

python/paddle/nn/functional/common.py

@@ -490,7 +490,7 @@ def interpolate(
         else:
             if in_dynamic_mode():
                 if isinstance(out_shape, Variable):
-                    out_shape = list(out_shape.numpy())
+                    out_shape = list(out_shape.numpy(False))
                 else:
                     out_shape = list(out_shape)
 

python/paddle/nn/functional/pooling.py

@@ -706,7 +706,7 @@ def _unpool_output_size(x, kernel_size, stride, padding, output_size):
         else:
             for i, var in enumerate(output_size):
                 if isinstance(var, Variable):
-                    output_size[i] = var.numpy().item()
+                    output_size[i] = var.item()
 
     if len(output_size) == len(kernel_size) + 2:
         output_size = output_size[2:]
@@ -1609,7 +1609,7 @@ def adaptive_avg_pool2d(x, output_size, data_format='NCHW', name=None):
 
     if in_dygraph_mode():
         output_size = [
-            item.numpy().item(0) if isinstance(item, Variable) else item
+            item.item(0) if isinstance(item, Variable) else item
             for item in output_size
         ]
     # output_size support Variable in static graph mode

python/paddle/optimizer/optimizer.py

@@ -382,9 +382,9 @@ class Optimizer:
                 load_para = state_dict[var_tmp.name]
 
                 if isinstance(load_para, Variable):
-                    load_para_np = load_para.numpy()
+                    load_para_np = np.array(load_para)
                 elif isinstance(load_para, core.VarBase):
-                    load_para_np = load_para.numpy()
+                    load_para_np = np.array(load_para)
                 elif isinstance(load_para, np.ndarray):
                     load_para_np = load_para
                 else:

python/paddle/quantization/imperative/ptq_quantizer.py

@@ -54,13 +54,13 @@ def combine_abs_max_and_hist(
         return origin_max, origin_hist
     elif origin_max == 0.0:
         new_hist, _ = np.histogram(
-            paddle.abs(tensor).numpy(), range=(0, new_max), bins=bins
+            paddle.abs(tensor).numpy(False), range=(0, new_max), bins=bins
         )
         new_hist = new_hist.astype(np.float32)
         return new_max, new_hist
     elif new_max <= origin_max:
         new_hist, _ = np.histogram(
-            paddle.abs(tensor).numpy(), range=(0, origin_max), bins=bins
+            paddle.abs(tensor).numpy(False), range=(0, origin_max), bins=bins
         )
         new_hist = new_hist.astype(np.float32)
         new_hist += origin_hist
@@ -84,7 +84,7 @@ def combine_abs_max_and_hist(
         sampled_hist = sampled_hist.astype(np.float32)
 
         new_hist, _ = np.histogram(
-            paddle.abs(tensor).numpy(), range=(0, new_max), bins=bins
+            paddle.abs(tensor).numpy(False), range=(0, new_max), bins=bins
         )
         new_hist = new_hist.astype(np.float32)
         new_hist += sampled_hist
@@ -189,7 +189,7 @@ class BaseHistQuantizer(BaseQuantizer, metaclass=abc.ABCMeta):
                     self.hists.append(None)
                 else:
                     hist, _ = np.histogram(
-                        paddle.abs(tensor).numpy(),
+                        paddle.abs(tensor).numpy(False),
                         range=(0.0, abs_max_vals[idx]),
                         bins=self.bins,
                     )

python/paddle/static/nn/metric.py

@@ -76,7 +76,7 @@ def accuracy(input, label, k=1, correct=None, total=None):
         if total is None:
             total = _varbase_creator(dtype="int32")
 
-        _k = k.numpy().item(0) if isinstance(k, Variable) else k
+        _k = k.item(0) if isinstance(k, Variable) else k
         topk_out, topk_indices = _legacy_C_ops.top_k_v2(
             input, 'k', _k, 'sorted', False
         )

python/paddle/tensor/array.py

@@ -119,7 +119,7 @@ def array_read(array, i):
         assert i.shape == [
             1
         ], "The shape of index 'i' should be [1] in dygraph mode"
-        i = i.numpy().item(0)
+        i = i.item(0)
         return array[i]
     else:
         check_variable_and_dtype(i, 'i', ['int64'], 'array_read')
@@ -179,7 +179,7 @@ def array_write(x, i, array=None):
         assert i.shape == [
             1
         ], "The shape of index 'i' should be [1] in dygraph mode"
-        i = i.numpy().item(0)
+        i = i.item(0)
         if array is None:
             array = create_array(x.dtype)
         assert isinstance(

python/paddle/tensor/manipulation.py

@@ -334,7 +334,7 @@ def slice(input, axes, starts, ends):
                 for item in starts
             ]
         elif isinstance(starts, tmp_tensor_type):
-            tensor_t = starts.numpy()
+            tensor_t = starts.numpy(False)
             starts = [ele for ele in tensor_t]
             infer_flags = [-1 for i in range(len(axes))]
 
@@ -344,7 +344,7 @@ def slice(input, axes, starts, ends):
                 for item in ends
             ]
         elif isinstance(ends, tmp_tensor_type):
-            tensor_t = ends.numpy()
+            tensor_t = ends.numpy(False)
             ends = [ele for ele in tensor_t]
             infer_flags = [-1 for i in range(len(axes))]
 

python/paddle/utils/layers_utils.py

@@ -456,12 +456,12 @@ def convert_shape_to_list(shape):
     if isinstance(shape, (list, tuple)):
         shape = list(
             map(
-                lambda x: x.numpy().flat[0] if isinstance(x, Variable) else x,
+                lambda x: x.item(0) if isinstance(x, Variable) else x,
                 shape,
             )
         )
     else:
-        shape = shape.numpy().astype(int).tolist()
+        shape = shape.astype(int).tolist()
     return shape