[Zero-Dim] fix Tensor.numpy, cntrol whether to hack process to 1D (#51757)

paddle/fluid/pybind/eager_method.cc

@@ -123,15 +123,29 @@ static PyObject* tensor_method_numpy(TensorObject* self,
   size_t py_rank = tensor_dims.size();
   size_t numel = 1;
   if (py_rank == 0) {
+    Py_ssize_t args_num = PyTuple_Size(args);
+    bool set_to_1d = true;
+    if (args_num == (Py_ssize_t)1) {
+      PyObject* obj = PyTuple_GET_ITEM(args, 0);
+      if (obj == Py_False) {
+        set_to_1d = false;
+      }
+    }
+    if (set_to_1d) {
       // 0D Tensor hack process to 1D numpy, will remove in future
-    VLOG(0) << "Warning:: 0D Tensor cannot be used as Tensor.numpy()[0], Now "
-               "0D will be changed to 1D numpy to avoid this problem, but it's "
-               "not correct and will be removed in future. Please change "
-               "'Tensor.numpy()[0]' to 'float(Tensor)' or "
-               "'Tensor.numpy().item()' as soon as possible.";
+      VLOG(0)
+          << "Warning:: 0D Tensor cannot be used as 'Tensor.numpy()[0]' . In "
+             "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 "
+             "possible, "
+             "otherwise 'Tensor.numpy()[0]' will raise error";
       py_rank = 1;
       py_dims[0] = 1;
       py_strides[0] = sizeof_dtype * numel;
+    }
   } else {
     for (int i = tensor_dims.size() - 1; i >= 0; --i) {
       py_dims[i] = static_cast<size_t>(tensor_dims[i]);
@@ -143,7 +157,7 @@ static PyObject* tensor_method_numpy(TensorObject* self,
   PyObject* array = api.PyArray_NewFromDescr_(
       api.PyArray_Type_,
       api.PyArray_DescrFromType_(numpy_dtype),
-      tensor_dims.size(),
+      py_rank,
       py_dims,
       py_strides,
       nullptr,

python/paddle/distributed/fleet/meta_parallel/pipeline_parallel.py

@@ -415,9 +415,9 @@ class PipelineParallel(MetaParallelBase):
             ), "train_batch() in last stage should obtain vaild loss"
             loss = self.total_loss.detach()
             is_fp32 = (
-                paddle.to_tensor(1)
+                paddle.full([], 1, 'int64')
                 if loss.dtype == paddle.float32
-                else paddle.to_tensor(0)
+                else paddle.full([], 0, 'int64')
             )
             paddle.distributed.broadcast(
                 is_fp32, src=self.global_rank, sync_op=True, group=self.pp_group
@@ -426,7 +426,7 @@ class PipelineParallel(MetaParallelBase):
                 loss, src=self.global_rank, sync_op=True, group=self.pp_group
             )
         else:
-            is_fp32 = paddle.to_tensor(1)
+            is_fp32 = paddle.full([], 1, 'int64')
             paddle.distributed.broadcast(
                 is_fp32,
                 src=self._hcg.get_rank_from_stage(self.num_stages - 1),
@@ -435,7 +435,7 @@ class PipelineParallel(MetaParallelBase):
             )
             loss = (
                 paddle.zeros(shape=[1], dtype="float32")
-                if is_fp32.numpy()[0]
+                if is_fp32.item()
                 else paddle.zeros(shape=[1], dtype="float16")
             )
             paddle.distributed.broadcast(

python/paddle/fluid/dygraph/varbase_patch_methods.py

@@ -764,7 +764,7 @@ def monkey_patch_varbase():
                 print(type(x_array))      #<class 'numpy.ndarray'>
                 print(x_array.shape)      #(2, 2)
         """
-        array = self.numpy()
+        array = self.numpy(False)
         if dtype:
             array = array.astype(dtype)
         return array

python/paddle/fluid/framework.py

@@ -773,7 +773,7 @@ def _var_base_to_np(var_base):
         "paddle.fluid.framework._var_base_to_np is deprecated, please use var_base.numpy() instead of _var_base_to_np(var_base)."
     )
 
-    return var_base.numpy()
+    return var_base.numpy(False)
 
 
 def _cpu_num():

python/paddle/fluid/layers/nn.py

@@ -698,10 +698,10 @@ def unsqueeze(input, axes, name=None):
         if isinstance(axes, int):
             axes = [axes]
         elif isinstance(axes, Variable):
-            axes = axes.numpy().tolist()
+            axes = axes.tolist()
         elif isinstance(axes, (list, tuple)):
             axes = [
-                item.numpy().item(0) if isinstance(item, Variable) else item
+                item.item(0) if isinstance(item, Variable) else item
                 for item in axes
             ]
         return _C_ops.unsqueeze(input, axes)

python/paddle/fluid/optimizer.py

@@ -2432,12 +2432,12 @@ class AdamOptimizer(Optimizer):
             _beta1 = (
                 self._beta1
                 if not isinstance(self._beta1, Variable)
-                else self._beta1.numpy().item(0)
+                else self._beta1.item(0)
             )
             _beta2 = (
                 self._beta2
                 if not isinstance(self._beta2, Variable)
-                else self._beta2.numpy().item(0)
+                else self._beta2.item(0)
             )
             master_weight = None
             _, _, _, _, _, _ = _legacy_C_ops.adam(

python/paddle/fluid/tests/unittests/test_zero_dim_tensor.py

@@ -952,7 +952,15 @@ class TestSundryAPI(unittest.TestCase):
 
     def test_numpy(self):
         x = paddle.full([], 0.5)
-        np.testing.assert_array_equal(x.numpy(), np.array(0.5))
+        # 0D Tensor hack to 1D Numpy defaut, will remove in future
+        x_np = x.numpy()
+        np.testing.assert_array_equal(x_np.shape, (1,))
+        np.testing.assert_array_equal(x_np, np.array([0.5]))
+
+        # return origin correct numpy
+        x_np = x.numpy(False)
+        np.testing.assert_array_equal(x_np.shape, ())
+        np.testing.assert_array_equal(x_np, np.array(0.5))
 
     def test_numel(self):
         out = paddle.numel(self.x)

python/paddle/geometric/message_passing/utils.py

@@ -29,7 +29,7 @@ def convert_out_size_to_list(out_size):
     elif isinstance(out_size, (int, np.int32, np.int64)):
         out_size = [out_size]
     else:
-        out_size = [out_size.numpy().astype(int)[0]]
+        out_size = [int(out_size)]
     return out_size
 
 

python/paddle/incubate/operators/graph_send_recv.py

@@ -182,7 +182,7 @@ def convert_out_size_to_list(out_size):
     elif isinstance(out_size, (int, np.int32, np.int64)):
         out_size = [out_size]
     else:
-        out_size = [out_size.numpy().astype(int)[0]]
+        out_size = [int(out_size)]
     return out_size
 
 

python/paddle/nn/functional/common.py

@@ -502,7 +502,7 @@ def interpolate(
 
                 for i, dim in enumerate(out_shape):
                     if isinstance(dim, Variable):
-                        out_shape[i] = dim.numpy().item()
+                        out_shape[i] = dim.item()
             if not (_is_list_or_turple_(out_shape)):
                 raise TypeError("size should be a list or tuple or Variable.")
             # Validate the shape
@@ -1692,7 +1692,7 @@ def pad(x, pad, mode='constant', value=0.0, data_format="NCHW", name=None):
 
     if in_dygraph_mode():
         if isinstance(pad, Variable):
-            pad = pad.numpy().tolist()
+            pad = pad.tolist()
         out = _C_ops.pad3d(x, pad, mode, value, data_format)
     else:
         attrs = {'mode': mode, 'value': value, 'data_format': data_format}

python/paddle/nn/functional/vision.py

@@ -86,17 +86,13 @@ def affine_grid(theta, out_shape, align_corners=True, name=None):
 
     if in_dygraph_mode():
         _out_shape = (
-            out_shape.numpy().tolist()
-            if isinstance(out_shape, Variable)
-            else out_shape
+            out_shape.tolist() if isinstance(out_shape, Variable) else out_shape
         )
         theta = theta._use_gpudnn(use_cudnn)
         return _C_ops.affine_grid(theta, _out_shape, align_corners)
     elif in_dynamic_mode():
         _out_shape = (
-            out_shape.numpy().tolist()
-            if isinstance(out_shape, Variable)
-            else out_shape
+            out_shape.tolist() if isinstance(out_shape, Variable) else out_shape
         )
         return _legacy_C_ops.affine_grid(
             theta,

python/paddle/nn/initializer/assign.py

@@ -211,6 +211,6 @@ class Assign(NumpyArrayInitializer):
 
         # TODO: value is already is a tensor, accounting efficiency maybe it does not need to convert tensor to numpy data and then initialized.
         if isinstance(value, paddle.static.Variable):
-            value = value.numpy()
+            value = value.numpy(False)
 
         super().__init__(value)

python/paddle/optimizer/adam.py

@@ -310,12 +310,12 @@ class Adam(Optimizer):
             _beta1 = (
                 self._beta1
                 if not isinstance(self._beta1, Variable)
-                else self._beta1.numpy().item(0)
+                else self._beta1.item(0)
             )
             _beta2 = (
                 self._beta2
                 if not isinstance(self._beta2, Variable)
-                else self._beta2.numpy().item(0)
+                else self._beta2.item(0)
             )
 
             _, _, _, _, _, _ = _C_ops.adam_(
@@ -623,12 +623,12 @@ class Adam(Optimizer):
                 _beta1 = (
                     self._beta1
                     if not isinstance(self._beta1, Variable)
-                    else self._beta1.numpy().item(0)
+                    else self._beta1.item(0)
                 )
                 _beta2 = (
                     self._beta2
                     if not isinstance(self._beta2, Variable)
-                    else self._beta2.numpy().item(0)
+                    else self._beta2.item(0)
                 )
 
                 if framework.in_dygraph_mode():

python/paddle/optimizer/adamw.py

@@ -434,12 +434,12 @@ class AdamW(Optimizer):
             _beta1 = (
                 self._beta1
                 if not isinstance(self._beta1, Variable)
-                else self._beta1.numpy().item(0)
+                else self._beta1.item(0)
             )
             _beta2 = (
                 self._beta2
                 if not isinstance(self._beta2, Variable)
-                else self._beta2.numpy().item(0)
+                else self._beta2.item(0)
             )
 
             _, _, _, _, _, _ = _C_ops.adamw_(

python/paddle/quantization/imperative/ptq_quantizer.py

@@ -24,7 +24,7 @@ from . import utils
 
 
 def abs_max_value(tensor):
-    return float(paddle.max(paddle.abs(tensor)).numpy())
+    return float(paddle.max(paddle.abs(tensor)))
 
 
 def merge_max_value(old, new):

python/paddle/static/nn/control_flow.py

@@ -466,7 +466,7 @@ def while_loop(cond, body, loop_vars, is_test=False, name=None):
         )
 
     if _non_static_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)):
@@ -476,7 +476,7 @@ 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
 
@@ -968,7 +968,7 @@ def cond(pred, true_fn=None, false_fn=None, name=None, return_names=None):
     if _non_static_mode():
         assert isinstance(pred, Variable), "The pred in cond must be Variable"
         assert pred.size == 1, "condition input's numel should be 1"
-        pred = pred.numpy().item()
+        pred = pred.item()
         if pred:
             if true_fn is not None:
                 if not callable(true_fn):

python/paddle/tensor/manipulation.py

@@ -330,9 +330,7 @@ def slice(input, axes, starts, ends):
 
         if isinstance(starts, (list, tuple)):
             starts = [
-                item.numpy().item(0)
-                if isinstance(item, tmp_tensor_type)
-                else item
+                item.item(0) if isinstance(item, tmp_tensor_type) else item
                 for item in starts
             ]
         elif isinstance(starts, tmp_tensor_type):
@@ -342,9 +340,7 @@ def slice(input, axes, starts, ends):
 
         if isinstance(ends, (list, tuple)):
             ends = [
-                item.numpy().item(0)
-                if isinstance(item, tmp_tensor_type)
-                else item
+                item.item(0) if isinstance(item, tmp_tensor_type) else item
                 for item in ends
             ]
         elif isinstance(ends, tmp_tensor_type):
@@ -1069,7 +1065,8 @@ def tolist(x):
             print(expectlist)   #[0, 1, 2, 3, 4]
 
     """
-    return x.numpy().tolist()
+    # TODO(zhouwei): will remove 0D Tensor.numpy() hack
+    return x.numpy(False).tolist()
 
 
 def concat(x, axis=0, name=None):
@@ -1118,7 +1115,6 @@ def concat(x, axis=0, name=None):
     input = x
     if in_dygraph_mode():
         if isinstance(axis, Variable):
-            axis = axis.numpy()
             axis = axis.item(0)
         if not isinstance(input, Variable):
             input = [t for t in input if t.shape.count(0) == 0]
@@ -1953,7 +1949,6 @@ def split(x, num_or_sections, axis=0, name=None):
     dim = axis
     if in_dygraph_mode():
         if isinstance(dim, Variable):
-            dim = dim.numpy()
             dim = dim.item(0)
         assert len(input.shape) + dim >= 0, "(rank(x) + axis) must >= 0"
         dim = (len(input.shape) + dim) if dim < 0 else dim
@@ -1962,9 +1957,7 @@ def split(x, num_or_sections, axis=0, name=None):
             if paddle.utils._contain_var(num_or_sections):
                 for index, item in enumerate(num_or_sections):
                     if isinstance(item, Variable):
-                        num_or_sections[index] = num_or_sections[index].numpy()[
-                            0
-                        ]
+                        num_or_sections[index] = num_or_sections[index].item()
         elif not isinstance(num_or_sections, int):
             raise TypeError(
                 "The type of 'num_or_sections' in split must be int, list or tuple in imperative mode, but "
@@ -2594,10 +2587,10 @@ def unsqueeze(x, axis, name=None):
         if isinstance(axes, int):
             axes = [axes]
         elif isinstance(axes, Variable):
-            axes = axes.numpy().tolist()
+            axes = axes.tolist()
         elif isinstance(axes, (list, tuple)):
             axes = [
-                item.numpy().item(0) if isinstance(item, Variable) else item
+                item.item(0) if isinstance(item, Variable) else item
                 for item in axes
             ]
         return _C_ops.unsqueeze(input, axes)
@@ -2660,10 +2653,10 @@ def unsqueeze_(x, axis, name=None):
     if isinstance(axes, int):
         axes = [axes]
     elif isinstance(axes, Variable):
-        axes = axes.numpy().tolist()
+        axes = axes.tolist()
     elif isinstance(axes, (list, tuple)):
         axes = [
-            item.numpy().item(0) if isinstance(item, Variable) else item
+            item.item(0) if isinstance(item, Variable) else item
             for item in axes
         ]
     return _C_ops.unsqueeze_(input, axes)
@@ -3149,7 +3142,7 @@ def tile(x, repeat_times, name=None):
             assert (
                 repeat_times.ndim == 1
             ), "Only support ndim == 1 while repeat_times is a Tensor."
-            repeat_times = repeat_times.numpy().tolist()
+            repeat_times = repeat_times.tolist()
 
         return _C_ops.tile(x, repeat_times)
     else:
@@ -3649,9 +3642,7 @@ def reshape_(x, shape, name=None):
         tmp_tensor_type = core.eager.Tensor
         if isinstance(shape, (list, tuple)):
             shape = [
-                item.numpy().item(0)
-                if isinstance(item, tmp_tensor_type)
-                else item
+                item.item(0) if isinstance(item, tmp_tensor_type) else item
                 for item in shape
             ]
             if shape == x.shape:

python/paddle/tensor/math.py

@@ -2872,9 +2872,9 @@ def clip(x, min=None, max=None, name=None):
 
     if in_dygraph_mode():
         if isinstance(min, Variable):
-            min = min.numpy().item(0)
+            min = min.item(0)
         if isinstance(max, Variable):
-            max = max.numpy().item(0)
+            max = max.item(0)
         min = min_ if min is None else min
         max = max_ if max is None else max
         return _C_ops.clip(x, min, max)
@@ -2939,9 +2939,9 @@ def clip_(x, min=None, max=None, name=None):
     fmin = float(np.finfo(np.float32).min)
     fmax = float(np.finfo(np.float32).max)
     if isinstance(min, Variable):
-        min = min.numpy().item(0)
+        min = min.item(0)
     if isinstance(max, Variable):
-        max = max.numpy().item(0)
+        max = max.item(0)
     min = fmin if min is None else min
     max = fmax if max is None else max
 

python/paddle/tensor/to_string.py

@@ -257,7 +257,7 @@ def to_string(var, prefix='Tensor'):
 
     if var.dtype == core.VarDesc.VarType.BF16:
         var = var.astype('float32')
-    np_var = var.numpy()
+    np_var = var.numpy(False)
 
     if len(var.shape) == 0:
         size = 0
@@ -291,7 +291,8 @@ def _format_dense_tensor(tensor, indent):
     if tensor.dtype == core.VarDesc.VarType.BF16:
         tensor = tensor.astype('float32')
 
-    np_tensor = tensor.numpy()
+    # TODO(zhouwei): will remove 0D Tensor.numpy() hack
+    np_tensor = tensor.numpy(False)
 
     if len(tensor.shape) == 0:
         size = 0