[Eager] Support test imperative basic in eager test_empty_grad (#38376)

* Rearranged Eager AutoCodeGen directory structure

* Removed USE_OP in Eager AutoCodeGen

* Enabled generation for Operators without Grad/Inputs/Outputs

* Resolved operators without input

* Fixed merge conflicts

* Enabled Eager AutoCodeGen for 10+ more operators

* Refactored Eager AutoCodeGen with more organized helper objects

* Enabled Eager AutoCodeGen for operators with multiple OpBases

* Adjusted Eager AutoCodeGen to Enable Passing Output Tensor as Input Argument

* Handled Dispensable Inputs/Outputs in Eager AutoCodeGen

* Adjusted function generation/call between Python-C API & Dygraph API

* Synchronized auto-generated Python-C API with Dygraph Forward Functions

* support more eager tensor api

* fix merge compile error

* fix compile error and fit develop code

* support pure CPU

* fix some logic error in eager_mode

* support _varbase_creator in eager mode

* Added safe_initialized interface to EagerTensor for use in processing dispensable inputs

* for eager mode

* refine

* support multiple constructor for eager tensor

* add place related code

* polish code

* specific randint with dtype of int64

* Support pure cpu test

* eager logic

* refine test in pure cpu

* eager logic

* eager logic

* eager logic, test=develop

* skip core.eager when in inference, test=develop

* refine, test=develop

* refine, test=develop

* call RetainGrad after run forward kernel, test=develop

* refine, test=develop

* support dygraph util, meta, guard test

* eager test case

* support inference test

* refine test and fix initializer failed

* modify eagertensor patch method

* add eagertensor.clear_grandint, test=develop

* refine, test=develop

* refine, test=develop

* refine, test=develop

* call monkey_patch_varbase in _test_eager_guard, test=develop

* split clear_gradient to clear_gradient and zero_grads, test=develop

* refine, test=develop

* refine, test=develop

* refine, test=develop
Co-authored-by: Njim19930609 <jim19930609@gmail.com>
Co-authored-by: NJiabinYang <360788950@qq.com>

paddle/fluid/pybind/eager_method.cc

@@ -27,6 +27,7 @@ limitations under the License. */
 #include "paddle/fluid/pybind/eager.h"
 #include "paddle/fluid/pybind/eager_utils.h"
 #include "paddle/fluid/pybind/exception.h"
+#include "paddle/pten/api/include/api.h"
 #include "paddle/pten/common/data_type.h"
 #include "paddle/pten/core/convert_utils.h"
 #include "paddle/pten/core/dense_tensor.h"
@@ -125,13 +126,17 @@ static PyObject* eager_tensor_method_copy_(EagerTensorObject* self,
   bool blocking = CastPyArg2AttrBoolean(PyTuple_GET_ITEM(args, 1), 1);
   VLOG(6) << "Start Copy Tensor " << src_tensor.name() << " to "
           << self->eager_tensor.name();
-  self->eager_tensor.copy_(src_tensor, blocking);
+  if (!self->eager_tensor.defined()) {
     egr::EagerUtils::autograd_meta(&(self->eager_tensor))
         ->SetStopGradient(
             egr::EagerUtils::autograd_meta(&(src_tensor))->StopGradient());
     egr::EagerUtils::autograd_meta(&(self->eager_tensor))
         ->SetPersistable(
             egr::EagerUtils::autograd_meta(&(src_tensor))->Persistable());
+  }
+
+  self->eager_tensor.copy_(src_tensor, blocking);
+
   VLOG(6) << "Finish Copy Tensor " << src_tensor.name() << " to "
           << self->eager_tensor.name();
   Py_INCREF(Py_None);
@@ -156,6 +161,74 @@ static PyObject* eager_tensor_retain_grads(EagerTensorObject* self,
   EAGER_CATCH_AND_THROW_RETURN_NULL
 }
 
+static PyObject* eager_tensor__clear_gradient(EagerTensorObject* self,
+                                              PyObject* args,
+                                              PyObject* kwargs) {
+  EAGER_SYNC_TRY
+  VLOG(4) << "ClearGradient " << self->eager_tensor.name();
+
+  egr::EagerTensor grad;
+  if (egr::egr_utils_api::IsLeafTensor(self->eager_tensor)) {
+    // Add RetainGrad as PostHook to AccumulationNode
+    std::shared_ptr<egr::GradNodeBase> grad_node =
+        egr::EagerUtils::grad_node(self->eager_tensor);
+    PADDLE_ENFORCE(
+        grad_node.get() != nullptr,
+        paddle::platform::errors::Fatal("Detected NULL grad_node"
+                                        "Leaf tensor should have had grad_node "
+                                        "with type: GradNodeAccumulation"));
+    auto accumulation_grad_node =
+        std::dynamic_pointer_cast<egr::GradNodeAccumulation>(grad_node);
+    grad = accumulation_grad_node->Grad();
+  } else {
+    auto meta = egr::EagerUtils::unsafe_autograd_meta(self->eager_tensor);
+    grad = meta->Grad();
+  }
+
+  if (grad.initialized()) {
+    VLOG(4) << "Gradient of " << self->eager_tensor.name()
+            << " is initialized, will be released.";
+    auto dense_tensor =
+        std::dynamic_pointer_cast<pten::DenseTensor>(grad.impl());
+    dense_tensor->release();
+  }
+  Py_INCREF(Py_None);
+  return Py_None;
+  EAGER_CATCH_AND_THROW_RETURN_NULL
+}
+
+static PyObject* eager_tensor__zero_grads(EagerTensorObject* self,
+                                          PyObject* args, PyObject* kwargs) {
+  EAGER_TRY
+  VLOG(4) << "ZeroGrads " << self->eager_tensor.name();
+
+  egr::EagerTensor grad;
+  if (egr::egr_utils_api::IsLeafTensor(self->eager_tensor)) {
+    // Add RetainGrad as PostHook to AccumulationNode
+    std::shared_ptr<egr::GradNodeBase> grad_node =
+        egr::EagerUtils::grad_node(self->eager_tensor);
+    PADDLE_ENFORCE(
+        grad_node.get() != nullptr,
+        paddle::platform::errors::Fatal("Detected NULL grad_node"
+                                        "Leaf tensor should have had grad_node "
+                                        "with type: GradNodeAccumulation"));
+    auto accumulation_grad_node =
+        std::dynamic_pointer_cast<egr::GradNodeAccumulation>(grad_node);
+    grad = accumulation_grad_node->Grad();
+  } else {
+    auto meta = egr::EagerUtils::unsafe_autograd_meta(self->eager_tensor);
+    grad = meta->Grad();
+  }
+
+  if (grad.initialized()) {
+    grad.set_tensor(std::make_shared<paddle::experimental::Tensor>(
+        paddle::experimental::zeros_like(*(grad.Tensor().get()))));
+  }
+  Py_INCREF(Py_None);
+  return Py_None;
+  EAGER_CATCH_AND_THROW_RETURN_NULL
+}
+
 PyMethodDef variable_methods[] = {
     {"numpy", (PyCFunction)(void (*)(void))eager_tensor_method_numpy,
      METH_VARARGS | METH_KEYWORDS, NULL},
@@ -168,6 +241,11 @@ PyMethodDef variable_methods[] = {
      METH_VARARGS | METH_KEYWORDS, NULL},
     {"retain_grads", (PyCFunction)(void (*)(void))eager_tensor_retain_grads,
      METH_VARARGS | METH_KEYWORDS, NULL},
+    {"_clear_gradient",
+     (PyCFunction)(void (*)(void))eager_tensor__clear_gradient,
+     METH_VARARGS | METH_KEYWORDS, NULL},
+    {"_zero_grads", (PyCFunction)(void (*)(void))eager_tensor__zero_grads,
+     METH_VARARGS | METH_KEYWORDS, NULL},
     {NULL, NULL, 0, NULL}};
 
 }  // namespace pybind

paddle/fluid/pybind/eager_utils.cc

@@ -222,6 +222,8 @@ std::vector<egr::EagerTensor> CastPyArg2VectorOfEagerTensor(PyObject* obj,
             reinterpret_cast<PyTypeObject*>(item->ob_type)->tp_name, i));
       }
     }
+  } else if (obj == Py_None) {
+    return {};
   } else {
     PADDLE_THROW(platform::errors::InvalidArgument(
         "argument (position %d) must be "
@@ -263,6 +265,8 @@ std::vector<int> CastPyArg2VectorOfInt(PyObject* obj, size_t arg_pos) {
             reinterpret_cast<PyTypeObject*>(item->ob_type)->tp_name, i));
       }
     }
+  } else if (obj == Py_None) {
+    return {};
   } else {
     PADDLE_THROW(platform::errors::InvalidArgument(
         "argument (position %d) must be "
@@ -557,6 +561,8 @@ std::vector<egr::EagerTensor> GetEagerTensorListFromArgs(
           reinterpret_cast<EagerTensorObject*>(PyTuple_GetItem(list, i))
               ->eager_tensor);
     }
+  } else if (list == Py_None) {
+    return {};
   } else {
     PADDLE_THROW(platform::errors::InvalidArgument(
         "%s(): argument '%s' (position %d) must be list of Tensors, but got "
@@ -634,6 +640,8 @@ std::vector<egr::EagerTensor*> GetEagerTensorPtrListFromArgs(
           &(reinterpret_cast<EagerTensorObject*>(PyTuple_GetItem(list, i))
                 ->eager_tensor));
     }
+  } else if (list == Py_None) {
+    return {};
   } else {
     PADDLE_THROW(platform::errors::InvalidArgument(
         "%s(): argument '%s' (position %d) must be list of Tensors, but got "
@@ -644,6 +652,5 @@ std::vector<egr::EagerTensor*> GetEagerTensorPtrListFromArgs(
 
   return result;
 }
-
 }  // namespace pybind
 }  // namespace paddle

python/paddle/fluid/__init__.py

@@ -55,7 +55,6 @@ from . import initializer
 from .initializer import set_global_initializer
 from . import layers
 from . import dygraph
-from . import eager
 from . import contrib
 from . import nets
 from . import optimizer
@@ -91,7 +90,6 @@ from .dygraph.base import enable_dygraph, disable_dygraph
 from .io import save, load, load_program_state, set_program_state
 from .dygraph.checkpoint import save_dygraph, load_dygraph
 from .dygraph.varbase_patch_methods import monkey_patch_varbase
-from .eager.eager_tensor_patch_methods import monkey_patch_eagertensor
 from . import generator
 from .core import _cuda_synchronize
 from .generator import Generator
@@ -115,7 +113,6 @@ __all__ = framework.__all__ + executor.__all__ + \
         'contrib',
         'data',
         'dygraph',
-        'eager',
         'enable_dygraph',
         'disable_dygraph',
         'enable_imperative',
@@ -221,7 +218,6 @@ def __bootstrap__():
 monkey_patch_variable()
 __bootstrap__()
 monkey_patch_varbase()
-monkey_patch_eagertensor()
 
 # NOTE(zhiqiu): register npu_finalize on the exit of Python,
 # do some clean up manually.

python/paddle/fluid/dygraph/math_op_patch.py

@@ -60,6 +60,7 @@ _complex_dtypes = [
 ]
 
 _already_patch_varbase = False
+_already_patch_eager_tensor = False
 
 
 def monkey_patch_math_varbase():
@@ -220,7 +221,11 @@ def monkey_patch_math_varbase():
 
             # 2. create varbase for scalar
             lhs_dtype = self.dtype
-            if not isinstance(other_var, core.VarBase):
+            if _in_eager_mode():
+                other_var_should_be = core.eager.EagerTensor
+            else:
+                other_var_should_be = core.VarBase
+            if not isinstance(other_var, other_var_should_be):
                 if isinstance(other_var, complex):
                     import paddle
                     other_var = paddle.to_tensor(other_var, dtype='complex64')
@@ -333,22 +338,30 @@ def monkey_patch_math_varbase():
     ]
 
     global _already_patch_varbase
+    global _already_patch_eager_tensor
+
+    if core._in_eager_mode():
+        local_already_patch = _already_patch_eager_tensor
+        _already_patch_eager_tensor = True
+        local_tensor = core.eager.EagerTensor
+    else:
+        local_already_patch = _already_patch_varbase
+        _already_patch_varbase = True
+        local_tensor = core.VarBase
 
-    if not _already_patch_varbase:
+    if not local_already_patch:
         for method in varbase_methods:
             method_name = method[0]
             method_impl = method[1]
-            setattr(core.VarBase, method_name, method_impl)
+            setattr(local_tensor, method_name, method_impl)
     else:
         import paddle.tensor
         # Tensor method from module paddle.tensor
         for method_name in paddle.tensor.tensor_method_func:
-            if hasattr(core.VarBase, method_name): continue
+            if hasattr(local_tensor, method_name): continue
             method_impl = getattr(paddle.tensor, method_name, None)
-            if method_impl: setattr(core.VarBase, method_name, method_impl)
+            if method_impl: setattr(local_tensor, method_name, method_impl)
 
         for magic_method, origin_method in paddle.tensor.magic_method_func:
             impl = getattr(paddle.tensor, origin_method, None)
-            if impl: setattr(core.VarBase, magic_method, impl)
-
-    _already_patch_varbase = True
+            if impl: setattr(local_tensor, magic_method, impl)

python/paddle/fluid/dygraph/varbase_patch_methods.py

@@ -22,7 +22,7 @@ import paddle
 from .. import framework
 from .. import core
 from .. import unique_name
-from ..framework import Variable, Parameter, ParamBase, _getitem_impl_, _setitem_impl_
+from ..framework import Variable, Parameter, ParamBase, _getitem_impl_, _setitem_impl_, _in_eager_mode
 from .base import switch_to_static_graph
 from .math_op_patch import monkey_patch_math_varbase
 from .parallel import scale_loss
@@ -58,6 +58,9 @@ class TensorHookRemoveHelper(object):
         return False
 
 
+_already_patch_repr = False
+
+
 def monkey_patch_varbase():
     @switch_to_static_graph
     def _to_static_var(self, to_parameter=False, **kwargs):
@@ -146,7 +149,11 @@ def monkey_patch_varbase():
                     out = linear(t)  # call with different weight
 
         """
-        assert isinstance(value, (np.ndarray, core.VarBase, dict, str)), \
+        if _in_eager_mode():
+            base_tensor = core.eager.EagerTensor
+        else:
+            base_tensor = core.VarBase
+        assert isinstance(value, (np.ndarray, base_tensor, dict, str)), \
             "Variable set_value function, arguments type only support Variable, numpy, VarBase, dict, string."
 
         if isinstance(value, (dict, str)):
@@ -160,7 +167,7 @@ def monkey_patch_varbase():
                 self.value().set_string_list(value)
         else:
             value_np = value
-            if isinstance(value, core.VarBase):
+            if isinstance(value, base_tensor):
                 value_np = value.numpy()
 
             self_tensor_np = self.numpy()
@@ -231,6 +238,11 @@ def monkey_patch_varbase():
         """
         if framework.in_dygraph_mode():
             if grad_tensor is not None:
+                if _in_eager_mode():
+                    assert isinstance(
+                        grad_tensor, core.eager.EagerTensor
+                    ), "The type of grad_tensor must be paddle.Tensor"
+                else:
                     assert isinstance(
                         grad_tensor, paddle.
                         Tensor), "The type of grad_tensor must be paddle.Tensor"
@@ -238,14 +250,27 @@ def monkey_patch_varbase():
                     "Tensor shape not match, Tensor of grad_tensor [ {} ] with shape {} mismatch Tensor [ {} ] with shape {}".format(
                     grad_tensor.name, grad_tensor.shape, self.name, self.shape)
 
+            if _in_eager_mode():
+                if grad_tensor is None:
+                    grad_tensor = []
+                else:
+                    grad_tensor = [grad_tensor]
             if paddle.is_compiled_with_xpu() or paddle.is_compiled_with_npu():
                 # TODO(liuyuhui): Currently only for xpu. Will be removed in the future.
                 scaled_loss = scale_loss(self)
+                if _in_eager_mode():
+                    core.eager.run_backward([scaled_loss], grad_tensor,
+                                            retain_graph)
+                else:
                     core.dygraph_run_backward([scaled_loss], [grad_tensor],
                                               retain_graph,
                                               framework._dygraph_tracer())
             else:
-                core.dygraph_run_backward([self], [grad_tensor], retain_graph,
+                if _in_eager_mode():
+                    core.eager.run_backward([self], grad_tensor, retain_graph)
+                else:
+                    core.dygraph_run_backward([self], [grad_tensor],
+                                              retain_graph,
                                               framework._dygraph_tracer())
         else:
             raise ValueError(
@@ -280,12 +305,19 @@ def monkey_patch_varbase():
                 # [500.]
 
         """
+        if _in_eager_mode():
+            if not self.grad._is_initialized():
+                return None
+            # TODO(wanghuancoder) support SELECTED_ROWS
+            return self.grad.numpy()
+        else:
             if self._grad_ivar() is None:
                 return None
 
             new_ivar = self._grad_ivar()._copy_to(core.CPUPlace(), True)
             if self._grad_ivar().type == core.VarDesc.VarType.SELECTED_ROWS:
-            return (np.array(new_ivar.value().get_selected_rows().get_tensor()),
+                return (
+                    np.array(new_ivar.value().get_selected_rows().get_tensor()),
                     np.array(new_ivar.value().get_selected_rows().rows()))
             else:
                 return np.array(new_ivar.value().get_tensor())
@@ -555,6 +587,10 @@ def monkey_patch_varbase():
                 #        [[0.30574632, 0.55739117, 0.30902600, 0.39413780, 0.44830436],
                 #         [0.79010487, 0.53972793, 0.09495186, 0.44267157, 0.72112119]])
         """
+        if _in_eager_mode():
+            from paddle.tensor.to_string import eager_tensor_to_string
+            return eager_tensor_to_string(self)
+        else:
             from paddle.tensor.to_string import to_string
             return to_string(self)
 
@@ -583,6 +619,9 @@ def monkey_patch_varbase():
             raise RuntimeError(
                 "Only Leaf Tensor support the deepcopy at the moment, non-Leaf Tensors contains graph information that does't support deepcopy"
             )
+        if _in_eager_mode():
+            new_varbase = core.eager.EagerTensor()
+        else:
             new_varbase = core.VarBase()
         new_varbase.name = self.name + unique_name.generate("_deepcopy")
         memo[id(self)] = new_varbase
@@ -717,19 +756,47 @@ def monkey_patch_varbase():
             # Call c++ func __setitem_varbase__ to speedup.
             return self.__setitem_varbase__(item, value)
 
+    @framework.dygraph_only
+    def _grad_ivar(self):
+        if self.grad._is_initialized():
+            return self.grad
+        else:
+            return None
+
+    @framework.dygraph_only
+    def clear_gradient(self, set_to_zero=True):
+        if set_to_zero:
+            self._zero_grads()
+        else:
+            self._clear_gradient()
+
+    if core._in_eager_mode() and not hasattr(core, "eager"):
+        return
+
     for method_name, method in (
         ("__bool__", __bool__), ("__nonzero__", __nonzero__),
         ("_to_static_var", _to_static_var), ("set_value", set_value),
         ("block", block), ("backward", backward), ("clear_grad", clear_grad),
-        ("inplace_version", inplace_version), ("grad", grad),
-        ("gradient", gradient), ("register_hook", register_hook),
-        ("__str__", __str__), ("__repr__", __str__),
-        ("__deepcopy__", __deepcopy__), ("__module__", "paddle"),
-        ("__name__", "Tensor"), ("__array__", __array__),
+        ("inplace_version", inplace_version), ("gradient", gradient),
+        ("register_hook", register_hook), ("__str__", __str__),
+        ("__repr__", __str__), ("__deepcopy__", __deepcopy__),
+        ("__module__", "paddle"), ("__array__", __array__),
         ("__getitem__", __getitem__), ("item", item),
         ("__setitem__", __setitem__), ("_to", _to)):
+        if core._in_eager_mode():
+            setattr(core.eager.EagerTensor, method_name, method)
+        else:
             setattr(core.VarBase, method_name, method)
 
+    if core._in_eager_mode():
+        setattr(core.eager.EagerTensor, "_grad_ivar", _grad_ivar)
+        setattr(core.eager.EagerTensor, "clear_gradient", clear_gradient)
+    else:
+        setattr(core.VarBase, "__name__", "Tensor")
+        setattr(core.VarBase, "grad", grad)
+
+    global _already_patch_repr
+    if not _already_patch_repr:
         # NOTE(zhiqiu): pybind11 will set a default __str__ method of enum class.
         # So, we need to overwrite it to a more readable one.
         # See details in https://github.com/pybind/pybind11/issues/2537.
@@ -744,6 +811,7 @@ def monkey_patch_varbase():
                 return origin(dtype)
 
         setattr(core.VarDesc.VarType, "__repr__", dtype_str)
+        _already_patch_repr = True
 
     # patch math methods for varbase
     monkey_patch_math_varbase()

python/paddle/fluid/eager/__init__.py

-#   Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-
-# incubate directory is mainly for internal use
-# after we have tested incubate APIs in industrial application for a period
-# we will move stable functions into fluid
-
-from . import eager_tensor_patch_methods
-
-__all__ = []

python/paddle/fluid/eager/eager_tensor_patch_methods.py

-#   Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-
-from .. import core as core
-from .. import framework as framework
-from ..dygraph.parallel import scale_loss
-import numpy as np
-
-
-def monkey_patch_eagertensor():
-    def __str__(self):
-        from paddle.tensor.to_string import eager_tensor_to_string
-        return eager_tensor_to_string(self)
-
-    @framework.dygraph_only
-    def backward(self, grad_tensor=None, retain_graph=False):
-        """
-        Run backward of current Graph which starts from current Tensor.
-
-        The new gradient will accumulat on previous gradient.
-
-        You can clear gradient by ``Tensor.clear_grad()`` .
-
-        Args:
-            grad_tensor(Tensor, optional): initial gradient values of the current Tensor. If `grad_tensor` is None, 
-            the initial gradient values of the current Tensor would be Tensor filled with 1.0; 
-            if `grad_tensor` is not None, it must have the same length as the current Tensor.
-            Teh default value is None.
-
-            retain_graph(bool, optional): If False, the graph used to compute grads will be freed. If you would
-                like to add more ops to the built graph after calling this method( :code:`backward` ), set the parameter
-                :code:`retain_graph` to True, then the grads will be retained. Thus, seting it to False is much more memory-efficient.
-                Defaults to False.
-        Returns:
-            NoneType: None
-
-        Examples:
-            .. code-block:: python
-
-                import paddle
-                x = paddle.to_tensor(5., stop_gradient=False)
-                for i in range(5):
-                    y = paddle.pow(x, 4.0)
-                    y.backward()
-                    print("{}: {}".format(i, x.grad))
-                # 0: [500.]
-                # 1: [1000.]
-                # 2: [1500.]
-                # 3: [2000.]
-                # 4: [2500.]
-
-                x.clear_grad()
-                print("{}".format(x.grad))
-                # 0.
-
-                grad_tensor=paddle.to_tensor(2.)
-                for i in range(5):
-                    y = paddle.pow(x, 4.0)
-                    y.backward(grad_tensor)
-                    print("{}: {}".format(i, x.grad))
-                # 0: [1000.]
-                # 1: [2000.]
-                # 2: [3000.]
-                # 3: [4000.]
-                # 4: [5000.]
-
-        """
-        if framework.in_dygraph_mode():
-            if grad_tensor is not None:
-                assert isinstance(
-                    grad_tensor, core.eager.EagerTensor
-                ), "The type of grad_tensor must be paddle.Tensor"
-                assert grad_tensor.shape == self.shape, \
-                    "Tensor shape not match, Tensor of grad_tensor [ {} ] with shape {} mismatch Tensor [ {} ] with shape {}".format(
-                    grad_tensor.name, grad_tensor.shape, self.name, self.shape)
-                grad_tensor = [grad_tensor]
-            else:
-                grad_tensor = []
-
-            if core.is_compiled_with_xpu() or core.is_compiled_with_npu():
-                # TODO(liuyuhui): Currently only for xpu. Will be removed in the future.
-                scaled_loss = scale_loss(self)
-                core.eager.run_backward([scaled_loss], grad_tensor,
-                                        retain_graph)
-            else:
-                core.eager.run_backward([self], grad_tensor, retain_graph)
-        else:
-            raise ValueError(
-                "Variable.backward() is only available in DyGraph mode")
-
-    @framework.dygraph_only
-    def gradient(self):
-        """
-        .. warning::
-          This API will be deprecated in the future, it is recommended to use
-          :code:`x.grad` which returns the tensor value of the gradient.
-
-        Get the Gradient of Current Tensor.
-
-        Returns:
-            ndarray: Numpy value of the gradient of current Tensor
-
-        Examples:
-            .. code-block:: python
-
-                import paddle
-
-                x = paddle.to_tensor(5., stop_gradient=False)
-                y = paddle.pow(x, 4.0)
-                y.backward()
-                print("grad of x: {}".format(x.gradient()))
-                # [500.]
-
-        """
-        if self.grad._is_initialized():
-            return self.grad.numpy()
-        else:
-            return None
-        # TODO(wanghuancoder) support SELECTED_ROWS
-
-    if hasattr(core, "eager"):
-        setattr(core.eager.EagerTensor, "__str__", __str__)
-        setattr(core.eager.EagerTensor, "backward", backward)
-        setattr(core.eager.EagerTensor, "gradient", gradient)

python/paddle/fluid/framework.py

@@ -77,6 +77,7 @@ _global_expected_place_ = None
 _current_device = None
 global_prog_seed = 0
 _current_pipeline_stage = None
+_already_patch_eager_tensor = False
 _global_flags_ = core.globals()
 core._disable_eager_mode()
 
@@ -85,6 +86,11 @@ core._disable_eager_mode()
 def _test_eager_guard(tracer=None):
     core._enable_eager_mode()
     _C_ops.switch_to_eager_ops()
+    global _already_patch_eager_tensor
+    if not _already_patch_eager_tensor:
+        from .dygraph.varbase_patch_methods import monkey_patch_varbase
+        monkey_patch_varbase()
+        _already_patch_eager_tensor = True
     if tracer is None:
         core._set_eager_tracer(_dygraph_tracer_)
     else:

python/paddle/fluid/initializer.py

@@ -604,7 +604,7 @@ class XavierInitializer(Initializer):
         if framework.in_dygraph_mode():
             if self._uniform:
                 limit = np.sqrt(6.0 / float(fan_in + fan_out))
-                out_var = _C_ops.uniform_random('shape', var.shape, 'min',
+                out_var = _C_ops.uniform_random('shape', out_var.shape, 'min',
                                                 -limit, 'max', limit, 'seed',
                                                 self._seed, 'dtype', out_dtype)
             else:

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

@@ -13,7 +13,6 @@
 # limitations under the License.
 
 import paddle.fluid.core as core
-import paddle.fluid.eager.eager_tensor_patch_methods as eager_tensor_patch_methods
 import paddle
 import numpy as np
 from paddle.fluid.framework import _test_eager_guard

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

@@ -13,7 +13,6 @@
 # limitations under the License.
 
 import paddle.fluid.core as core
-import paddle.fluid.eager.eager_tensor_patch_methods as eager_tensor_patch_methods
 import paddle
 import numpy as np
 from paddle.fluid.framework import _test_eager_guard, EagerParamBase, _in_eager_mode
@@ -621,7 +620,7 @@ class EagerTensorPropertiesTestCase(unittest.TestCase):
             self.assertTrue(np.array_equal(tensor.numpy(), arr))
             print("Test copy_")
             tensor.copy_(tensor1, True)
-            self.assertEqual(tensor.persistable, True)
+            self.assertEqual(tensor.persistable, False)
             self.assertEqual(tensor.shape, [4, 16])
             self.assertEqual(tensor.dtype, core.VarDesc.VarType.FP32)
             self.assertTrue(np.array_equal(tensor.numpy(), arr1))
@@ -764,6 +763,7 @@ class EagerParamBaseUsageTestCase(unittest.TestCase):
         self.assertTrue(np.array_equal(res3, res4))
 
     def test_backward_with_single_tensor(self):
+        with _test_eager_guard():
             arr4 = np.random.rand(4, 16, 16, 32).astype('float32')
             egr_tensor12 = core.eager.EagerTensor(arr4, core.CPUPlace())
             egr_tensor12.retain_grads()

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

@@ -267,7 +267,7 @@ class TestImperative(unittest.TestCase):
                 tmp = l1.weight * 2
                 self.assertTrue(tmp.stop_gradient)
             x = paddle.to_tensor(data)
-            y = l0(x) + tmp
+            y = paddle.add(l0(x), tmp)
             o = l1(y)
             o.backward()
 
@@ -285,7 +285,7 @@ class TestImperative(unittest.TestCase):
                 tmp = l1.weight * 2
                 self.assertTrue(tmp.stop_gradient)
             x = paddle.to_tensor(data)
-            y = l0(x) + tmp
+            y = paddle.add(l0(x), tmp)
             o = l1(y)
             o.backward()
 
@@ -306,7 +306,7 @@ class TestImperative(unittest.TestCase):
                 self.assertTrue(tmp.stop_gradient)
                 self.assertTrue(tmp2.stop_gradient is False)
             x = paddle.to_tensor(data)
-            y = l0(x) + tmp2
+            y = paddle.add(l0(x), tmp2)
             o = l1(y)
             o.backward()
 
@@ -329,7 +329,7 @@ class TestImperative(unittest.TestCase):
                 tmp = paddle.to_tensor(x)
                 tmp.stop_gradient = False
                 inputs.append(tmp)
-            ret = fluid.layers.sums(inputs)
+            ret = paddle.add_n(inputs)
             loss = fluid.layers.reduce_sum(ret)
             loss.backward()
         with fluid.dygraph.guard():

python/setup.py.in

@@ -308,7 +308,6 @@ packages=['paddle',
           'paddle.fluid.dygraph',
           'paddle.fluid.dygraph.dygraph_to_static',
           'paddle.fluid.dygraph.amp',
-          'paddle.fluid.eager',
           'paddle.fluid.proto',
           'paddle.fluid.proto.profiler',
           'paddle.fluid.distributed',