Support test imperative basic with fixed retain grad interface (#38548)

* 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

* support inference test

* refine test and fix initializer failed

* support create varbase and fix retain grad error

* fix windows error

* support test_imperative_basic test in eager mode

* remove additional log in variable.h

* remove additional log in variable.h

* remove additional code create in merge
Co-authored-by: Njim19930609 <jim19930609@gmail.com>
Co-authored-by: NWang Huan <wanghuan29@baidu.com>

paddle/fluid/eager/api/utils/hook_utils.cc

@@ -43,25 +43,25 @@ void RegisterReduceHookForTensor(const egr::EagerTensor& tensor,
 void RetainGradForTensor(const egr::EagerTensor& tensor) {
   // TODO(jiabin): Support More Tensor type here
   AutogradMeta* meta = EagerUtils::unsafe_autograd_meta(tensor);
-  egr::EagerTensor* grad_tensor = meta->MutableGrad();
+  std::weak_ptr<egr::EagerTensor> weak_grad_tensor = meta->WeakGrad();
 
   // Define Hook
   std::function<egr::EagerTensor(const egr::EagerTensor&)> hook =
-      [grad_tensor](const egr::EagerTensor& t) {
-        if (!grad_tensor) {
-          PADDLE_THROW(paddle::platform::errors::Fatal(
-              "Detected null grad_tensor."
-              "Grad tensor in AutogradMeta of should not be nullptr"));
-        }
+      [weak_grad_tensor](const egr::EagerTensor& t) {
+        if (!weak_grad_tensor.expired()) {
+          auto grad_tensor = weak_grad_tensor.lock();
           if (t.defined()) {
+            VLOG(7) << "Set impl for RetainGrad Hook for tensor: " << t.name();
             // Simply Copy impl() to grad_tensor
             grad_tensor->set_impl(t.impl());
-          return *grad_tensor;
+            return *grad_tensor.get();
           } else {
+            VLOG(7) << "Set Var for RetainGrad Hook for tensor: " << t.name();
             PADDLE_ENFORCE_EQ(
                 t.Var().IsInitialized(), true,
                 paddle::platform::errors::Fatal(
-                  "Detected uninitialized variable, causing segmentation fault "
+                    "Detected uninitialized variable, causing segmentation "
+                    "fault "
                     "inside the hook."
                     "Variable %s has to be initialized while we need to set it."
                     "please check tensor initialization status.",
@@ -69,7 +69,11 @@ void RetainGradForTensor(const egr::EagerTensor& tensor) {
             grad_tensor->MutableVar()
                 ->GetMutable<paddle::framework::LoDTensor>()
                 ->ShareDataWith(t.Var().Get<paddle::framework::LoDTensor>());
-          return *grad_tensor;
+            return *grad_tensor.get();
+          }
+        } else {
+          VLOG(7) << "Retain NULL EagerTensor in Grad Hook";
+          return EagerTensor();
         }
       };
 

paddle/fluid/eager/auto_code_generator/eager_generator.cc

@@ -972,11 +972,16 @@ static std::string GenerateGradNodeCreationContent(
         iter.GetGradInsFwdSlotnameMap();
     for (auto& kv : grad_ins_fwd_slotname_map) {
       const std::string& tensor_wrapper_name = kv.second;
+      std::string full_reserved = "false";
+      if (fwd_outputs_name_pos_map.find(tensor_wrapper_name) ==
+          fwd_outputs_name_pos_map.end()) {
+        full_reserved = "true";
+      }
       const char* SET_TENSOR_WRAPPER_TEMPLATE =
-          "    grad_node->SetTensorWrapper%s(%s);\n";
-      grad_node_creation_str +=
-          paddle::string::Sprintf(SET_TENSOR_WRAPPER_TEMPLATE,
-                                  tensor_wrapper_name, tensor_wrapper_name);
+          "    grad_node->SetTensorWrapper%s(%s, %s);\n";
+      grad_node_creation_str += paddle::string::Sprintf(
+          SET_TENSOR_WRAPPER_TEMPLATE, tensor_wrapper_name, tensor_wrapper_name,
+          full_reserved);
     }
   }
   grad_node_creation_str += "\n";
@@ -1017,11 +1022,6 @@ static std::string GenerateGradNodeCreationContent(
       grad_node_creation_str += paddle::string::Sprintf(
           ADD_EDGES_TEMPLATE, input_autograd_name, input_position);
     }
-    VLOG(6) << "Generated Call RetainGradForTensor";
-    const char* RETAIN_GRAD_TEMPLATE =
-        "    egr::EagerUtils::CheckAndRetainGrad(%s);\n";
-    grad_node_creation_str +=
-        paddle::string::Sprintf(RETAIN_GRAD_TEMPLATE, input_name);
   }
 
   // [GradOpNode] SetGradInMeta
@@ -1048,6 +1048,12 @@ static std::string GenerateGradNodeCreationContent(
         "    egr::EagerUtils::SetHistory(&%s, grad_node);\n";
     grad_node_creation_str +=
         paddle::string::Sprintf(SET_HISTORY_TEMPLATE, output_autograd_name);
+
+    VLOG(6) << "Generated Call RetainGradForTensor";
+    const char* RETAIN_GRAD_TEMPLATE =
+        "    egr::EagerUtils::CheckAndRetainGrad(%s);\n";
+    grad_node_creation_str +=
+        paddle::string::Sprintf(RETAIN_GRAD_TEMPLATE, output_name);
   }
   VLOG(6) << "Generated SetGradIn/OutMeta";
 
@@ -1771,6 +1777,7 @@ static std::string GenerateGradNodeHeaderContents(
 
       std::string tensor_wrapper_arg_str;
       std::string tensor_wrapper_body_str;
+      std::string full_reserved_str = "full_reserved";
       if (duplicable_tensors.count(tensor_wrapper_name)) {
         const char* ATTR_TENSOR_WRAPPER_ARG_TEMPLATE =
             "const std::vector<egr::EagerTensor>& %s";
@@ -1803,17 +1810,18 @@ static std::string GenerateGradNodeHeaderContents(
             TENSOR_WRAPPER_MEMBER_TEMPLATE, struct_tensor_wrapper_name);
 
         const char* SET_TENSOR_WRAPPER_BODY_TEMPLATE =
-            "%s = egr::TensorWrapper(%s, true /*full_reserved*/);";
+            "%s = egr::TensorWrapper(%s, %s /*full_reserved*/);";
         tensor_wrapper_body_str = paddle::string::Sprintf(
             SET_TENSOR_WRAPPER_BODY_TEMPLATE, struct_tensor_wrapper_name,
-            tensor_wrapper_name);
+            tensor_wrapper_name, full_reserved_str);
       }
-
+      std::string full_reserved_signature_str = "bool full_reserved";
       const char* SET_TENSOR_WRAPPER_TEMPLATE =
-          "   void SetTensorWrapper%s(%s) {\n     %s\n   }\n";
+          "   void SetTensorWrapper%s(%s, %s) {\n     %s\n   }\n";
       set_tensor_wrappers_str += paddle::string::Sprintf(
           SET_TENSOR_WRAPPER_TEMPLATE, tensor_wrapper_name,
-          tensor_wrapper_arg_str, tensor_wrapper_body_str);
+          tensor_wrapper_arg_str, full_reserved_signature_str,
+          tensor_wrapper_body_str);
     }
   }
   VLOG(6) << "Generated TensorWrapper";

paddle/fluid/eager/autograd_meta.h

@@ -14,8 +14,8 @@
 
 #pragma once
 
+#include "paddle/fluid/eager/api/utils/global_utils.h"
 #include "paddle/fluid/eager/grad_node_info.h"
-
 namespace egr {
 
 using AbstractAutogradMeta = paddle::experimental::AbstractAutogradMeta;
@@ -75,9 +75,20 @@ class AutogradMeta : public AbstractAutogradMeta {
 
   ~AutogradMeta() override = default;
 
-  const egr::EagerTensor& Grad() const { return grad_; }
+  const egr::EagerTensor& Grad() const {
+    PADDLE_ENFORCE_NOT_NULL(
+        grad_.get(),
+        paddle::platform::errors::InvalidArgument(
+            "Should Not get NULL from Grad pointer, since "
+            "we should have default EagerTensor once we init AutoGradMeta. "
+            "if you got this error may indicates framework error in "
+            "PaddlePaddle"));
+    return *(grad_.get());
+  }
+
+  egr::EagerTensor* MutableGrad() { return grad_.get(); }
 
-  egr::EagerTensor* MutableGrad() { return &grad_; }
+  std::weak_ptr<egr::EagerTensor> WeakGrad() { return grad_; }
 
   void SetGradNode(const std::shared_ptr<GradNodeBase>& grad_node) {
     PADDLE_ENFORCE_NOT_NULL(
@@ -126,12 +137,13 @@ class AutogradMeta : public AbstractAutogradMeta {
 
  private:
   // TODO(jiabin) :Should we use pointer instead of object?
-  egr::EagerTensor grad_;
+  std::shared_ptr<egr::EagerTensor> grad_{std::make_shared<egr::EagerTensor>(
+      egr::Controller::Instance().GenerateUniqueName("@grad"))};
 
   // GradNodeBase is base class of all grad op which is a
   // wrapper for grad op. This class will make grad op easy
   // to be traced.
-  std::shared_ptr<GradNodeBase> grad_node_;
+  std::shared_ptr<GradNodeBase> grad_node_ = nullptr;
 
   /**
    * Why we need slot id here?

paddle/fluid/eager/backward.cc

@@ -71,6 +71,14 @@ std::unordered_map<GradNodeBase*, int> getInDegreeMap(
   return node_in_degree_map;
 }
 
+void RunBackwardHooks(
+    const std::vector<std::vector<egr::EagerTensor>>& grad_tensors,
+    egr::GradNodeBase* grad_node) {
+  grad_node->ApplyGradientHooks(grad_tensors);
+  VLOG(6) << "Apply Reduce Hooks for node";
+  grad_node->ApplyReduceHooks();
+}
+
 void RunBackward(const std::vector<egr::EagerTensor>& tensors,
                  const std::vector<egr::EagerTensor>& grad_tensors,
                  bool retain_graph) {
@@ -157,7 +165,11 @@ void RunBackward(const std::vector<egr::EagerTensor>& tensors,
     std::unique_ptr<GradTensorHolder> node_input_buffer =
         std::move(node_input_buffers_dict[node]);
     VLOG(6) << "Run Backward Kernel with input_buffer";
-    // Run Backward Node and get outputs
+
+    RunBackwardHooks(node_input_buffer->Buffers(), node);
+    // TODO(jiabin): Support post hook here and make hook run in seperate
+    // operator
+    // Run Pre Backward Node and get outputs
     std::vector<std::vector<egr::EagerTensor>> grad_output_tensors =
         (*node)(node_input_buffer->Buffers());
     // TODO(jiabin): Should we erase it or find a more efficient way.

paddle/fluid/eager/grad_node_info.cc

@@ -47,16 +47,20 @@ void GradNodeBase::AddEdges(std::vector<AutogradMeta*>* metas, size_t slot_id) {
     // adj_edges has as same rank as fwd inputs, and record it's output rank
     // from
     // its pre-ops
+    if (meta) {
       auto node = meta->GetMutableGradNode();
       if (node) {
         adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
                                          meta->OutRankInfo());
       } else {
+        if (!meta->StopGradient()) {
           meta->SetGradNode(std::make_shared<egr::GradNodeAccumulation>());
           adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
                                            meta->OutRankInfo());
         }
       }
+    }
+  }
 }
 
 void GradNodeBase::AddEdges(const std::vector<AutogradMeta*>& metas,
@@ -71,16 +75,20 @@ void GradNodeBase::AddEdges(const std::vector<AutogradMeta*>& metas,
     // adj_edges has as same rank as fwd inputs, and record it's output rank
     // from
     // its pre-ops
+    if (meta) {
       auto node = meta->GetMutableGradNode();
       if (node) {
         adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
                                          meta->OutRankInfo());
       } else {
+        if (!meta->StopGradient()) {
           meta->SetGradNode(std::make_shared<egr::GradNodeAccumulation>());
           adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
                                            meta->OutRankInfo());
         }
       }
+    }
+  }
 }
 
 void GradNodeBase::AddEdges(AutogradMeta* meta, size_t slot_id) {
@@ -90,15 +98,19 @@ void GradNodeBase::AddEdges(AutogradMeta* meta, size_t slot_id) {
           "Given slot id is out of range of adj_edges outter size, "
           "adj_edges is designed to has the same size of grad "
           "inputs's slot num."));
+  if (meta) {
     auto node = meta->GetMutableGradNode();
     if (node) {
       adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
                                        meta->OutRankInfo());
     } else {
+      if (!meta->StopGradient()) {
         meta->SetGradNode(std::make_shared<egr::GradNodeAccumulation>());
         adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
                                          meta->OutRankInfo());
       }
+    }
+  }
 }
 
 const std::vector<GradSlotMeta>& GradNodeBase::InputMeta() const {
@@ -127,6 +139,11 @@ void GradNodeBase::SetGradInMeta(const std::vector<AutogradMeta*>& fwd_out,
   // Init stop gradient vector before use to avoid push back
   meta.Init(slot_size);
   for (size_t i = 0; i < slot_size; i++) {
+    PADDLE_ENFORCE_NOT_NULL(fwd_out[i],
+                            paddle::platform::errors::PreconditionNotMet(
+                                "Bwd_in_meta should only be called while "
+                                "autograd_meta is not null. If you got this "
+                                "error, it indicates bugs in framework."));
     if (fwd_out[i]->StopGradient()) {
       // Set Stop Gradient only when its true or non-initialized autograd_meta,
       // since all default value is false.
@@ -173,6 +190,10 @@ void GradNodeBase::SetGradOutMeta(const std::vector<AutogradMeta*>& fwd_in,
   // Init stop gradient vector before use to avoid push back
   meta.Init(slot_size);
   for (size_t i = 0; i < slot_size; i++) {
+    if (!fwd_in[i]) {
+      meta.SetStopGradient(i, true);
+      continue;
+    }
     if (fwd_in[i]->StopGradient()) {
       // Set Stop Gradient only when its true or non-initialized autograd_meta,
       // since all default value is false.
@@ -249,6 +270,7 @@ std::vector<std::vector<egr::EagerTensor>> GradNodeBase::ApplyGradientHooks(
     slot_out.resize(tensors[slot_id].size());
     egr::EagerTensor& out = slot_out[rank];
     if (!out.defined() || !out.initialized()) {
+      VLOG(8) << "Run Hook for tensor: " << tensors[slot_id][rank].name();
       out = hook(tensors[slot_id][rank]);
     } else {
       // TODO(jiabin): Why this?

paddle/fluid/eager/utils.cc

@@ -266,6 +266,7 @@ std::vector<EagerTensor> EagerUtils::RecoverTensorWrapper(
 void EagerUtils::CheckAndRetainGrad(const egr::EagerTensor& tensor) {
   VLOG(6) << "Check RetainGradForTensor: " << tensor.name();
   if (FLAGS_retain_grad_for_all_tensor) {
+    VLOG(6) << "RetainGradForTensor: " << tensor.name();
     egr::egr_utils_api::RetainGradForTensor(tensor);
   }
 }
@@ -274,7 +275,7 @@ void EagerUtils::CheckAndRetainGrad(
     const std::vector<egr::EagerTensor>& tensors) {
   if (FLAGS_retain_grad_for_all_tensor) {
     for (auto& tensor : tensors) {
-      VLOG(6) << "Check RetainGradForTensor: " << tensor.name();
+      VLOG(6) << "RetainGradForTensor: " << tensor.name();
       egr::egr_utils_api::RetainGradForTensor(tensor);
     }
   }

paddle/fluid/pybind/eager.cc

@@ -62,6 +62,13 @@ void EmptyEagerTensorInitializer(
     const std::vector<int>& dims = {},
     framework::proto::VarType::Type var_type =
         paddle::framework::proto::VarType::LOD_TENSOR) {
+  auto ddims = paddle::framework::make_ddim(dims);
+  PADDLE_ENFORCE_GE(
+      paddle::framework::product(ddims), 0,
+      paddle::platform::errors::InvalidArgument(
+          "Create Eager Tensor with dims contain minus num is ilegal"
+          "Please check your code and make sure you new a "
+          "eager tensor with fixed shape instead of using -1."));
   self->eager_tensor.set_name(name);
   auto autograd_meta = egr::EagerUtils::autograd_meta(&(self->eager_tensor));
   autograd_meta->SetPersistable(persistable);
@@ -71,8 +78,7 @@ void EmptyEagerTensorInitializer(
     std::shared_ptr<pten::DenseTensor> dense_tensor =
         std::make_shared<pten::DenseTensor>(
             pten::make_intrusive<paddle::experimental::SharedStorage>(place),
-            pten::DenseTensorMeta(pten::TransToPtenDataType(dtype),
-                                  paddle::framework::make_ddim(dims)));
+            pten::DenseTensorMeta(pten::TransToPtenDataType(dtype), ddims));
     self->eager_tensor.set_impl(dense_tensor);
   } else {
     PADDLE_THROW(platform::errors::InvalidArgument(

paddle/fluid/pybind/eager_method.cc

@@ -39,10 +39,12 @@ extern PyTypeObject* pEagerTensorType;
 static PyObject* eager_tensor_method_numpy(EagerTensorObject* self,
                                            PyObject* args, PyObject* kwargs) {
   EAGER_SYNC_TRY
-  if (!self->eager_tensor.initialized()) {
-    Py_INCREF(Py_None);
-    return Py_None;
-  }
+  PADDLE_ENFORCE_EQ(
+      self->eager_tensor.initialized(), true,
+      platform::errors::InvalidArgument(
+          "Tensor data of %s is Empty that indicates we have null tensor for "
+          "now, please check if it has no data and initialize it first.",
+          self->eager_tensor.name()));
   auto tensor_dims = self->eager_tensor.shape();
   auto numpy_dtype = TensorDtype2NumpyDtype(self->eager_tensor.type());
   auto sizeof_dtype = pten::DataTypeSize(self->eager_tensor.type());

paddle/fluid/pybind/eager_properties.cc

@@ -75,6 +75,7 @@ PyObject* eager_tensor_properties_get_grad(EagerTensorObject* self,
         std::dynamic_pointer_cast<egr::GradNodeAccumulation>(grad_node);
     return ToPyObject(accumulation_grad_node->Grad());
   } else {
+    VLOG(6) << "Get grad for tensor: " << self->eager_tensor.name();
     auto meta = egr::EagerUtils::unsafe_autograd_meta(self->eager_tensor);
     return ToPyObject(meta->Grad());
   }

python/paddle/fluid/eager/eager_tensor_patch_methods.py

@@ -123,10 +123,11 @@ def monkey_patch_eagertensor():
                 # [500.]
 
         """
-        if self.grad is None:
+        if self.grad._is_initialized():
+            return self.grad.numpy()
+        else:
             return None
         # TODO(wanghuancoder) support SELECTED_ROWS
-        return self.grad.numpy()
 
     if hasattr(core, "eager"):
         setattr(core.eager.EagerTensor, "__str__", __str__)

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

@@ -452,6 +452,22 @@ class EagerParamBaseUsageTestCase(unittest.TestCase):
         self.assertTrue(np.array_equal(res1, res2))
         self.assertTrue(np.array_equal(res3, res4))
 
+    def test_backward_with_single_tensor(self):
+        arr4 = np.random.rand(4, 16, 16, 32).astype('float32')
+        egr_tensor12 = core.eager.EagerTensor(arr4, core.CPUPlace())
+        egr_tensor12.retain_grads()
+        arr = np.ones([4, 16, 16, 32]).astype('float32')
+        self.assertEqual(egr_tensor12.persistable, False)
+        self.assertTrue("generated_tensor" in egr_tensor12.name)
+        self.assertEqual(egr_tensor12.shape, [4, 16, 16, 32])
+        self.assertEqual(egr_tensor12.dtype, core.VarDesc.VarType.FP32)
+        self.assertEqual(egr_tensor12.stop_gradient, True)
+        self.assertTrue(egr_tensor12.place._equals(paddle.fluid.CPUPlace()))
+        self.assertTrue(np.array_equal(egr_tensor12.numpy(), arr4))
+        self.assertTrue(np.array_equal(egr_tensor12.gradient(), None))
+        egr_tensor12.backward()
+        self.assertTrue(np.array_equal(egr_tensor12.gradient(), arr))
+
 
 class EagerGuardTestCase(unittest.TestCase):
     def test__test_eager_guard(self):

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

@@ -24,7 +24,7 @@ from test_imperative_base import new_program_scope
 import paddle.fluid.dygraph_utils as dygraph_utils
 from paddle.fluid.dygraph.layer_object_helper import LayerObjectHelper
 import paddle
-from paddle.fluid.framework import _test_eager_guard, _in_eager_mode
+from paddle.fluid.framework import _test_eager_guard, _in_eager_mode, in_dygraph_mode
 
 
 class MyLayer(fluid.Layer):
@@ -94,58 +94,13 @@ class SimpleRNNCell(fluid.Layer):
             is_bias=False)
 
     def forward(self, input, pre_hidden):
-        tmp_i2h = self.create_variable(dtype=self._dtype)
-        tmp_h2h = self.create_variable(dtype=self._dtype)
-        hidden = self.create_variable(dtype=self._dtype)
-        out = self.create_variable(dtype=self._dtype)
-        softmax_out = self.create_variable(dtype=self._dtype)
-        reduce_out = self.create_variable(dtype=self._dtype)
-        self._helper.append_op(
-            type="mul",
-            inputs={"X": input,
-                    "Y": self._i2h_w},
-            outputs={"Out": tmp_i2h},
-            attrs={"x_num_col_dims": 1,
-                   "y_num_col_dims": 1})
-
-        self._helper.append_op(
-            type="mul",
-            inputs={"X": pre_hidden,
-                    "Y": self._h2h_w},
-            outputs={"Out": tmp_h2h},
-            attrs={"x_num_col_dims": 1,
-                   "y_num_col_dims": 1})
-
-        self._helper.append_op(
-            type="elementwise_add",
-            inputs={'X': tmp_h2h,
-                    'Y': tmp_i2h},
-            outputs={'Out': hidden},
-            attrs={'axis': -1,
-                   'use_mkldnn': False})
+        tmp_i2h = paddle.fluid.layers.nn.mul(input, self._i2h_w)
+        tmp_h2h = paddle.fluid.layers.nn.mul(pre_hidden, self._h2h_w)
+        hidden = paddle.add(tmp_h2h, tmp_i2h)
         hidden = self._helper.append_activation(hidden, act='tanh')
-
-        self._helper.append_op(
-            type="mul",
-            inputs={"X": hidden,
-                    "Y": self._h2o_w},
-            outputs={"Out": out},
-            attrs={"x_num_col_dims": 1,
-                   "y_num_col_dims": 1})
-
-        self._helper.append_op(
-            type="softmax",
-            inputs={"X": out},
-            outputs={"Out": softmax_out},
-            attrs={"use_cudnn": False})
-
-        self._helper.append_op(
-            type='reduce_sum',
-            inputs={'X': softmax_out},
-            outputs={'Out': reduce_out},
-            attrs={'keep_dim': False,
-                   'reduce_all': True})
-
+        out = paddle.fluid.layers.nn.mul(hidden, self._h2o_w)
+        softmax_out = paddle.nn.functional.softmax(out)
+        reduce_out = paddle.fluid.layers.nn.reduce_sum(softmax_out)
         return reduce_out, hidden
 
 
@@ -394,12 +349,17 @@ class TestImperative(unittest.TestCase):
             a = inputs2[0].gradient()
             self.assertTrue(np.allclose(inputs2[0].gradient(), x))
 
-    def test_empty_var(self):
+    def func_empty_var(self):
         with fluid.dygraph.guard():
             cur_program = fluid.Program()
             cur_block = cur_program.current_block()
+            # Normally, we don't allow tensor with -1 shape being created in dygraph mode, this test is not good.
+            if not _in_eager_mode():
                 new_variable = cur_block.create_var(
                     name="X", shape=[-1, 23, 48], dtype='float32')
+            else:
+                new_variable = cur_block.create_var(
+                    name="X", shape=[1, 23, 48], dtype='float32')
             try:
                 new_variable.numpy()
             except Exception as e:
@@ -409,21 +369,25 @@ class TestImperative(unittest.TestCase):
                 new_variable.backward()
             except Exception as e:
                 assert type(e) == core.EnforceNotMet
-
+            # TODO(jiabin): Support clear_gradient in eager mode later and remove this if statement
+            if not _in_eager_mode():
                 try:
                     new_variable.clear_gradient()
                 except Exception as e:
                     assert type(e) == core.EnforceNotMet
 
-    def test_empty_grad(self):
+    def test_empty_var(self):
+        with _test_eager_guard():
+            self.func_empty_var()
+        self.func_empty_var()
+
+    def func_empty_grad(self):
         with fluid.dygraph.guard():
             x = np.ones([2, 2], np.float32)
             new_var = paddle.to_tensor(x)
-            try:
-                new_var.gradient()
-            except Exception as e:
-                assert type(e) == ValueError
-
+            self.assertIsNone(new_var.gradient())
+            # TODO(jiabin): Support clear_gradient in eager mode later and remove this if statement
+            if not _in_eager_mode():
                 try:
                     new_var.clear_gradient()
                 except Exception as e:
@@ -432,14 +396,24 @@ class TestImperative(unittest.TestCase):
         with fluid.dygraph.guard():
             cur_program = fluid.Program()
             cur_block = cur_program.current_block()
+            # Normally, we don't allow tensor with -1 shape being created in dygraph mode, this test is not good.
+            if not _in_eager_mode():
                 new_variable = cur_block.create_var(
                     name="X", shape=[-1, 23, 48], dtype='float32')
+            else:
+                new_variable = cur_block.create_var(
+                    name="X", shape=[1, 23, 48], dtype='float32')
             try:
                 new_variable.gradient()
             except Exception as e:
                 assert type(e) == ValueError
 
-    def test_set_persistable(self):
+    def test_empty_grad(self):
+        with _test_eager_guard():
+            self.func_empty_grad()
+        self.func_empty_grad()
+
+    def func_set_persistable(self):
         with fluid.dygraph.guard():
             x = np.ones([2, 2], np.float32)
             new_var = paddle.to_tensor(x)
@@ -447,12 +421,22 @@ class TestImperative(unittest.TestCase):
             new_var.persistable = True
             self.assertTrue(new_var.persistable)
 
-    def test_layer(self):
+    def test_set_persistable(self):
+        with _test_eager_guard():
+            self.func_set_persistable()
+        self.func_set_persistable()
+
+    def func_layer(self):
         with fluid.dygraph.guard():
             l = fluid.Layer("l")
             self.assertRaises(NotImplementedError, l.forward, [])
 
-    def test_layer_in_out(self):
+    def test_layer(self):
+        with _test_eager_guard():
+            self.func_layer()
+        self.func_layer()
+
+    def func_layer_in_out(self):
         np_inp = np.array([1.0, 2.0, -1.0], dtype=np.float32)
         with fluid.dygraph.guard():
             var_inp = paddle.to_tensor(np_inp)
@@ -489,12 +473,17 @@ class TestImperative(unittest.TestCase):
                 feed={inp.name: np_inp},
                 fetch_list=[x.name, param_grads[1].name])
 
-        self.assertTrue(np.allclose(dy_out, static_out))
-        self.assertTrue(np.allclose(dy_grad, static_grad))
-        self.assertTrue(np.allclose(dy_out2, static_out))
-        self.assertTrue(np.allclose(dy_grad2, static_grad))
+        self.assertTrue(np.array_equal(dy_out, static_out))
+        self.assertTrue(np.array_equal(dy_grad, static_grad))
+        self.assertTrue(np.array_equal(dy_out2, static_out))
+        self.assertTrue(np.array_equal(dy_grad2, static_grad))
 
-    def test_mlp(self):
+    def test_layer_in_out(self):
+        with _test_eager_guard():
+            self.func_layer_in_out()
+        self.func_layer_in_out()
+
+    def func_mlp(self):
         np_inp = np.array([[1.0, 2.0], [3.0, 4.0]], dtype=np.float32)
         with fluid.dygraph.guard():
             var_inp = paddle.to_tensor(np_inp)
@@ -545,6 +534,11 @@ class TestImperative(unittest.TestCase):
         self.assertEqual(mlp._linear2, sublayers[1])
         self.assertEqual(len(sublayers), 2)
 
+    def test_mlp(self):
+        with _test_eager_guard():
+            self.func_mlp()
+        self.func_mlp()
+
     def test_gradient_accumulation(self):
         def test_single_api(sort_sum_gradient):
             fluid.set_flags({'FLAGS_sort_sum_gradient': sort_sum_gradient})
@@ -677,7 +671,7 @@ class TestImperative(unittest.TestCase):
             test_mlp(False)
             test_mlp(True)
 
-    def test_dygraph_vs_static(self):
+    def func_dygraph_vs_static(self):
         np_inp1 = np.random.rand(4, 3, 3)
         np_inp2 = np.random.rand(4, 3, 3)
 
@@ -728,7 +722,12 @@ class TestImperative(unittest.TestCase):
                                     fetch_list=out)[0]
         self.assertTrue(np.allclose(dygraph_result, static_result))
 
-    def test_rnn(self):
+    def test_dygraph_vs_static(self):
+        with _test_eager_guard():
+            self.func_dygraph_vs_static()
+        self.func_dygraph_vs_static()
+
+    def func_rnn(self):
         np_inp = np.array([[1.0, 2.0, 3.0], [4.0, 5.0, 6.0], [7.0, 8.0, 9.0],
                            [10.0, 11.0, 12.0]])
         np_inp = np_inp.reshape((1, 4, 3))
@@ -771,14 +770,19 @@ class TestImperative(unittest.TestCase):
                     param_grads[1][1].name, param_grads[2][1].name
                 ])
 
-        self.assertTrue(np.allclose(dy_out, static_out))
-        self.assertTrue(np.allclose(dy_grad_h2o, static_grad_h2o))
-        self.assertTrue(np.allclose(dy_grad_h2h, static_grad_h2h))
-        self.assertTrue(np.allclose(dy_grad_i2h, static_grad_i2h))
-        self.assertTrue(np.allclose(dy_out2, static_out))
-        self.assertTrue(np.allclose(dy_grad_h2o2, static_grad_h2o))
-        self.assertTrue(np.allclose(dy_grad_h2h2, static_grad_h2h))
-        self.assertTrue(np.allclose(dy_grad_i2h2, static_grad_i2h))
+        self.assertTrue(np.array_equal(dy_out, static_out))
+        self.assertTrue(np.array_equal(dy_grad_h2o, static_grad_h2o))
+        self.assertTrue(np.array_equal(dy_grad_h2h, static_grad_h2h))
+        self.assertTrue(np.array_equal(dy_grad_i2h, static_grad_i2h))
+        self.assertTrue(np.array_equal(dy_out2, static_out))
+        self.assertTrue(np.array_equal(dy_grad_h2o2, static_grad_h2o))
+        self.assertTrue(np.array_equal(dy_grad_h2h2, static_grad_h2h))
+        self.assertTrue(np.array_equal(dy_grad_i2h2, static_grad_i2h))
+
+    def test_rnn(self):
+        with _test_eager_guard():
+            self.func_rnn()
+        self.func_rnn()
 
     def func_layer_attrs(self):
         layer = fluid.dygraph.Layer("test")