Support slim eager (#39874)

* eager, test=develop

* fix bug, test=develop

* eager, test=develop

* merge legacy to fluid

* eager, test=develop

* eager, test=develop

* Refactor TensorAdd func by template and remove gradient_accumulation in eager

* Remove needless target name

* eager, test=develop

* eager, test=develop

* Use overload instead of template

* Remove legacy code

* Remove legacy code

* selectedrows, test=develop

* Remove DataType test

* eager, test=develop

* eager, test=develop

* support gan, test=develop

* Using Tensor directly instead of using EagerTensor

* support gradient_accumulation

* make test_imperative_lod_tensor_to_selected_rows longer

* make test_imperative_lod_tensor_to_selected_rows longer

* refine code

* ptb, test=develop

* Rename all EagerTensor to Tensor

* Rename some EagerTensor to Tensor

* rename EagerTensor to EagerVariable

* eager, test=develop

* eager, test=develop

* eager, test=develop

* eager, test=develop

* add more test

* eager, test=develop

* Support copiable selected rows and merge develop

* save load, eager, test=develop

* save load, eager, test=develop

* refine, test=develop

* remove useless _set_value method

* refine, test=develop

* refine, test=develop

* revert static_runner, test=develop

* EagerTensor to Tensor, test=develop

* refine, test=develop

* refine, test=develop

* clear grad, test=develop

* merge, develop

* merge, develop

* merge, test=develop

* merge, test=develop

* Support quant and part of slice

* support legacy static save

* extend slim tests time

* remove imperative on inference

* remove imperative on inference

* merge develop

* fix typo

* fix typo

* split slice related code into 2 part for imperative and eager

* split slice from inference

* split slice from inference

* fix test_tensor_register_hook
Co-authored-by: NWang Huan <wanghuan29@baidu.com>
Co-authored-by: NWeilong Wu <veyron_wu@163.com>
Co-authored-by: Nwanghuancoder <wanghuancoder@163.com>

paddle/fluid/eager/accumulation/accumulation_node.h

@@ -24,11 +24,14 @@ class GradNodeAccumulation : public GradNodeBase {
  public:
   // Constructor: configure fwd input tensors to grad node
   explicit GradNodeAccumulation(AutogradMeta* meta) : GradNodeBase(1, 1) {
+    VLOG(6) << "Construct GradNodeAccumulation";
     weak_grad_ = meta->WeakGrad();
     SetDefaultGradInOutMeta();
   }
 
-  ~GradNodeAccumulation() override = default;
+  ~GradNodeAccumulation() override {
+    VLOG(6) << "Destruct GradNodeAccumulation";
+  }
 
   // Functor: perform backward computations
   virtual std::vector<std::vector<paddle::experimental::Tensor>> operator()(

paddle/fluid/eager/api/generated/eager_generated/backwards/scale_node.h

@@ -46,7 +46,7 @@ class GradNodeScale : public GradNodeBase {
       const std::vector<paddle::experimental::Tensor>& tensors);
 
   void SetAttributes_scale(float scale);
-
+  std::string name() override { return ""; }
   // Members: define fwd input tensors
   // For Scale there is no fwd input tensor needed
  private:

paddle/fluid/eager/auto_code_generator/eager_generator.cc

@@ -996,6 +996,29 @@ static std::string GenerateGradNodeCreationContent(
   // then generate: "egr::AutogradMeta* p_autograd_out =
   // egr::EagerUtils::autograd_meta("op_proto->outputs()[0].name()")"
   std::string get_autograd_meta_str = "  // Prepare Autograd Meta \n";
+  // If single output slotname and not duplicable,
+  // then generate: "egr::AutogradMeta* p_autograd_out =
+  // egr::EagerUtils::autograd_meta("op_proto.outputs()[0].name()")"
+  for (const proto::OpProto::Var& output : out_vars) {
+    const std::string& output_name = output.name();
+    const std::string& output_autograd_name = "p_autograd_" + output_name;
+
+    if (output.duplicable()) {
+      const char* GET_MULTI_AUTOGRAD_META_TEMPLATE =
+          "  std::vector<egr::AutogradMeta*> %s = "
+          "egr::EagerUtils::autograd_meta(&%s);\n";
+      get_autograd_meta_str += paddle::string::Sprintf(
+          GET_MULTI_AUTOGRAD_META_TEMPLATE, output_autograd_name, output_name);
+    } else {
+      const char* GET_SINGLE_AUTOGRAD_META_TEMPLATE =
+          "  egr::AutogradMeta* %s = "
+          "egr::EagerUtils::autograd_meta(&%s);\n";
+      get_autograd_meta_str += paddle::string::Sprintf(
+          GET_SINGLE_AUTOGRAD_META_TEMPLATE, output_autograd_name, output_name);
+    }
+  }
+  VLOG(6) << "Generated outputs autograd_meta";
+
   for (const proto::OpProto::Var& input : in_vars) {
     const std::string& input_name = input.name();
     const std::string& input_autograd_name = "p_autograd_" + input_name;
@@ -1024,31 +1047,6 @@ static std::string GenerateGradNodeCreationContent(
   }
   VLOG(6) << "Generated inputs autograd_meta";
 
-  // If single output slotname and not duplicable,
-  // then generate: "egr::AutogradMeta* p_autograd_out =
-  // egr::EagerUtils::autograd_meta("op_proto.outputs()[0].name()")"
-  for (const proto::OpProto::Var& output : out_vars) {
-    const std::string& output_name = output.name();
-    const std::string& output_autograd_name = "p_autograd_" + output_name;
-
-    // Skip Intermediate Tensor
-
-    if (output.duplicable()) {
-      const char* GET_MULTI_AUTOGRAD_META_TEMPLATE =
-          "  std::vector<egr::AutogradMeta*> %s = "
-          "egr::EagerUtils::autograd_meta(&%s);\n";
-      get_autograd_meta_str += paddle::string::Sprintf(
-          GET_MULTI_AUTOGRAD_META_TEMPLATE, output_autograd_name, output_name);
-    } else {
-      const char* GET_SINGLE_AUTOGRAD_META_TEMPLATE =
-          "  egr::AutogradMeta* %s = "
-          "egr::EagerUtils::autograd_meta(&%s);\n";
-      get_autograd_meta_str += paddle::string::Sprintf(
-          GET_SINGLE_AUTOGRAD_META_TEMPLATE, output_autograd_name, output_name);
-    }
-  }
-  VLOG(6) << "Generated outputs autograd_meta";
-
   std::string prepare_autograd_meta_str = "";
   prepare_autograd_meta_str += get_autograd_meta_str;
   prepare_autograd_meta_str += "\n";
@@ -1204,11 +1202,12 @@ static std::string GenerateGradNodeCreationContent(
       "  %s"
       "  bool require_any_grad = egr::EagerUtils::ComputeRequireGrad(%s);\n"
       "  if(require_any_grad) {\n"
+      "    VLOG(6) << \" Construct Grad for %s \"; \n"
       "    egr::EagerUtils::PassStopGradient(%s);\n"
       "%s\n  }";
   std::string grad_node_creation_body_str = paddle::string::Sprintf(
       GRAD_NODE_CREATION_TEMPLATE, prepare_autograd_meta_str,
-      compute_require_grad_args, pass_stop_gradient_args,
+      compute_require_grad_args, op_type, pass_stop_gradient_args,
       grad_node_creation_str);
 
   return grad_node_creation_body_str;
@@ -2083,22 +2082,24 @@ static std::string GenerateGradNodeHeaderContents(
   const char* GRAD_NODE_TEMPLATE =
       "class GradNode%s : public egr::GradNodeBase {\n"
       " public:\n"
-      "  GradNode%s() : egr::GradNodeBase() {}\n"
+      "  GradNode%s() : egr::GradNodeBase() { VLOG(7) << \" Construct "
+      "GradNode%s \"; }\n"
       "  GradNode%s(size_t bwd_in_slot_num, size_t bwd_out_slot_num) : "
-      "egr::GradNodeBase(bwd_in_slot_num, bwd_out_slot_num) {}\n"
-      "  ~GradNode%s() override = default;\n"
+      "egr::GradNodeBase(bwd_in_slot_num, bwd_out_slot_num) { VLOG(7) << \" "
+      "Construct GradNode%s \"; }\n"
+      "  ~GradNode%s() override { VLOG(6) << \" Destruct GradNode%s \"; }\n"
       "\n"
       "  virtual std::vector<std::vector<paddle::experimental::Tensor>> "
       "operator()(const "
       "std::vector<std::vector<paddle::experimental::Tensor>>& grads) "
       "override;\n"
       "\n"
+      "  std::string name() override { return \" GradNode%s \"; } \n "
+      "\n"
       "  // SetX, SetY, ...\n"
       "%s\n"
       "  // SetAttrMap\n"
       "%s\n"
-      "  std::string name() { return \"GradNode%s\"; }\n"
-      "\n"
       " private:\n"
       "   // TensorWrappers\n"
       "%s\n"
@@ -2195,8 +2196,8 @@ static std::string GenerateGradNodeHeaderContents(
   VLOG(6) << "Generated TensorWrapper";
 
   std::string grad_node_str = paddle::string::Sprintf(
-      GRAD_NODE_TEMPLATE, op_type, op_type, op_type, op_type,
-      set_tensor_wrappers_str, set_attr_map_str, op_type,
+      GRAD_NODE_TEMPLATE, op_type, op_type, op_type, op_type, op_type, op_type,
+      op_type, op_type, set_tensor_wrappers_str, set_attr_map_str,
       tensor_wrapper_members_str, attr_members_str);
 
   return grad_node_str;

paddle/fluid/eager/auto_code_generator/final_state_generator/eager_gen.py

@@ -538,7 +538,7 @@ class {} : public egr::GradNodeBase {{
 
   virtual std::vector<std::vector<paddle::experimental::Tensor>> operator()(
       const std::vector<std::vector<paddle::experimental::Tensor>>& grads) override;
-  
+  std::string name() override {{ return \" {} \"; }}
   // SetTensorWrapperX, SetTensorWrapperY, ...
   {}
   // SetAttributes
@@ -553,8 +553,9 @@ class {} : public egr::GradNodeBase {{
 """
     node_declaration_str = NODE_DECLARATION_TEMPLATE.format(
         grad_node_name, grad_node_name, grad_node_name, grad_node_name,
-        set_tensor_wrapper_methods_str, set_attribute_methods_str,
-        tensor_wrapper_members_str, attribute_members_str)
+        grad_node_name, set_tensor_wrapper_methods_str,
+        set_attribute_methods_str, tensor_wrapper_members_str,
+        attribute_members_str)
 
     return node_declaration_str
 

paddle/fluid/eager/backward.cc

@@ -48,12 +48,16 @@ std::unordered_map<GradNodeBase*, int> getInDegreeMap(
     }
     visited.insert(node);
 
+    PADDLE_ENFORCE_NOT_NULL(
+        node,
+        paddle::platform::errors::Fatal(
+            "We got null node when we traverse the backward graph, and this "
+            "should not happened please check your code and contact us."));
     // Find and append next nodes
     const std::vector<std::vector<Edge>>& edges = node->GetEdges();
     for (const auto& edge_list : edges) {
       for (const Edge& edge : edge_list) {
         GradNodeBase* next_node = edge.GetMutableGradNode().get();
-
         // Next node could be nullptr if it is leaf tensor with no
         // AccumulationNode attached
         // Or it could also originated from dispensable inputs
@@ -67,7 +71,6 @@ std::unordered_map<GradNodeBase*, int> getInDegreeMap(
       }
     }
   }
-
   return node_in_degree_map;
 }
 

paddle/fluid/eager/grad_node_info.cc

@@ -30,6 +30,7 @@
 namespace egr {
 
 GradNodeBase::GradNodeBase(size_t bwd_in_slot_num, size_t bwd_out_slot_num) {
+  VLOG(6) << "Construct GradNodeBase";
   bwd_in_meta_.resize(bwd_in_slot_num);
   bwd_out_meta_.resize(bwd_out_slot_num);
   // adj_edges has the same num as backward outputs
@@ -49,11 +50,15 @@ void GradNodeBase::AddEdges(std::vector<AutogradMeta*>* metas, size_t slot_id) {
     // its pre-ops
     if (meta && !meta->StopGradient()) {
       auto node = meta->GetMutableGradNode();
-      if (node) {
+      if (node && node.get()) {
+        VLOG(6) << "Add Edges for slot: " << slot_id
+                << " which is: " << meta->GetMutableGradNode()->name();
         adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
                                          meta->OutRankInfo());
       } else {
         meta->SetGradNode(std::make_shared<egr::GradNodeAccumulation>(meta));
+        VLOG(6) << "Add Edges for slot: " << slot_id
+                << " which is: " << meta->GetMutableGradNode()->name();
         adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),
                                          meta->OutRankInfo());
       }
@@ -70,7 +75,7 @@ void GradNodeBase::AddEdges(AutogradMeta* meta, size_t slot_id) {
           "inputs's slot num."));
   if (meta && !meta->StopGradient()) {
     auto node = meta->GetMutableGradNode();
-    if (node) {
+    if (node && node.get()) {
       VLOG(6) << "Add Edges for slot: " << slot_id << ", the Edge is from "
               << this->name() << " to " << meta->GetMutableGradNode()->name();
       adj_edges_[slot_id].emplace_back(meta->GetMutableGradNode(),

paddle/fluid/eager/grad_node_info.h

@@ -76,10 +76,10 @@ class GradSlotMeta {
 
 class GradNodeBase {
  public:
-  GradNodeBase() = default;
+  GradNodeBase() { VLOG(6) << "Construct GradNodeBase"; }
   GradNodeBase(size_t bwd_in_slot_num, size_t bwd_out_slot_num);
   // TODO(jiabin): Should we have other constructor here?
-  virtual ~GradNodeBase() = default;
+  virtual ~GradNodeBase() { VLOG(6) << "Destruct GradNodeBase"; }
 
   /**
    * operator() designed to contian the real backward execution logic, it should

paddle/fluid/eager/tests/data_structure_tests/grad_node_test.h

@@ -30,6 +30,7 @@ class GradTestNode : public egr::GradNodeBase {
   GradTestNode(float val, int in_num, int out_num)
       : GradNodeBase(in_num, out_num), val_(val) {}
   GradTestNode() : GradNodeBase() { val_ = 1.0; }
+  std::string name() override { return "GradTestNode"; }
   std::vector<std::vector<paddle::experimental::Tensor>> operator()(
       const std::vector<std::vector<paddle::experimental::Tensor>>& grads)
       override {

paddle/fluid/eager/utils.cc

@@ -122,9 +122,10 @@ paddle::experimental::Tensor* EagerUtils::mutable_grad(
 void EagerUtils::SetHistory(std::vector<AutogradMeta*>* autograd_metas,
                             const std::shared_ptr<GradNodeBase>& grad_node) {
   for (const auto& autograd_meta : *autograd_metas) {
-    if (dynamic_cast<GradNodeAccumulation*>(autograd_meta->GradNode())) {
-      VLOG(6) << "Warning: Reseting GradNodeAccumulation for leaf tensor is "
-                 "detected";
+    if (autograd_meta->GradNode()) {
+      VLOG(7) << "Should not set grad node twice, original node is:"
+              << autograd_meta->GradNode()->name()
+              << "current is: " << grad_node->name();
     }
     autograd_meta->SetGradNode(grad_node);
   }
@@ -132,11 +133,11 @@ void EagerUtils::SetHistory(std::vector<AutogradMeta*>* autograd_metas,
 
 void EagerUtils::SetHistory(AutogradMeta* autograd_meta,
                             const std::shared_ptr<GradNodeBase>& grad_node) {
-  if (dynamic_cast<GradNodeAccumulation*>(autograd_meta->GradNode())) {
-    VLOG(6)
-        << "Warning: Reseting GradNodeAccumulation for leaf tensor is detected";
+  if (autograd_meta->GradNode()) {
+    VLOG(7) << "Should not set grad node twice, original node is:"
+            << autograd_meta->GradNode()->name()
+            << "current is: " << grad_node->name();
   }
-
   autograd_meta->SetGradNode(grad_node);
 }
 

paddle/fluid/pybind/eager_method.cc

@@ -19,6 +19,7 @@ limitations under the License. */
 
 #include "paddle/fluid/eager/accumulation/accumulation_node.h"
 #include "paddle/fluid/eager/api/all.h"
+#include "paddle/fluid/eager/api/generated/fluid_generated/dygraph_forward_api.h"
 #include "paddle/fluid/eager/autograd_meta.h"
 #include "paddle/fluid/eager/grad_node_info.h"
 #include "paddle/fluid/eager/hooks.h"
@@ -30,10 +31,12 @@ 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/fluid/pybind/slice_utils.h"
 #include "paddle/phi/api/include/api.h"
 #include "paddle/phi/common/data_type.h"
 #include "paddle/phi/core/compat/convert_utils.h"
 #include "paddle/phi/core/dense_tensor.h"
+
 namespace paddle {
 namespace pybind {
 
@@ -119,6 +122,29 @@ extern void InitTensorWithNumpyValue(TensorObject* self,
 
 extern PyTypeObject* p_tensor_type;
 
+Py_ssize_t GetSliceIndexFromPyObject(PyObject* obj) {
+  if (PyObject_IsInstance(obj, reinterpret_cast<PyObject*>(p_tensor_type))) {
+    VLOG(6) << "Call GetSliceIndexFromTensor in Eager";
+    paddle::experimental::Tensor tensor = CastPyArg2Tensor(obj, 0);
+    PADDLE_ENFORCE_EQ(
+        tensor.initialized(), true,
+        paddle::platform::errors::InvalidArgument(
+            "We can only support initialized tensor in slice, however we got "
+            "uninitialized tensor %s, please check your code.",
+            tensor.name()));
+    return GetSliceIndexFromTensor((*static_cast<phi::DenseTensor*>(
+        CastPyArg2Tensor(obj, 0).impl().get())));
+  } else {
+    PADDLE_THROW(platform::errors::InvalidArgument(
+        "We should only get paddle::experimental::Tensor or VarBase in this "
+        "method, when you reach this means we got another type index."));
+  }
+}
+
+bool PyCheckTensor(PyObject* obj) {
+  return PyObject_IsInstance(obj, reinterpret_cast<PyObject*>(p_tensor_type));
+}
+
 static PyObject* tensor_method_numpy(TensorObject* self, PyObject* args,
                                      PyObject* kwargs) {
   EAGER_TRY
@@ -468,16 +494,111 @@ static PyObject* tensor_method_get_underline_tensor(TensorObject* self,
   EAGER_CATCH_AND_THROW_RETURN_NULL
 }
 
-// NOTE(wuweilong): Set value and not change self's original place
-static PyObject* tensor_method_set_value(TensorObject* self, PyObject* args,
-                                         PyObject* kwargs) {
+static PyObject* tensor__getitem_index_not_tensor(TensorObject* self,
+                                                  PyObject* args,
+                                                  PyObject* kwargs) {
   EAGER_TRY
-  VLOG(4) << "Value " << self->tensor.name();
-  pybind11::object numpy_value =
-      pybind11::object(pybind11::handle(PyTuple_GET_ITEM(args, 0)), true);
-  InitTensorWithNumpyValue(self, numpy_value, false);
-  Py_INCREF(Py_None);
-  return Py_None;
+  PyObject* _index = PyTuple_GET_ITEM(args, 0);
+  VLOG(4) << "Call _getitem_index_not_tensor";
+  std::vector<int> slice_axes, slice_starts, slice_ends, slice_strides,
+      decrease_axis, none_axes, infer_flags, list_select_idxs;
+  // if index is a list, list_select_flag will be true
+  bool list_select_flag = false;
+  PADDLE_ENFORCE_EQ(
+      self->tensor.is_initialized(), true,
+      platform::errors::InvalidArgument(
+          "tensor %s has not been initialized, we can only slice initialized "
+          "tensor please init it first with numpy or other tensor.",
+          self->tensor.name()));
+  auto tensor = static_cast<phi::DenseTensor*>(self->tensor.impl().get());
+  ParseIndexingSlice(tensor, _index, &slice_axes, &slice_starts, &slice_ends,
+                     &slice_strides, &decrease_axis, &none_axes, &infer_flags,
+                     &list_select_idxs, &list_select_flag);
+
+  auto out = slice_axes.empty() && !list_select_flag
+                 ? self->tensor
+                 : paddle::experimental::Tensor(
+                       egr::Controller::Instance().GenerateUniqueName());
+
+  if (!slice_axes.empty()) {
+    framework::AttributeMap attrs = {{"axes", slice_axes},
+                                     {"starts", slice_starts},
+                                     {"ends", slice_ends},
+                                     {"infer_flags", infer_flags},
+                                     {"decrease_axis", decrease_axis}};
+    std::string op_type = "slice";
+    for (auto stride : slice_strides) {
+      if (stride != 1) {
+        op_type = "strided_slice";
+        attrs.insert({"strides", slice_strides});
+        attrs.erase("decrease_axis");
+        break;
+      }
+    }
+    if (op_type == "slice") {
+      out = slice_dygraph_function(self->tensor, paddle::experimental::Tensor(),
+                                   paddle::experimental::Tensor(),
+                                   std::move(attrs));
+    } else if (op_type == "strided_slice") {
+      out = strided_slice_dygraph_function(self->tensor, attrs);
+    } else {
+      PADDLE_THROW(platform::errors::InvalidArgument(
+          "Slice is only support slice and strided_slice, but we got %s which "
+          "is impossible, please check your code first or contact us by "
+          "issue. ",
+          op_type));
+    }
+  }
+
+  if (!none_axes.empty()) {
+    // Deal with cases when all axes are decreased.
+    // After slice, the shape of out is [1], which should have been
+    // [], but Paddle doesn't support scalar.
+    // In order to ensure the correctness of the final shape of out,
+    // one dimension of out needs to be decreased.
+    // For example:
+    // # x.shape: (2,3,4)
+    // out = x[0, 1, 1, None] # out.shape : (1)
+    if (static_cast<int>(decrease_axis.size()) == tensor->dims().size()) {
+      none_axes.pop_back();
+    }
+    if (!none_axes.empty()) {
+      // Deal with cases that decrease_axes is not empty
+      // For example:
+      // # x.shape: (2,3,4)
+      // out = x[0, 0:2, None] # out.shape : (2, 1, 4)
+      for (auto& axis : none_axes) {
+        int len = 0;
+        for (int da : decrease_axis) {
+          if (da < axis) {
+            len++;
+          }
+        }
+        axis -= len;
+      }
+
+      paddle::experimental::Tensor new_out;
+      framework::AttributeMap attrs = {{"axes", none_axes}};
+      new_out = std::get<0>(unsqueeze2_dygraph_function(out, std::move(attrs)));
+      return ToPyObject(new_out);
+    }
+  }
+
+  // the index is a list
+  if (list_select_flag) {
+    auto select_index = paddle::experimental::Tensor(
+        egr::Controller::Instance().GenerateUniqueName());
+    auto idx_tensor = std::make_shared<phi::DenseTensor>();
+    auto* dev_ctx = platform::DeviceContextPool::Instance().Get(
+        egr::Controller::Instance().GetExpectedPlace());
+    paddle::framework::TensorFromVector(list_select_idxs, *dev_ctx,
+                                        idx_tensor.get());
+    framework::AttributeMap attrs = {{"dim", 0}};
+    out = index_select_dygraph_function(self->tensor, select_index,
+                                        std::move(attrs));
+  }
+
+  return ToPyObject(out);
   EAGER_CATCH_AND_THROW_RETURN_NULL
 }
 
@@ -602,7 +723,8 @@ PyMethodDef variable_methods[] = {
     {"get_tensor",
      (PyCFunction)(void (*)(void))tensor_method_get_underline_tensor,
      METH_VARARGS | METH_KEYWORDS, NULL},
-    {"_set_value", (PyCFunction)(void (*)(void))tensor_method_set_value,
+    {"_getitem_index_not_tensor",
+     (PyCFunction)(void (*)(void))tensor__getitem_index_not_tensor,
      METH_VARARGS | METH_KEYWORDS, NULL},
     {"_register_grad_hook",
      (PyCFunction)(void (*)(void))tensor_register_grad_hook,

paddle/fluid/pybind/eager_utils.cc

@@ -16,8 +16,11 @@ limitations under the License. */
 
 #include "paddle/fluid/eager/api/all.h"
 #include "paddle/fluid/eager/autograd_meta.h"
+#include "paddle/fluid/framework/convert_utils.h"
+#include "paddle/fluid/framework/scope_guard.h"
 #include "paddle/fluid/memory/allocation/allocator.h"
 #include "paddle/fluid/operators/py_func_op.h"
+#include "paddle/fluid/operators/utils.h"
 #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/pybind/eager.h"
 #include "paddle/fluid/pybind/eager_utils.h"
@@ -184,6 +187,11 @@ paddle::experimental::Tensor CastPyArg2Tensor(PyObject* obj, ssize_t arg_pos) {
   }
 }
 
+std::shared_ptr<imperative::VarBase> CastPyArg2VarBase(PyObject* obj,
+                                                       ssize_t arg_pos) {
+  return py::cast<std::shared_ptr<imperative::VarBase>>(obj);
+}
+
 std::vector<paddle::experimental::Tensor> CastPyArg2VectorOfTensor(
     PyObject* obj, ssize_t arg_pos) {
   std::vector<paddle::experimental::Tensor> result;
@@ -737,5 +745,6 @@ std::vector<paddle::experimental::Tensor*> GetTensorPtrListFromArgs(
 
   return result;
 }
+
 }  // namespace pybind
 }  // namespace paddle

paddle/fluid/pybind/eager_utils.h

@@ -14,7 +14,6 @@ limitations under the License. */
 #include "paddle/phi/core/dense_tensor.h"
 #include "pybind11/pybind11.h"
 #include "pybind11/stl.h"
-
 namespace paddle {
 namespace pybind {
 
@@ -33,6 +32,8 @@ int64_t CastPyArg2AttrLong(PyObject* obj, ssize_t arg_pos);
 float CastPyArg2AttrFloat(PyObject* obj, ssize_t arg_pos);
 std::string CastPyArg2AttrString(PyObject* obj, ssize_t arg_pos);
 paddle::experimental::Tensor CastPyArg2Tensor(PyObject* obj, ssize_t arg_pos);
+std::shared_ptr<imperative::VarBase> CastPyArg2VarBase(PyObject* obj,
+                                                       ssize_t arg_pos);
 std::vector<paddle::experimental::Tensor> CastPyArg2VectorOfTensor(
     PyObject* obj, ssize_t arg_pos);
 platform::Place CastPyArg2Place(PyObject* obj, ssize_t arg_pos);
@@ -112,5 +113,7 @@ std::vector<paddle::experimental::Tensor*> GetTensorPtrListFromArgs(
     const std::string& op_type, const std::string& arg_name, PyObject* args,
     ssize_t arg_idx, bool dispensable = false);
 
+// end of Slice related methods
+
 }  // namespace pybind
 }  // namespace paddle

paddle/fluid/pybind/imperative.cc

@@ -54,6 +54,7 @@ limitations under the License. */
 #include "paddle/fluid/operators/utils.h"
 #include "paddle/fluid/pybind/op_function.h"
 #include "paddle/fluid/pybind/pybind_boost_headers.h"
+#include "paddle/fluid/pybind/slice_utils.h"
 #include "paddle/fluid/pybind/tensor_py.h"
 
 namespace paddle {
@@ -319,6 +320,23 @@ static std::string GetTypeName(const imperative::VarBase &var) {
   }
 }
 
+Py_ssize_t GetSliceIndexFromPyObject(PyObject *obj) {
+  if (py::isinstance<imperative::VarBase>(obj)) {
+    VLOG(6) << "Call GetSliceIndexFromTensor in Imperative";
+    return GetSliceIndexFromTensor(
+        py::cast<std::shared_ptr<imperative::VarBase>>(obj)
+            ->Var()
+            .Get<framework::LoDTensor>());
+  } else {
+    PADDLE_THROW(platform::errors::InvalidArgument(
+        "We should only get paddle::experimental::Tensor or VarBase in this "
+        "method, when you reach this means we got another type index."));
+  }
+}
+
+bool PyCheckTensor(PyObject *obj) {
+  return py::isinstance<imperative::VarBase>(obj);
+}
 using PyNameVarBaseMap = std::unordered_map<std::string, py::handle>;
 
 // NOTE(zjl): py::handle is a very light wrapper of PyObject *.
@@ -360,18 +378,6 @@ GetVarBaseListFromPyHandle(const py::handle &handle) {
 
   return result;
 }
-static bool IsNumpyType(PyObject *obj) {
-  // It is not a good way to judge the type of obj by its type'name. Maybe using
-  // `PyArray_IsScalar` will be better. However, this interface cannot be used
-  // by including pybind11, and it needs to compile with numpy.
-  auto type_name = std::string(Py_TYPE(obj)->tp_name);
-  return type_name == "numpy.int64" || type_name == "numpy.longlong" ||
-         type_name == "numpy.int32" || type_name == "numpy.int16";
-}
-
-static bool PyCheckTensor(PyObject *obj) {
-  return py::isinstance<imperative::VarBase>(obj);
-}
 
 // cast numpy type form S to T, this may allocate new memory
 template <class T, class S>
@@ -429,260 +435,6 @@ static imperative::NameVarBaseMap ConvertToNameVarBaseMap(
   return result;
 }
 
-static bool PyCheckInteger(PyObject *obj) {
-#if PY_VERSION_HEX < 0x03000000
-  return (PyLong_Check(obj) || PyInt_Check(obj)) && !PyBool_Check(obj);
-#else
-  return PyLong_Check(obj) && !PyBool_Check(obj);
-#endif
-}
-
-static Py_ssize_t GetSliceIndexFromTensor(
-    const std::shared_ptr<imperative::VarBase> &tensor_index) {
-  const auto &tensor = tensor_index->Var().Get<framework::LoDTensor>();
-  if (tensor.numel() == 1) {
-    if (framework::TransToProtoVarType(tensor.dtype()) ==
-        framework::proto::VarType::INT32) {
-      return static_cast<Py_ssize_t>(operators::GetValue<int32_t>(&tensor));
-    } else if (framework::TransToProtoVarType(tensor.dtype()) ==
-               framework::proto::VarType::INT64) {
-      return static_cast<Py_ssize_t>(operators::GetValue<int64_t>(&tensor));
-    } else {
-      PADDLE_THROW(platform::errors::InvalidArgument(
-          "Currently, the type of tensor in slice indices only allows "
-          "int32 and int64, please check the type of index tensor."));
-    }
-  } else {
-    PADDLE_THROW(platform::errors::InvalidArgument(
-        "Currently, tensor in slice indices only allows 1 element, "
-        "but received %d.",
-        tensor.numel()));
-  }
-}
-
-// NOTE(zhiqiu): Revised version of PySlice_GetIndices. From:
-// https://github.com/python/cpython/blob/8d21aa21f2cbc6d50aab3f420bb23be1d081dac4/Objects/sliceobject.c#L103
-// Original PySlice_GetIndices return wrong result when
-// slice_item contains long int, such as arr[:180L].
-// NOT sure why this happens !!!
-// Besides, PySlice_GetIndices cannot raise error when float in slice item.
-// So, I make a revised version of PySlice_GetIndices, named to
-// _PySlice_GetIndices. Try to use _PySlice_Unpack which is more robust than
-// PySlice_GetIndices in the future.
-static int _PySlice_GetIndices(PySliceObject *r, Py_ssize_t length,
-                               Py_ssize_t *start, Py_ssize_t *stop,
-                               Py_ssize_t *step) {
-  /* XXX support long ints */
-  if (r->step == Py_None) {
-    *step = 1;
-  } else {
-    if (PyCheckInteger(r->step) || IsNumpyType(r->step)) {
-      *step = PyLong_AsLong(r->step);
-    } else if (PyCheckTensor(r->step)) {
-      *step = GetSliceIndexFromTensor(
-          py::cast<std::shared_ptr<imperative::VarBase>>(r->step));
-    } else {
-      PADDLE_THROW(platform::errors::InvalidArgument(
-          "Currently, slice indices only allows None, integers, "
-          "tensor(int) and numpy(int) in slice item, but received %s.",
-          std::string(Py_TYPE(r->step)->tp_name)));
-    }
-  }
-  if (r->start == Py_None) {
-    *start = *step < 0 ? length - 1 : 0;
-  } else {
-    if (PyCheckInteger(r->start) || IsNumpyType(r->start)) {
-      *start = PyLong_AsLong(r->start);
-    } else if (PyCheckTensor(r->start)) {
-      *start = GetSliceIndexFromTensor(
-          py::cast<std::shared_ptr<imperative::VarBase>>(r->start));
-    } else {
-      PADDLE_THROW(platform::errors::InvalidArgument(
-          "Currently, slice indices only allows None, integers, "
-          "tensor(int) and numpy(int) in slice item, but received %s.",
-          std::string(Py_TYPE(r->start)->tp_name)));
-    }
-    if (*start < 0) *start += length;
-    *start = std::max(*start, static_cast<Py_ssize_t>(0));
-  }
-  if (r->stop == Py_None) {
-    *stop = *step < 0 ? -1 : length;
-  } else {
-    if (PyCheckInteger(r->stop) || IsNumpyType(r->stop)) {
-      *stop = PyLong_AsLong(r->stop);
-    } else if (PyCheckTensor(r->stop)) {
-      *stop = GetSliceIndexFromTensor(
-          py::cast<std::shared_ptr<imperative::VarBase>>(r->stop));
-    } else {
-      PADDLE_THROW(platform::errors::InvalidArgument(
-          "Currently, slice indices only allows None, integers, "
-          "tensor(int) and numpy(int) in slice item, but received %s.",
-          std::string(Py_TYPE(r->stop)->tp_name)));
-    }
-    if (0 < *step && *stop < 0) *stop += length;
-    *stop = std::min(*stop, length);
-  }
-  if (*stop > length) return -1;
-  if (*start >= length) return -1;
-  if (*step == 0) return -1;
-  return 0;
-}
-
-static void ParseIndexingSlice(
-    framework::LoDTensor *tensor, PyObject *_index,
-    std::vector<int> *slice_axes, std::vector<int> *slice_starts,
-    std::vector<int> *slice_ends, std::vector<int> *slice_strides,
-    std::vector<int> *decrease_axis, std::vector<int> *none_axes,
-    std::vector<int> *infer_flags, std::vector<int> *list_select_idxs,
-    bool *list_select_flag) {
-  // We allow indexing by Integers, Slices, Ellipsis, None, tuples of those
-  // types, and list of Bool and Integers.
-  // wrap to tuple
-
-  // NOTE(zhiqiu): PyTuple_Pack increases refcount.
-  PyObject *index = !PyTuple_Check(_index) ? PyTuple_Pack(1, _index) : _index;
-  DEFINE_PADDLE_SCOPE_GUARD([index, _index]() {
-    if (!PyTuple_Check(_index)) {
-      Py_DECREF(index);
-      VLOG(4) << "Call Py_DECREF";
-    }
-  });
-  PADDLE_ENFORCE_EQ(
-      tensor->IsInitialized(), true,
-      platform::errors::InvalidArgument("tensor has not been initialized"));
-  const auto &shape = tensor->dims();
-  const int rank = shape.size();
-  const int size = PyTuple_GET_SIZE(index);
-
-  // specified_dims is the number of dimensions which indexed by Interger,
-  // Slices.
-  int specified_dims = 0;
-  int ell_count = 0;
-  for (int dim = 0; dim < size; ++dim) {
-    PyObject *slice_item = PyTuple_GetItem(index, dim);
-    if (PyCheckInteger(slice_item) || PySlice_Check(slice_item)) {
-      specified_dims++;
-    } else if (slice_item == Py_Ellipsis) {
-      ell_count++;
-    }
-  }
-
-  PADDLE_ENFORCE_LE(ell_count, 1,
-                    platform::errors::InvalidArgument(
-                        "An index can only have a single ellipsis ('...')"));
-  int none_count = 0;
-  for (int i = 0, dim = 0; i < size; ++i) {
-    PyObject *slice_item = PyTuple_GetItem(index, i);
-
-    infer_flags->push_back(1);
-    int dim_len = shape[dim];
-    if (PyCheckInteger(slice_item) || IsNumpyType(slice_item)) {
-      // integer, PyLong_AsLong supports both int and long
-      int start = static_cast<int>(PyLong_AsLong(slice_item));
-      auto s_t = start;
-      start = start < 0 ? start + dim_len : start;
-      if (start >= dim_len || start < 0) {
-        std::string str_error_message =
-            "The starting index " + std::to_string(s_t) +
-            " of slice is out of bounds in tensor " + std::to_string(dim) +
-            "-th axis, it shound be in the range of [" +
-            std::to_string(-dim_len) + ", " + std::to_string(dim_len) + ")";
-        // py::index_error is corresponding to IndexError in Python
-        // Used to indicate out of bounds access in __getitem__, __setitem__
-        throw py::index_error(str_error_message);
-      }
-      slice_axes->push_back(dim);
-      slice_starts->push_back(start);
-      slice_ends->push_back(start + 1);
-      slice_strides->push_back(1);
-      decrease_axis->push_back(dim);
-      dim++;
-    } else if (PySlice_Check(slice_item)) {
-      // slice item
-      Py_ssize_t start, end, step;
-      PySliceObject *p = reinterpret_cast<PySliceObject *>(slice_item);
-      _PySlice_GetIndices(p, dim_len, &start, &end, &step);
-
-      // :: or : or 0:dim_len:1
-      if (start == 0 && end == dim_len && step == 1) {
-        dim++;
-        continue;
-      }
-      slice_axes->push_back(dim);
-      slice_starts->push_back(start);
-      slice_ends->push_back(end);
-      slice_strides->push_back(step);
-      dim++;
-    } else if (slice_item == Py_Ellipsis) {
-      dim += rank - specified_dims;
-    } else if (slice_item == Py_None) {
-      none_axes->push_back(dim + none_count);
-      none_count++;
-    } else if (PyList_Check(slice_item)) {
-      *list_select_flag = true;
-      PADDLE_ENFORCE_EQ(
-          size, 1,
-          platform::errors::InvalidArgument(
-              "When index contains a list, its length is excepted to 1, "
-              "but received %d",
-              size));
-      bool all_bool = true;
-      int list_size = PyList_GET_SIZE(slice_item);
-      for (int j = 0; j < list_size; ++j) {
-        PyObject *list_item = PyList_GetItem(slice_item, j);
-        if (PyCheckInteger(list_item)) {
-          all_bool = false;
-        } else if (!PyBool_Check(list_item)) {
-          PADDLE_THROW(platform::errors::InvalidArgument(
-              "Only support int or bool in index list."));
-        }
-      }
-      if (all_bool) {
-        PADDLE_ENFORCE_EQ(
-            list_size, shape[0],
-            platform::errors::InvalidArgument(
-                "The dimension of bool index doesn't match indexed array along "
-                "dimension 0, the target dimension is %d, but received %d.",
-                shape[0], list_size));
-
-        for (int j = 0; j < list_size; ++j) {
-          PyObject *list_item = PyList_GetItem(slice_item, j);
-          if (list_item == Py_True) {
-            list_select_idxs->push_back(j);
-          }
-        }
-      } else {
-        for (int j = 0; j < list_size; ++j) {
-          PyObject *list_item = PyList_GetItem(slice_item, j);
-          if (PyCheckInteger(list_item)) {
-            list_select_idxs->push_back(
-                static_cast<int>(PyLong_AsLong(list_item)));
-          } else if (list_item == Py_True) {
-            list_select_idxs->push_back(1);
-          } else {
-            list_select_idxs->push_back(0);
-          }
-        }
-      }
-
-    } else {
-      PADDLE_THROW(platform::errors::InvalidArgument(
-          "Currently, Tensor.__indices__() only allows indexing "
-          "by Integers, Slices, Ellipsis, None, tuples of these types "
-          "and list of Bool and Integers, but received "
-          "%s in %dth slice item",
-          std::string(Py_TYPE(slice_item)->tp_name), i + 1));
-    }
-  }
-
-  // valid_index is the number of dimensions exclude None index
-  const int valid_indexs = size - none_axes->size() - ell_count;
-  PADDLE_ENFORCE_EQ(valid_indexs <= rank, true,
-                    platform::errors::InvalidArgument(
-                        "Too many indices (%d) for tensor of dimension %d.",
-                        valid_indexs, rank));
-}
-
 template <typename P>
 static void VarBaseCopy(std::shared_ptr<imperative::VarBase> &src,  // NOLINT
                         imperative::VarBase &dst,                   // NOLINT

paddle/fluid/pybind/pybind.cc

@@ -80,6 +80,7 @@ limitations under the License. */
 #include "paddle/fluid/pybind/cuda_streams_py.h"
 #include "paddle/fluid/pybind/distributed_py.h"
 #include "paddle/fluid/pybind/eager.h"
+#include "paddle/fluid/pybind/imperative.h"
 #include "paddle/fluid/pybind/io.h"
 #include "paddle/phi/core/compat/convert_utils.h"
 #include "paddle/phi/core/lod_utils.h"
@@ -101,7 +102,6 @@ limitations under the License. */
 #include "paddle/fluid/pybind/gloo_context_py.h"
 #include "paddle/fluid/pybind/gloo_wrapper_py.h"
 #include "paddle/fluid/pybind/heter_wrapper_py.h"
-#include "paddle/fluid/pybind/imperative.h"
 #include "paddle/fluid/pybind/inference_api.h"
 #include "paddle/fluid/pybind/ir.h"
 #include "paddle/fluid/pybind/metrics_py.h"
@@ -527,6 +527,7 @@ PYBIND11_MODULE(core_avx, m) {
 PYBIND11_MODULE(core_noavx, m) {
 #endif
 
+  BindImperative(&m);
   BindEager(&m);
   BindCudaStream(&m);
 
@@ -741,8 +742,6 @@ PYBIND11_MODULE(core_noavx, m) {
   m.def("_promote_types_if_complex_exists",
         &paddle::framework::PromoteTypesIfComplexExists);
 
-  BindImperative(&m);
-
   py::class_<framework::Tensor> framework_tensor(m, "Tensor",
                                                  py::buffer_protocol());
   g_framework_tensor_pytype =

paddle/fluid/pybind/slice_utils.h

+// Copyright (c) 2022 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.
+
+#pragma once
+
+#include <Python.h>
+#include "paddle/fluid/framework/convert_utils.h"
+#include "paddle/fluid/framework/scope_guard.h"
+#include "paddle/fluid/operators/utils.h"
+#include "paddle/phi/common/data_type.h"
+#include "paddle/phi/core/compat/convert_utils.h"
+#include "paddle/phi/core/dense_tensor.h"
+#include "pybind11/pybind11.h"
+#include "pybind11/stl.h"
+
+namespace py = pybind11;
+
+namespace paddle {
+namespace pybind {
+
+static bool PyCheckTensor(PyObject* obj);
+static Py_ssize_t GetSliceIndexFromPyObject(PyObject* obj);
+// Slice related methods
+static bool PyCheckInteger(PyObject* obj) {
+#if PY_VERSION_HEX < 0x03000000
+  return (PyLong_Check(obj) || PyInt_Check(obj)) && !PyBool_Check(obj);
+#else
+  return PyLong_Check(obj) && !PyBool_Check(obj);
+#endif
+}
+
+static bool IsNumpyType(PyObject* obj) {
+  // It is not a good way to judge the type of obj by its type'name. Maybe using
+  // `PyArray_IsScalar` will be better. However, this interface cannot be used
+  // by including pybind11, and it needs to compile with numpy.
+  auto type_name = std::string(Py_TYPE(obj)->tp_name);
+  return type_name == "numpy.int64" || type_name == "numpy.longlong" ||
+         type_name == "numpy.int32" || type_name == "numpy.int16";
+}
+
+static Py_ssize_t GetSliceIndexFromTensor(const phi::DenseTensor& tensor) {
+  if (tensor.numel() == 1) {
+    if (framework::TransToProtoVarType(tensor.type()) ==
+        framework::proto::VarType::INT32) {
+      return static_cast<Py_ssize_t>(operators::GetValue<int32_t>(&tensor));
+    } else if (framework::TransToProtoVarType(tensor.type()) ==
+               framework::proto::VarType::INT64) {
+      return static_cast<Py_ssize_t>(operators::GetValue<int64_t>(&tensor));
+    } else {
+      PADDLE_THROW(platform::errors::InvalidArgument(
+          "Currently, the type of tensor in slice indices only allows "
+          "int32 and int64, please check the type of index tensor."));
+    }
+  } else {
+    PADDLE_THROW(platform::errors::InvalidArgument(
+        "Currently, tensor in slice indices only allows 1 element, "
+        "but received %d.",
+        tensor.numel()));
+  }
+}
+
+// NOTE(zhiqiu): Revised version of PySlice_GetIndices. From:
+// https://github.com/python/cpython/blob/8d21aa21f2cbc6d50aab3f420bb23be1d081dac4/Objects/sliceobject.c#L103
+// Original PySlice_GetIndices return wrong result when
+// slice_item contains long int, such as arr[:180L].
+// NOT sure why this happens !!!
+// Besides, PySlice_GetIndices cannot raise error when float in slice item.
+// So, I make a revised version of PySlice_GetIndices, named to
+// _PySlice_GetIndices. Try to use _PySlice_Unpack which is more robust than
+// PySlice_GetIndices in the future.
+static int _PySlice_GetIndices(PySliceObject* r, Py_ssize_t length,
+                               Py_ssize_t* start, Py_ssize_t* stop,
+                               Py_ssize_t* step) {
+  /* XXX support long ints */
+  if (r->step == Py_None) {
+    *step = 1;
+  } else {
+    if (PyCheckInteger(r->step) || IsNumpyType(r->step)) {
+      *step = PyLong_AsLong(r->step);
+    } else if (PyCheckTensor(r->step)) {
+      *step = GetSliceIndexFromPyObject(r->step);
+    } else {
+      PADDLE_THROW(platform::errors::InvalidArgument(
+          "Currently, slice indices only allows None, integers, "
+          "tensor(int) and numpy(int) in slice item, but received %s.",
+          std::string(Py_TYPE(r->step)->tp_name)));
+    }
+  }
+  if (r->start == Py_None) {
+    *start = *step < 0 ? length - 1 : 0;
+  } else {
+    if (PyCheckInteger(r->start) || IsNumpyType(r->start)) {
+      *start = PyLong_AsLong(r->start);
+    } else if (PyCheckTensor(r->start)) {
+      *start = GetSliceIndexFromPyObject(r->start);
+    } else {
+      PADDLE_THROW(platform::errors::InvalidArgument(
+          "Currently, slice indices only allows None, integers, "
+          "tensor(int) and numpy(int) in slice item, but received %s.",
+          std::string(Py_TYPE(r->start)->tp_name)));
+    }
+    if (*start < 0) *start += length;
+    *start = std::max(*start, static_cast<Py_ssize_t>(0));
+  }
+  if (r->stop == Py_None) {
+    *stop = *step < 0 ? -1 : length;
+  } else {
+    if (PyCheckInteger(r->stop) || IsNumpyType(r->stop)) {
+      *stop = PyLong_AsLong(r->stop);
+    } else if (PyCheckTensor(r->stop)) {
+      *stop = GetSliceIndexFromPyObject(r->stop);
+    } else {
+      PADDLE_THROW(platform::errors::InvalidArgument(
+          "Currently, slice indices only allows None, integers, "
+          "tensor(int) and numpy(int) in slice item, but received %s.",
+          std::string(Py_TYPE(r->stop)->tp_name)));
+    }
+    if (0 < *step && *stop < 0) *stop += length;
+    *stop = std::min(*stop, length);
+  }
+  if (*stop > length) return -1;
+  if (*start >= length) return -1;
+  if (*step == 0) return -1;
+  return 0;
+}
+
+static void ParseIndexingSlice(
+    framework::LoDTensor* tensor, PyObject* _index,
+    std::vector<int>* slice_axes, std::vector<int>* slice_starts,
+    std::vector<int>* slice_ends, std::vector<int>* slice_strides,
+    std::vector<int>* decrease_axis, std::vector<int>* none_axes,
+    std::vector<int>* infer_flags, std::vector<int>* list_select_idxs,
+    bool* list_select_flag) {
+  // We allow indexing by Integers, Slices, Ellipsis, None, tuples of those
+  // types, and list of Bool and Integers.
+  // wrap to tuple
+
+  // NOTE(zhiqiu): PyTuple_Pack increases refcount.
+  PyObject* index = !PyTuple_Check(_index) ? PyTuple_Pack(1, _index) : _index;
+  DEFINE_PADDLE_SCOPE_GUARD([index, _index]() {
+    if (!PyTuple_Check(_index)) {
+      Py_DECREF(index);
+      VLOG(4) << "Call Py_DECREF";
+    }
+  });
+  PADDLE_ENFORCE_EQ(
+      tensor->IsInitialized(), true,
+      platform::errors::InvalidArgument("tensor has not been initialized"));
+  const auto& shape = tensor->dims();
+  const int rank = shape.size();
+  const int size = PyTuple_GET_SIZE(index);
+
+  // specified_dims is the number of dimensions which indexed by Interger,
+  // Slices.
+  int specified_dims = 0;
+  int ell_count = 0;
+  for (int dim = 0; dim < size; ++dim) {
+    PyObject* slice_item = PyTuple_GetItem(index, dim);
+    if (PyCheckInteger(slice_item) || PySlice_Check(slice_item)) {
+      specified_dims++;
+    } else if (slice_item == Py_Ellipsis) {
+      ell_count++;
+    }
+  }
+
+  PADDLE_ENFORCE_LE(ell_count, 1,
+                    platform::errors::InvalidArgument(
+                        "An index can only have a single ellipsis ('...')"));
+  int none_count = 0;
+  for (int i = 0, dim = 0; i < size; ++i) {
+    PyObject* slice_item = PyTuple_GetItem(index, i);
+
+    infer_flags->push_back(1);
+    int dim_len = shape[dim];
+    if (PyCheckInteger(slice_item) || IsNumpyType(slice_item)) {
+      // integer, PyLong_AsLong supports both int and long
+      int start = static_cast<int>(PyLong_AsLong(slice_item));
+      auto s_t = start;
+      start = start < 0 ? start + dim_len : start;
+      if (start >= dim_len || start < 0) {
+        std::string str_error_message =
+            "The starting index " + std::to_string(s_t) +
+            " of slice is out of bounds in tensor " + std::to_string(dim) +
+            "-th axis, it shound be in the range of [" +
+            std::to_string(-dim_len) + ", " + std::to_string(dim_len) + ")";
+        // py::index_error is corresponding to IndexError in Python
+        // Used to indicate out of bounds access in __getitem__, __setitem__
+        throw py::index_error(str_error_message);
+      }
+      slice_axes->push_back(dim);
+      slice_starts->push_back(start);
+      slice_ends->push_back(start + 1);
+      slice_strides->push_back(1);
+      decrease_axis->push_back(dim);
+      dim++;
+    } else if (PySlice_Check(slice_item)) {
+      // slice item
+      Py_ssize_t start, end, step;
+      PySliceObject* p = reinterpret_cast<PySliceObject*>(slice_item);
+      _PySlice_GetIndices(p, dim_len, &start, &end, &step);
+
+      // :: or : or 0:dim_len:1
+      if (start == 0 && end == dim_len && step == 1) {
+        dim++;
+        continue;
+      }
+      slice_axes->push_back(dim);
+      slice_starts->push_back(start);
+      slice_ends->push_back(end);
+      slice_strides->push_back(step);
+      dim++;
+    } else if (slice_item == Py_Ellipsis) {
+      dim += rank - specified_dims;
+    } else if (slice_item == Py_None) {
+      none_axes->push_back(dim + none_count);
+      none_count++;
+    } else if (PyList_Check(slice_item)) {
+      *list_select_flag = true;
+      PADDLE_ENFORCE_EQ(
+          size, 1,
+          platform::errors::InvalidArgument(
+              "When index contains a list, its length is excepted to 1, "
+              "but received %d",
+              size));
+      bool all_bool = true;
+      int list_size = PyList_GET_SIZE(slice_item);
+      for (int j = 0; j < list_size; ++j) {
+        PyObject* list_item = PyList_GetItem(slice_item, j);
+        if (PyCheckInteger(list_item)) {
+          all_bool = false;
+        } else if (!PyBool_Check(list_item)) {
+          PADDLE_THROW(platform::errors::InvalidArgument(
+              "Only support int or bool in index list."));
+        }
+      }
+      if (all_bool) {
+        PADDLE_ENFORCE_EQ(
+            list_size, shape[0],
+            platform::errors::InvalidArgument(
+                "The dimension of bool index doesn't match indexed array along "
+                "dimension 0, the target dimension is %d, but received %d.",
+                shape[0], list_size));
+
+        for (int j = 0; j < list_size; ++j) {
+          PyObject* list_item = PyList_GetItem(slice_item, j);
+          if (list_item == Py_True) {
+            list_select_idxs->push_back(j);
+          }
+        }
+      } else {
+        for (int j = 0; j < list_size; ++j) {
+          PyObject* list_item = PyList_GetItem(slice_item, j);
+          if (PyCheckInteger(list_item)) {
+            list_select_idxs->push_back(
+                static_cast<int>(PyLong_AsLong(list_item)));
+          } else if (list_item == Py_True) {
+            list_select_idxs->push_back(1);
+          } else {
+            list_select_idxs->push_back(0);
+          }
+        }
+      }
+
+    } else {
+      PADDLE_THROW(platform::errors::InvalidArgument(
+          "Currently, Tensor.__indices__() only allows indexing "
+          "by Integers, Slices, Ellipsis, None, tuples of these types "
+          "and list of Bool and Integers, but received "
+          "%s in %dth slice item",
+          std::string(Py_TYPE(slice_item)->tp_name), i + 1));
+    }
+  }
+
+  // valid_index is the number of dimensions exclude None index
+  const int valid_indexs = size - none_axes->size() - ell_count;
+  PADDLE_ENFORCE_EQ(valid_indexs <= rank, true,
+                    platform::errors::InvalidArgument(
+                        "Too many indices (%d) for tensor of dimension %d.",
+                        valid_indexs, rank));
+}
+
+}  // namespace pybind
+}  // namespace paddle

python/paddle/fluid/contrib/slim/tests/CMakeLists.txt

@@ -351,10 +351,10 @@ endif()
 
 set_tests_properties(test_graph PROPERTIES TIMEOUT 120)
 set_tests_properties(test_quantization_pass PROPERTIES TIMEOUT 120)
-set_tests_properties(test_imperative_qat_channelwise PROPERTIES TIMEOUT 120)
-set_tests_properties(test_user_defined_quantization PROPERTIES TIMEOUT 120)
-set_tests_properties(test_imperative_qat PROPERTIES TIMEOUT 120)
-set_tests_properties(test_imperative_out_scale PROPERTIES TIMEOUT 120)
+set_tests_properties(test_imperative_qat_channelwise PROPERTIES TIMEOUT 200)
+set_tests_properties(test_user_defined_quantization PROPERTIES TIMEOUT 200)
+set_tests_properties(test_imperative_qat PROPERTIES TIMEOUT 200)
+set_tests_properties(test_imperative_out_scale PROPERTIES TIMEOUT 200)
 if(LINUX AND WITH_MKLDNN)
     set_tests_properties(test_quant2_int8_mobilenetv1_mkldnn PROPERTIES TIMEOUT 120)
     set_tests_properties(convert_model2dot_ernie PROPERTIES TIMEOUT 120)

python/paddle/fluid/contrib/slim/tests/test_imperative_out_scale.py

@@ -26,7 +26,7 @@ import paddle.fluid as fluid
 import paddle.fluid.layers as layers
 from paddle.fluid import core
 from paddle.fluid.optimizer import AdamOptimizer
-from paddle.fluid.framework import IrGraph
+from paddle.fluid.framework import IrGraph, _test_eager_guard
 from paddle.fluid.contrib.slim.quantization import ImperativeQuantAware
 from paddle.fluid.dygraph.container import Sequential
 from paddle.fluid.dygraph.io import INFER_MODEL_SUFFIX, INFER_PARAMS_SUFFIX
@@ -122,7 +122,7 @@ class ImperativeLenet(fluid.dygraph.Layer):
 
 
 class TestImperativeOutSclae(unittest.TestCase):
-    def test_out_scale_acc(self):
+    def func_out_scale_acc(self):
         seed = 1000
         lr = 0.001
 
@@ -166,9 +166,14 @@ class TestImperativeOutSclae(unittest.TestCase):
                 loss_list[i] > loss_list[i + 1],
                 msg='Failed to do the imperative qat.')
 
+    def test_out_scale_acc(self):
+        with _test_eager_guard():
+            self.func_out_scale_acc()
+        self.func_out_scale_acc()
+
 
 class TestSaveQuanztizedModelFromCheckPoint(unittest.TestCase):
-    def test_save_quantized_model(self):
+    def func_save_quantized_model(self):
         lr = 0.001
 
         load_param_path = "test_save_quantized_model/lenet.pdparams"
@@ -206,6 +211,11 @@ class TestSaveQuanztizedModelFromCheckPoint(unittest.TestCase):
                 loss_list[i] > loss_list[i + 1],
                 msg='Failed to do the imperative qat.')
 
+    def test_save_quantized_model(self):
+        with _test_eager_guard():
+            self.func_save_quantized_model()
+        self.func_save_quantized_model()
+
 
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/contrib/slim/tests/test_imperative_ptq.py

@@ -29,6 +29,7 @@ import paddle.fluid as fluid
 from paddle.fluid.contrib.slim.quantization import *
 from paddle.fluid.log_helper import get_logger
 from paddle.dataset.common import download
+from paddle.fluid.framework import _test_eager_guard
 
 from imperative_test_utils import fix_model_dict, ImperativeLenet, ImperativeLinearBn
 from imperative_test_utils import ImperativeLinearBn_hook
@@ -194,7 +195,7 @@ class TestImperativePTQ(unittest.TestCase):
                 break
         return top1_correct_num / total_num
 
-    def test_ptq(self):
+    def func_ptq(self):
         start_time = time.time()
 
         self.set_vars()
@@ -244,9 +245,14 @@ class TestImperativePTQ(unittest.TestCase):
         end_time = time.time()
         print("total time: %ss \n" % (end_time - start_time))
 
+    def test_ptq(self):
+        with _test_eager_guard():
+            self.func_ptq()
+        self.func_ptq()
+
 
 class TestImperativePTQfuse(TestImperativePTQ):
-    def test_ptq(self):
+    def func_ptq(self):
         start_time = time.time()
 
         self.set_vars()
@@ -305,6 +311,11 @@ class TestImperativePTQfuse(TestImperativePTQ):
         end_time = time.time()
         print("total time: %ss \n" % (end_time - start_time))
 
+    def test_ptq(self):
+        with _test_eager_guard():
+            self.func_ptq()
+        self.func_ptq()
+
 
 class TestImperativePTQHist(TestImperativePTQ):
     def set_vars(self):

python/paddle/fluid/contrib/slim/tests/test_imperative_qat.py

@@ -32,7 +32,7 @@ from paddle.nn import Linear, Conv2D, Softmax, Conv2DTranspose
 from paddle.fluid.log_helper import get_logger
 from paddle.fluid.dygraph.io import INFER_MODEL_SUFFIX, INFER_PARAMS_SUFFIX
 from paddle.nn.quant.quant_layers import QuantizedConv2D, QuantizedConv2DTranspose
-
+from paddle.fluid.framework import _test_eager_guard
 from imperative_test_utils import fix_model_dict, ImperativeLenet
 
 paddle.enable_static()
@@ -55,7 +55,7 @@ class TestImperativeQat(unittest.TestCase):
         self.activation_quantize_type = 'moving_average_abs_max'
         print('weight_quantize_type', self.weight_quantize_type)
 
-    def test_qat(self):
+    def func_qat(self):
         self.set_vars()
 
         imperative_qat = ImperativeQuantAware(
@@ -193,6 +193,11 @@ class TestImperativeQat(unittest.TestCase):
                 np.allclose(after_save, before_save.numpy()),
                 msg='Failed to save the inference quantized model.')
 
+    def test_qat(self):
+        with _test_eager_guard():
+            self.func_qat()
+        self.func_qat()
+
 
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/contrib/slim/tests/test_imperative_qat_amp.py

@@ -27,7 +27,7 @@ import paddle.fluid as fluid
 from paddle.fluid.contrib.slim.quantization import ImperativeQuantAware
 from paddle.fluid.log_helper import get_logger
 from paddle.dataset.common import download
-
+from paddle.fluid.framework import _test_eager_guard
 from imperative_test_utils import fix_model_dict, ImperativeLenet
 
 os.environ["CPU_NUM"] = "1"

python/paddle/fluid/contrib/slim/tests/test_imperative_qat_user_defined.py

@@ -30,7 +30,7 @@ from paddle.fluid.dygraph import Pool2D
 from paddle.fluid.dygraph import Linear
 from paddle.nn.quant.quant_layers import QuantizedConv2DTranspose
 from paddle.fluid.log_helper import get_logger
-
+from paddle.fluid.framework import _test_eager_guard
 os.environ["CPU_NUM"] = "1"
 
 _logger = get_logger(
@@ -157,7 +157,7 @@ class TestUserDefinedActPreprocess(unittest.TestCase):
         _logger.info("test act_preprocess")
         self.imperative_qat = ImperativeQuantAware(act_preprocess_layer=PACT)
 
-    def test_quant_aware_training(self):
+    def func_quant_aware_training(self):
         imperative_qat = self.imperative_qat
         seed = 1
         np.random.seed(seed)
@@ -243,6 +243,11 @@ class TestUserDefinedActPreprocess(unittest.TestCase):
         train(lenet)
         test(lenet)
 
+    def test_quant_aware_training(self):
+        with _test_eager_guard():
+            self.func_quant_aware_training()
+        self.func_quant_aware_training()
+
 
 class TestUserDefinedWeightPreprocess(TestUserDefinedActPreprocess):
     def setUp(self):

python/paddle/fluid/contrib/slim/tests/test_imperative_skip_op.py

@@ -32,6 +32,7 @@ from paddle.fluid.dygraph.nn import Pool2D
 from paddle.fluid.log_helper import get_logger
 
 from imperative_test_utils import fix_model_dict, train_lenet, ImperativeLenetWithSkipQuant
+from paddle.fluid.framework import _test_eager_guard
 
 os.environ["CPU_NUM"] = "1"
 if core.is_compiled_with_cuda():
@@ -42,7 +43,8 @@ _logger = get_logger(
 
 
 class TestImperativeOutSclae(unittest.TestCase):
-    def test_out_scale_acc(self):
+    def func_out_scale_acc(self):
+        paddle.disable_static()
         seed = 1000
         lr = 0.1
 
@@ -125,6 +127,11 @@ class TestImperativeOutSclae(unittest.TestCase):
         if find_matmul:
             self.assertTrue(matmul_skip_count == 1)
 
+    def test_out_scale_acc(self):
+        with _test_eager_guard():
+            self.func_out_scale_acc()
+        self.func_out_scale_acc()
+
 
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/dygraph/base.py

@@ -99,18 +99,19 @@ def param_guard(parameters):
         yield
 
 
-def _convert_into_variable(var_base):
+def _convert_into_variable(tensor):
     """
     Convert Varbase into Variable.
     """
-    if isinstance(var_base, core.VarBase):
+    if isinstance(tensor, (core.eager.Tensor, core.VarBase)):
         # Check whether has been created before.
-        new_var = var_base.block._find_var_recursive(var_base.name)
+        new_var = tensor.block._find_var_recursive(tensor.name)
         if new_var is not None:
             assert isinstance(new_var, framework.Variable)
         # Convert ParamBase into Parameter with same attributes in dy2stat.
-        elif isinstance(var_base, framework.ParamBase):
-            new_var = var_base._to_static_var(to_parameter=True)
+        elif isinstance(tensor,
+                        (framework.EagerParamBase, framework.ParamBase)):
+            new_var = tensor._to_static_var(to_parameter=True)
         else:
             # Note(Aurelius84): Convert VarBase in self._buffers into Variable with
             # same attributes and set persistable=True to allow saving this var.
@@ -120,13 +121,13 @@ def _convert_into_variable(var_base):
 
             # But if its shape is empty while created from `create_variable()`, we consider this buffer
             # non-persistable. See case of `drop_state` in lstm api.
-            is_persistable = len(var_base.shape) > 0
+            is_persistable = len(tensor.shape) > 0
 
-            new_var = var_base._to_static_var(
+            new_var = tensor._to_static_var(
                 to_parameter=False, persistable=is_persistable)
         return new_var
     else:
-        return var_base
+        return tensor
 
 
 def enabled():

python/paddle/fluid/dygraph/dygraph_to_static/partial_program.py

@@ -61,7 +61,8 @@ class NestSequence(object):
     def _get_var_ids(self):
         var_ids = []
         for idx, var in enumerate(self.__input_list):
-            if isinstance(var, (framework.Variable, core.VarBase)):
+            if isinstance(var, (framework.Variable, core.VarBase,
+                                core.eager.Tensor)):
                 var_ids.append(idx)
 
         return var_ids
@@ -73,7 +74,8 @@ class NestSequence(object):
         if need_check:
             warning_types = set()
             for var in self.__input_list:
-                if not isinstance(var, (framework.Variable, core.VarBase)):
+                if not isinstance(var, (framework.Variable, core.VarBase,
+                                        core.eager.Tensor)):
                     warning_types.add(type(var))
             if warning_types:
                 logging_utils.warn(
@@ -301,10 +303,17 @@ class PartialProgramLayer:
             for name in block.vars:
                 if "@GRAD" in name:
                     var_desc = block.vars[name].desc
-                    var_base = core.VarBase(var_desc.dtype(),
-                                            var_desc.shape(),
-                                            var_desc.name(),
-                                            var_desc.type(), False)
+                    var_base = None
+                    if not core._in_eager_mode():
+                        var_base = core.VarBase(var_desc.dtype(),
+                                                var_desc.shape(),
+                                                var_desc.name(),
+                                                var_desc.type(), False)
+                    else:
+                        var_base = core.eager.Tensor(var_desc.dtype(),
+                                                     var_desc.shape(),
+                                                     var_desc.name(),
+                                                     var_desc.type(), False)
                     double_grads.append(var_base)
         return self._valid_vars(double_grads)
 
@@ -386,13 +395,22 @@ class PartialProgramLayer:
         expected_place = framework._current_expected_place()
         for i, value in enumerate(flatten_inputs):
             if isinstance(value, np.ndarray):
-                var = core.VarBase(
-                    value=value,
-                    name=self._inputs[i].desc.name(),
-                    persistable=False,
-                    place=expected_place,
-                    zero_copy=True)
-            elif isinstance(value, core.VarBase):
+                var = None
+                if not core._in_eager_mode():
+                    var = core.VarBase(
+                        value=value,
+                        name=self._inputs[i].desc.name(),
+                        persistable=False,
+                        place=expected_place,
+                        zero_copy=True)
+                else:
+                    var = core.eager.Tensor(
+                        value=value,
+                        name=self._inputs[i].desc.name(),
+                        persistable=False,
+                        place=expected_place,
+                        zero_copy=True)
+            elif isinstance(value, (core.VarBase, core.eager.Tensor)):
                 # NOTE(Aurelius84): If var is on CPUPlace, it will be transformed multi times
                 # into CUDAPlace when it's as input of multi Ops. so we move it in advance
                 # to avoid this problem.
@@ -411,9 +429,16 @@ class PartialProgramLayer:
             var = self._outputs[var_id]
             assert isinstance(var, framework.Variable)
             var_desc = var.desc
-            var_base = core.VarBase(var_desc.dtype(),
-                                    var_desc.shape(),
-                                    var_desc.name(), var_desc.type(), False)
+            varbase = None
+            if not core._in_eager_mode():
+                var_base = core.VarBase(var_desc.dtype(),
+                                        var_desc.shape(),
+                                        var_desc.name(), var_desc.type(), False)
+            else:
+                var_base = core.eager.Tensor(var_desc.dtype(),
+                                             var_desc.shape(),
+                                             var_desc.name(),
+                                             var_desc.type(), False)
             return var_base
 
         # Create VarBase to receive output data.
@@ -423,12 +448,19 @@ class PartialProgramLayer:
 
     def _create_scope_vec(self):
         # Hold forward variables
-        tmp_scope_vec = core.VarBase(core.VarDesc.VarType.FP32, [],
-                                     "program_out_scope",
-                                     core.VarDesc.VarType.STEP_SCOPES, True)
-
-        inner_scope = core.Scope()
-        tmp_scope_vec.value().set_scope(inner_scope)
+        tmp_scope_vec = None
+        if not core._in_eager_mode():
+            tmp_scope_vec = core.VarBase(core.VarDesc.VarType.FP32, [],
+                                         "program_out_scope",
+                                         core.VarDesc.VarType.STEP_SCOPES, True)
+            # TODO(jiabin): Support this later.
+            # else:
+            #     tmp_scope_vec = core.eager.Tensor(core.VarDesc.VarType.FP32, [],
+            #                                 "program_out_scope",
+            #                                 core.VarDesc.VarType.STEP_SCOPES, True)
+
+            inner_scope = core.Scope()
+            tmp_scope_vec.value().set_scope(inner_scope)
         return tmp_scope_vec
 
     def _restore_out(self, out_vars):
@@ -450,7 +482,8 @@ class PartialProgramLayer:
         return main_program.clone(for_test=True)
 
     def _is_no_value(self, var):
-        if isinstance(var, core.VarBase) and var.shape == [1]:
+        if isinstance(var,
+                      (core.VarBase, core.eager.Tensor)) and var.shape == [1]:
             # NOTE: .numpy() will insert MemcpySync operation, it hits performance.
             if var.numpy()[0] == RETURN_NO_VALUE_MAGIC_NUM:
                 return True
@@ -460,7 +493,7 @@ class PartialProgramLayer:
         """
         Removes invalid value for various-length return statement
         """
-        if isinstance(out_vars, core.VarBase):
+        if isinstance(out_vars, (core.VarBase, core.eager.Tensor)):
             if self._is_no_value(out_vars):
                 return None
             return out_vars
@@ -527,7 +560,7 @@ class PartialProgramLayer:
         param_and_buffer_names_set = set()
         for i, var in enumerate(self._params):
             # self._params constains parameters and buffers with persistable=True.
-            if not isinstance(var, core.VarBase):
+            if not isinstance(var, (core.VarBase, core.eager.Tensor)):
                 raise TypeError(
                     'Type of self._params[{}] in PartialProgramLayer should be Parameter or Variable, but received {}.'.
                     format(i, type(var)))
@@ -559,10 +592,18 @@ def _create_fake_var():
     """
     Create a fake_var (force on CPU) to handle empty input or output
     """
-    return [
-        core.VarBase(core.VarDesc.VarType.FP32, [], "Fake_var",
-                     core.VarDesc.VarType.RAW, False)
-    ]
+    if not core._in_eager_mode():
+        return [
+            core.VarBase(core.VarDesc.VarType.FP32, [], "Fake_var",
+                         core.VarDesc.VarType.RAW, False)
+        ]
+    else:
+        return []
+        # TODO(jiabin): Support this later
+        # return [
+        #     core.eager.Tensor(core.VarDesc.VarType.FP32, [], "Fake_var",
+        #                 core.VarDesc.VarType.RAW, False)
+        # ]
 
 
 def partial_program_from(concrete_program):

python/paddle/fluid/dygraph/jit.py

@@ -25,7 +25,7 @@ import threading
 
 import six
 import paddle
-from paddle.fluid import core
+from paddle.fluid import core, dygraph
 from paddle.fluid.compiler import BuildStrategy, CompiledProgram, ExecutionStrategy
 from paddle.fluid.data_feeder import check_type
 from paddle.fluid.layers.utils import flatten, pack_sequence_as
@@ -898,30 +898,33 @@ def save(layer, path, input_spec=None, **configs):
                 state_var_dict[var.name] = var
 
             # 3. share parameters from Layer to scope & record var info
-            for param_or_buffer in concrete_program.parameters:
-                # share to scope
-                if param_or_buffer.type == core.VarDesc.VarType.VOCAB:
-                    scr_tensor = param_or_buffer.value().get_map_tensor()
-                    tgt_var = scope.var(param_or_buffer.name)
-                    tgt_var.set_vocab(scr_tensor)
-                else:
-                    param_or_buffer_tensor = scope.var(
-                        param_or_buffer.name).get_tensor()
-                    #src_tensor = param_or_buffer.value().get_tensor()
-                    src_tensor = state_var_dict[param_or_buffer.name].value(
-                    ).get_tensor()
-                    param_or_buffer_tensor._share_data_with(src_tensor)
-                # record var info
-                if param_or_buffer.name not in extra_var_info:
-                    extra_info_dict = dict()
-                    if param_or_buffer.name in state_names_dict:
-                        extra_info_dict['structured_name'] = state_names_dict[
-                            param_or_buffer.name]
-                    extra_info_dict[
-                        'stop_gradient'] = param_or_buffer.stop_gradient
-                    if isinstance(param_or_buffer, ParamBase):
-                        extra_info_dict['trainable'] = param_or_buffer.trainable
-                    extra_var_info[param_or_buffer.name] = extra_info_dict
+            with dygraph.guard():
+                for param_or_buffer in concrete_program.parameters:
+                    # share to scope
+                    if param_or_buffer.type == core.VarDesc.VarType.VOCAB:
+                        scr_tensor = param_or_buffer.value().get_map_tensor()
+                        tgt_var = scope.var(param_or_buffer.name)
+                        tgt_var.set_vocab(scr_tensor)
+                    else:
+                        param_or_buffer_tensor = scope.var(
+                            param_or_buffer.name).get_tensor()
+                        #src_tensor = param_or_buffer.value().get_tensor()
+                        src_tensor = state_var_dict[param_or_buffer.name].value(
+                        ).get_tensor()
+                        param_or_buffer_tensor._share_data_with(src_tensor)
+                    # record var info
+                    if param_or_buffer.name not in extra_var_info:
+                        extra_info_dict = dict()
+                        if param_or_buffer.name in state_names_dict:
+                            extra_info_dict[
+                                'structured_name'] = state_names_dict[
+                                    param_or_buffer.name]
+                        extra_info_dict[
+                            'stop_gradient'] = param_or_buffer.stop_gradient
+                        if isinstance(param_or_buffer, ParamBase):
+                            extra_info_dict[
+                                'trainable'] = param_or_buffer.trainable
+                        extra_var_info[param_or_buffer.name] = extra_info_dict
 
         # 4. build input & output of save_infernece_model
         # NOTE(chenweihang): [ Get input variables name ]

python/paddle/fluid/dygraph/varbase_patch_methods.py

@@ -94,7 +94,7 @@ def monkey_patch_varbase():
         # Note: getattr(self, attr, None) will call x.grad=x.gradient(), but gradient() only available in dygraph.
         # It will fail. So, for propery that different between dynamic and static graph, should not getattr(self, attr, None).
         attr_not_need_keys = ['grad', 'T']
-        if isinstance(self, ParamBase):
+        if isinstance(self, (ParamBase, EagerParamBase)):
             attr_kwargs = self.__dict__.copy()
         else:
             attr_names = []
@@ -111,7 +111,7 @@ def monkey_patch_varbase():
 
         attr_kwargs.update(kwargs)
 
-        if to_parameter or isinstance(self, ParamBase):
+        if to_parameter or isinstance(self, (ParamBase, EagerParamBase)):
             del attr_kwargs['persistable']
             # NOTE(Aurelius84): All parameters should be placed into global block.
             attr_kwargs['block'] = attr_kwargs['block'].program.global_block()

python/paddle/fluid/io.py

@@ -1821,7 +1821,7 @@ def _pack_loaded_dict(load_obj):
 @static_only
 def _legacy_save(param_dict, model_path, protocol=2):
     def get_tensor(var):
-        if isinstance(var, core.VarBase):
+        if isinstance(var, (core.VarBase, core.eager.Tensor)):
             return var.numpy()
         elif isinstance(var, core.LoDTensor):
             return np.array(var)

python/paddle/fluid/layers/nn.py

@@ -10148,6 +10148,9 @@ def flatten(x, axis=1, name=None):
     check_variable_and_dtype(
         x, 'x', ['float32', 'float64', 'int8', 'int32', 'int64', 'uint8'],
         'flatten')
+    if in_dygraph_mode():
+        return _C_ops.flatten2(x, 'axis', axis)[0]
+
     helper = LayerHelper('flatten', **locals())
 
     if not (isinstance(x, Variable)):

python/paddle/fluid/layers/tensor.py

@@ -663,7 +663,9 @@ def assign(input, output=None):
             })
 
     if is_inplace and in_dygraph_mode():
-        output._bump_inplace_version()
+        # TODO(jiabin): Remove this when we support inplace
+        if not core._in_eager_mode():
+            output._bump_inplace_version()
 
     return output
 

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

@@ -771,13 +771,13 @@ class EagerVariablePropertiesAndMethodsTestCase(unittest.TestCase):
             self.assertTrue(np.array_equal(egr_tensor.numpy(), ori_arr))
             ori_place = egr_tensor.place
 
-            new_arr = np.random.rand(4, 4, 16, 32).astype('float32')
+            new_arr = np.random.rand(4, 16, 16, 32).astype('float32')
             self.assertFalse(np.array_equal(egr_tensor.numpy(), new_arr))
 
-            egr_tensor._set_value(new_arr)
+            egr_tensor.set_value(new_arr)
             self.assertEqual(egr_tensor.stop_gradient, True)
             self.assertTrue(egr_tensor.place._equals(ori_place))
-            self.assertEqual(egr_tensor.shape, [4, 4, 16, 32])
+            self.assertEqual(egr_tensor.shape, [4, 16, 16, 32])
             self.assertTrue(np.array_equal(egr_tensor.numpy(), new_arr))
 
 
@@ -880,7 +880,7 @@ class EagerParamBaseUsageTestCase(unittest.TestCase):
             new_weight = np.ones([1, 3]).astype('float32')
             self.assertFalse(np.array_equal(linear.weight.numpy(), new_weight))
 
-            linear.weight._set_value(new_weight)
+            linear.weight.set_value(new_weight)
             self.assertTrue(np.array_equal(linear.weight.numpy(), new_weight))
             self.assertTrue(linear.weight.place._equals(ori_place))
 

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

@@ -533,12 +533,8 @@ class TestTensorRegisterHook(unittest.TestCase):
             size=[self.batch_size, self.in_size]).astype('float32')
         data_t = paddle.to_tensor(data)
 
-        if _in_eager_mode():
-            with self.assertRaises(TypeError):
-                out = jit_net(data_t)
-        else:
-            with self.assertRaises(AssertionError):
-                out = jit_net(data_t)
+        with self.assertRaises(AssertionError):
+            out = jit_net(data_t)
 
     def test_register_hook_in_dy2static_mode(self):
         with _test_eager_guard():