Merge branch 'develop' of github.com:baidu/Paddle into feature/pybind_for_protobuf_desc

doc/howto/dev/new_op_en.md

@@ -1,14 +1,17 @@
 # How to write a new operator
 
- - [Background](#Background)
- - [Implementing C++ Types](#Implementing_C++_Types)
-   - [Defining ProtoMaker](#Defining_ProtoMaker)
-   - [Defining Operator](#Defining_Operator)
-   - [Registering Operator](#Registering_Operator)
-   - [Compilation](#Compilation)
- - [Python Binding](#Python_Binding)
- - [Unit Tests](#Unit_Tests)
-
+ - [Background](#background)
+ - [Implementing C++ Types](#implementing-c++-types)
+   - [Defining ProtoMaker](#defining-protoMaker)
+   - [Defining Operator](#defining-operator)
+   - [Registering Operator](#registering-operator)
+   - [Compilation](#compilation)
+ - [Python Binding](#python-binding)
+ - [Unit Tests](#unit-tests)
+   - [Testing Forward Operators](#testing-forward-operators)
+   - [Testing Backward Operators](#testing-backward-operators)
+   - [Compiling and Running](#compiling-and-running)
+ - [Remarks](#remarks)
 ## Background
 
 Here are the base types needed. For details, please refer to the design docs.
@@ -232,4 +235,122 @@ The system will automatically bind to Python and link it to a generated library.
 
 ## Unit Tests
 
-Unit tests include comparing a forward operator's implementations on different devices, comparing a backward operator's implementation on different devices, and a scaling test for the backward operator. Here, we introduce the [unit tests for `MulOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py).
+Unit tests for an operator include
+
+1. comparing a forward operator's implementations on different devices,
+
+2. comparing a backward operator's implementation on different devices, and
+
+3. a scaling test for the backward operator.
+
+Here, we introduce the [unit tests for `MulOp`](https://github.com/PaddlePaddle/Paddle/blob/develop/python/paddle/v2/framework/tests/test_mul_op.py).
+
+### Testing Forward Operators
+
+A forward operator unit test inherits `unittest.TestCase` and defines metaclass `__metaclass__ = OpTestMeta`. More concrete tests are performed in `OpTestMeta`. Testing a forward operator requires the following:
+
+1. Defining input, output and relevant attributes in `setUp` method.
+
+2. Generating random input data.
+
+3. Implementing the same computation logic in a Python script:
+
+  ```python
+  import unittest
+  import numpy as np
+  from gradient_checker import GradientChecker, create_op
+  from op_test_util import OpTestMeta
+
+  class TestMulOp(unittest.TestCase):
+      __metaclass__ = OpTestMeta
+
+      def setUp(self):
+          self.type = "mul"
+          self.inputs = {
+              'X': np.random.random((32, 84)).astype("float32"),
+              'Y': np.random.random((84, 100)).astype("float32")
+          }
+          self.outputs = {'Out': np.dot(self.inputs['X'], self.inputs['Y'])}
+  ```
+Get its output, and compare it with the forward operator's own output.
+
+The code above first loads required packages. In addition, we have
+
+- `self.type = "mul" ` defines the type that is identical to what the operator's registered type.
+- `self.inputs` defines input, with type `numpy.array` and initializes it.
+- `self.outputs` defines output and completes the same operator computation in the Python script, and returns its result from the Python script.
+
+### Testing Backward Operators
+
+A backward operator unit test inherits `GradientChecker`, which inherits `unittest.TestCase`. As a result, **a backward operator unit test needs to be have the prefix `test_`**.
+
+```python
+class TestMulGradOp(GradientChecker):
+    def setUp(self):
+        self.op = create_op("mul")
+        self.inputs = {
+            'X': np.random.random((32, 84)).astype("float32"),
+            'Y': np.random.random((84, 100)).astype("float32")
+        }
+
+    def test_cpu_gpu_compare(self):
+        self.compare_grad(self.op, self.inputs)
+
+    def test_normal(self):
+        # mul op will enlarge the relative error
+        self.check_grad(
+            self.op, self.inputs, ["X", "Y"], "Out", max_relative_error=0.5)
+
+    def test_ignore_x(self):
+        self.check_grad(
+            self.op,
+            self.inputs, ["Y"],
+            "Out",
+            max_relative_error=0.5,
+            no_grad_set={"X"})
+
+    def test_ignore_y(self):
+        self.check_grad(
+            self.op,
+            self.inputs, ["X"],
+            "Out",
+            max_relative_error=0.5,
+            no_grad_set={"Y"})
+```
+
+Some key points in the code above include:
+
+- `create_op("mul")` creates the backward operator's corresponding forward operator.
+- `compare_grad` compares results between utilizing the CPU and the GPU.
+- `test_normal` calls `check_grad` to validate scaling tests' correctness and stability through numeric methods.
+  - The first variable `self.op` denotes the forward operator.
+  - The second variable `self.inputs` denotes the input dictionary, which has its key value identical to its `ProtoMaker` definitions.
+  - The third variable `["X", "Y"]` appoints `X` and `Y` to be scale tested.
+  - The fourth variable `"Out"` points to the network's final output target `Out`.
+- `test_ignore_x` and `test_ignore_y`branches test the cases where there is only one scaling input.
+
+### Compiling and Running
+
+
+Any new unit testing file of the format `test_*.py`  added to the director `python/paddle/v2/framework/tests` is automatically added to the project to compile.
+
+Note that **unlike the compile test for Ops, running unit tests requires compiling the entire project** and requires compiling with flag `WITH_TESTING` on i.e. `cmake paddle_dir -DWITH_TESTING=ON`.
+
+After successfully compiling the project, run the following command to run unit tests:
+
+```bash
+make test ARGS="-R test_mul_op -V"
+```
+
+Or,
+
+```bash
+ctest -R test_mul_op
+```
+
+## Remarks
+
+- Every `*_op.h` (if applicable), `*_op.cc`, and `*_op.cu` (if applicable) must be created for a unique Op. Compiling will fail if multiple operators are included per file.
+- The type with which an operator is registered needs to be identical to the Op's name. Registering `REGISTER_OP(B, ...)` in `A_op.cc` will cause unit testing failures.
+- If the operator does not implement a GPU kernel, please refrain from creating an empty `*_op.cu` file, or else unit tests will fail.
+- If multiple operators rely on some shared methods, a file NOT named `*_op.*` can be created to store them, such as `gather.h`.

paddle/framework/CMakeLists.txt

@@ -26,7 +26,7 @@ cc_library(operator SRCS operator.cc DEPS op_info device_context tensor scope)
 cc_test(operator_test SRCS operator_test.cc DEPS operator op_registry)
 
 cc_library(grad_op_builder SRCS grad_op_builder.cc DEPS operator)
-cc_library(op_registry SRCS op_registry.cc DEPS grad_op_builder op_proto_maker)
+cc_library(op_registry SRCS op_registry.cc DEPS grad_op_builder op_proto_maker op_info)
 cc_test(op_registry_test SRCS op_registry_test.cc DEPS op_registry)
 cc_test(grad_op_builder_test SRCS grad_op_builder_test.cc DEPS grad_op_builder op_registry add_op)
 

paddle/framework/attribute.h

@@ -45,6 +45,21 @@ inline AttrType AttrTypeID() {
 
 Attribute GetAttrValue(const OpDesc::Attr& attr_desc);
 
+class AttrReader {
+ public:
+  explicit AttrReader(const AttributeMap& attrs) : attrs_(attrs) {}
+
+  template <typename T>
+  inline const T& Get(const std::string& name) const {
+    PADDLE_ENFORCE(attrs_.count(name) != 0, "%s should be in AttributeMap",
+                   name);
+    return boost::get<T>(attrs_.at(name));
+  }
+
+ private:
+  const AttributeMap& attrs_;
+};
+
 // check whether a value(attribute) fit a certain limit
 template <typename T>
 class GreaterThanChecker {

paddle/framework/backward.md

@@ -2,7 +2,7 @@
 
 ## Motivation
 
-In Neural Network, many model is solved by the the backpropagation algorithm(known as BP) at present. Technically it caculates the gradient of the loss function, then distributed back through the networks. Follows the chain rule, so we need a module chains the gradient operators/expressions together with to construct the backward pass. Every forward network needs a backward network to construct the full computation graph, the operator/expression's backward pass will be generated respect to forward pass. 
+In Neural Network, most models are solved by the backpropagation algorithm(known as **BP**) at present. Technically, BP calculates the gradient of the loss function, then propagates it back through the networks following the chain rule. Hence we need a module that chains the gradient operators/expressions together to construct the backward pass. Every forward network needs a backward network to construct the full computation graph. The operator/expression's backward pass will be generated with respect to the forward pass. 
 
 ## Implementation
 
@@ -24,9 +24,9 @@ A backward network is built up with several backward operators. Backward operato
 | **Operator::inputs_**  | Inputs       | Inputs, Outputs, OutputGradients |	
 | **Operator::outputs_** | Outputs          | InputGradients            |
 
- In most cases, there is a one-to-one correspondence between the forward and backward operators. These correspondences are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and make operators pluggable, the registry mechanism is introduced.
+ In most cases, there is a one-to-one relation between the forward and backward operators. These relations are recorded by a global hash map(`OpInfoMap`). To follow the philosophy of minimum core and to make operators pluggable, the registry mechanism is introduced.
 
-For example, we have got a `mul_op`, and we can register its information and corresponding backward operator by the following macro:
+For example, we have `mul_op`, and we can register its information and corresponding backward operator by the following macro:
 
 ```cpp
 REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad);
@@ -48,7 +48,7 @@ The function `BuildGradOp` will sequentially execute following processes:
 
 1. Get the `type_` of given forward operator, and then get the corresponding backward operator's type by looking up the `OpInfoMap`.
 
-2. Build two maps named `inputs` and `outputs` to temporary storage backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these, are not necessary for gradient computing.
+2. Build two maps named `inputs` and `outputs` to temporarily store backward operator's inputs and outputs. Copy forward operator's `inputs_` and `outputs_` to map `inputs`, except these, are not necessary for gradient computing.
 
 3. Add forward inputs' gradient variables into map `output`, adding forward outputs' gradient variables into map `input`.
 
@@ -56,11 +56,11 @@ The function `BuildGradOp` will sequentially execute following processes:
 
 ### Backward Network Building
 
-A backward network is a series of backward operators. The main idea of building a backward network is creating backward operators in the inverted sequence and append them together one by one. There is some corner case need to process specially.
+A backward network is a series of backward operators. The main idea of building a backward network is creating backward operators in the inverted sequence and appending them together one by one. There are some corner cases that need special processing.
 
 1. Op 
 
-   When the input forward network is an Op, return its gradient Operator Immediately. If all of its outputs are in no gradient set, then return a special `NOP`.
+   When the input forward network is an Op, return its gradient Operator immediately. If all of its outputs are in no gradient set, then return a special `NOP`.
 
 2. NetOp 
 
@@ -68,33 +68,33 @@ A backward network is a series of backward operators. The main idea of building
 
 3. RnnOp
 
-   RnnOp is a nested stepnet operator.  Backward module need to recusively call `Backward` for every stepnet.
+   RnnOp is a nested stepnet operator.  Backward module needs to recusively call `Backward` for every stepnet.
 
 4. Sharing Variables
 
-   **sharing variables**. As illustrated in the pictures, two operator's share the same variable name of W@GRAD, which will overwrite their sharing input variable. 
+   As illustrated in the figure 1 and figure 2, two operators share the same variable name **W@GRAD**, which will overwrite their shared input variable. 
 
 <p align="center">
 <img src="./images/duplicate_op.png" width="50%" ><br/>
 
-​	pic 1. Sharing variables in operators. 
+​	Figure 1. Sharing variables in operators. 
 
 </p>
 
-​	Sharing variable between operators or same input variable used in multiple operators leads to a duplicate gradient variable. As demo show above, we need to rename gradient name recursively and add a generic add operator to replace the overwrite links. 
+​	Sharing variable between operators or same input variable used in multiple operators can lead to duplicate gradient variables. As illustrated in figure 2, we need to rename the gradient names recursively and add a generic add operator to prevent overwriting. 
 
 <p align="center">
 <img src="images/duplicate_op2.png" width="40%" ><br/>
 
-​	pic 2. Replace sharing variable's gradient with `Add` operator.
+​	Figure 2. Replace sharing variable's gradient with `Add` operator.
 
 </p>
 
-​	Because our framework finds variables accord to their names, we need to rename the output links. We add a suffix of number to represent its position in clockwise. 
+​	Because the framework finds variables according to their names, we need to rename the output links. We add an integer suffix to represent its position in the clockwise direction. 
 
-5. Part of Gradient is Zero.
+5. Part of the Gradient is Zero.
 
-   In the whole graph, there is some case of that one operator's gradient is not needed, but its input's gradient is a dependency link of other operator,  we need to fill a same shape gradient matrix in the position. In our implement, we insert a special `fillZeroLike` operator.
+   In the whole graph, there is some case of that one operator's gradient is not needed, but its input's gradient is a dependency link of other operator,  we need to fill a same shape gradient matrix in the position. In our implementation, we insert a special `fillZeroLike` operator.
 
 
 Follow these rules above, then collect the sub graph `OutputGradients`/`InputGradients` as the NetOp's and return it.

paddle/framework/op_registry_test.cc

@@ -174,4 +174,4 @@ TEST(OpRegistry, CustomChecker) {
   op->Run(scope, dev_ctx);
   int test_attr = op->Attr<int>("test_attr");
   ASSERT_EQ(test_attr, 4);
-}
+}

paddle/framework/operator.cc

@@ -14,7 +14,6 @@ limitations under the License. */
 
 #include "paddle/framework/operator.h"
 #include <algorithm>
-#include "paddle/framework/op_registry.h"
 
 namespace paddle {
 namespace framework {
@@ -33,6 +32,24 @@ ExecutionContext::GetEigenDevice<platform::GPUPlace, Eigen::GpuDevice>() const {
 }
 #endif
 
+const Tensor* GetTensorFromVar(const Variable* var) {
+  if (var->IsType<LoDTensor>()) {
+    return &var->Get<LoDTensor>();
+  }
+  PADDLE_ENFORCE(var->IsType<Tensor>(),
+                 "The Input must be LoDTensor or Tensor.");
+  return &var->Get<Tensor>();
+}
+
+Tensor* GetTensorFromVar(Variable* var) {
+  if (var->IsType<LoDTensor>()) {
+    return var->GetMutable<LoDTensor>();
+  }
+  PADDLE_ENFORCE(var->IsType<Tensor>(),
+                 "The Input must be LoDTensor or Tensor.");
+  return var->GetMutable<Tensor>();
+}
+
 std::string OperatorBase::Input(const std::string& name) const {
   auto& ins = Inputs(name);
   PADDLE_ENFORCE_LE(ins.size(), 1UL,

paddle/framework/operator.h

@@ -24,6 +24,7 @@ limitations under the License. */
 #include "paddle/framework/framework.pb.h"
 #include "paddle/framework/lod_tensor.h"
 #include "paddle/framework/scope.h"
+#include "paddle/framework/shape_inference.h"
 #include "paddle/framework/tensor.h"
 #include "paddle/platform/device_context.h"
 #include "paddle/platform/place.h"
@@ -56,6 +57,9 @@ class OperatorBase;
 class InferShapeContext;
 class ExecutionContext;
 
+extern const Tensor* GetTensorFromVar(const Variable* var);
+extern Tensor* GetTensorFromVar(Variable* var);
+
 /**
  * OperatorBase has the basic element that Net will call to do computation.
  * Only CreateOperator from OpRegistry will new Operator directly. User
@@ -262,15 +266,6 @@ class InferShapeContext {
     return res;
   }
 
-  const Tensor* GetTensorFromVar(const Variable* var) const {
-    if (var->IsType<LoDTensor>()) {
-      return &var->Get<LoDTensor>();
-    }
-    PADDLE_ENFORCE(var->IsType<Tensor>(),
-                   "The Input(%s) must be LoDTensor or Tensor.");
-    return &var->Get<Tensor>();
-  }
-
   void ShareLoD(const std::string& in, const std::string& out, size_t i = 0,
                 size_t j = 0) const {
     PADDLE_ENFORCE_LT(i, InputSize(in));
@@ -340,6 +335,78 @@ class ExecutionContext : public InferShapeContext {
   const platform::DeviceContext& device_context_;
 };
 
+class RuntimeInferShapeContext : public InferShapeContextBase {
+ public:
+  RuntimeInferShapeContext(const OperatorBase& op, const Scope& scope)
+      : op_(op), scope_(scope) {}
+
+  bool HasInput(const std::string& name) const {
+    auto ipt = op_.Input(name);
+    auto* var = ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt);
+    return var != nullptr;
+  }
+
+  bool HasOutput(const std::string& name) const {
+    auto ipt = op_.Output(name);
+    auto* var = ipt == kEmptyVarName ? nullptr : scope_.FindVar(ipt);
+    return var != nullptr;
+  }
+
+  DDim GetInputDim(const std::string& name) const {
+    return GetDim(op_.Input(name));
+  }
+
+  void SetInputDim(const std::string& name, const DDim& dim) {
+    SetDim(op_.Input(name), dim);
+  }
+
+  DDim GetOutputDim(const std::string& name) const {
+    return GetDim(op_.Output(name));
+  }
+
+  void SetOutputDim(const std::string& name, const DDim& dim) {
+    SetDim(op_.Output(name), dim);
+  }
+
+  AttrReader Attrs() const { return AttrReader(op_.Attrs()); }
+
+  const std::vector<std::string>& Inputs(const std::string& name) const {
+    return op_.Inputs(name);
+  }
+
+  const std::vector<std::string>& Outputs(const std::string& name) const {
+    return op_.Outputs(name);
+  }
+
+ private:
+  template <bool Allocate>
+  Tensor* GetTensor(const std::string& name) const {
+    Tensor* t = nullptr;
+    auto* var = scope_.FindVar(name);
+    if (!var->IsType<LoDTensor>() && !var->IsType<Tensor>()) {
+      if (Allocate) {
+        t = var->GetMutable<LoDTensor>();
+      } else {
+        PADDLE_THROW("Variable(%s) should be tensor", name);
+      }
+    } else {
+      t = GetTensorFromVar(scope_.FindVar(name));
+    }
+    return t;
+  }
+
+  DDim GetDim(const std::string& name) const {
+    return GetTensor<false>(name)->dims();
+  }
+
+  void SetDim(const std::string& name, const DDim& dim) {
+    GetTensor<true>(name)->Resize(dim);
+  }
+
+  const OperatorBase& op_;
+  const Scope& scope_;
+};
+
 class OpKernel {
  public:
   /**
@@ -383,8 +450,10 @@ class OperatorWithKernel : public OperatorBase {
                      const VariableNameMap& outputs, const AttributeMap& attrs)
       : OperatorBase(type, inputs, outputs, attrs) {}
 
+  // runtime infershape
   void InferShape(const Scope& scope) const override {
-    InferShape(InferShapeContext(*this, scope));
+    auto c = RuntimeInferShapeContext(*this, scope);
+    InferShape(&c);
   }
 
   void Run(const Scope& scope,
@@ -406,7 +475,7 @@ class OperatorWithKernel : public OperatorBase {
   }
 
  protected:
-  virtual void InferShape(const InferShapeContext& ctx) const = 0;
+  virtual void InferShape(InferShapeContextBase* ctx) const = 0;
 };
 
 }  // namespace framework

paddle/framework/operator_test.cc

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #include "paddle/framework/operator.h"
 #include "gtest/gtest.h"
+#include "paddle/framework/op_info.h"
 #include "paddle/framework/op_registry.h"
 
 namespace paddle {
@@ -114,7 +115,7 @@ class OpWithKernelTest : public OperatorWithKernel {
   using OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext& ctx) const override {}
+  void InferShape(framework::InferShapeContextBase* ctx) const override {}
 };
 
 template <typename T1, typename T2>

paddle/framework/shape_inference.h

@@ -0,0 +1,82 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+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 "paddle/framework/ddim.h"
+
+namespace paddle {
+namespace framework {
+
+class InferShapeContextBase {
+ public:
+  virtual ~InferShapeContextBase() {}
+  virtual bool HasInput(const std::string &name) const = 0;
+  virtual bool HasOutput(const std::string &name) const = 0;
+  virtual framework::DDim GetInputDim(const std::string &name) const = 0;
+  std::vector<framework::DDim> GetInputsDim(const std::string &name) const {
+    const std::vector<std::string> &names = Inputs(name);
+    return GetDims(names);
+  }
+  virtual void SetInputDim(const std::string &name,
+                           const framework::DDim &dim) = 0;
+  void SetInputsDim(const std::string &name,
+                    const std::vector<framework::DDim> &dims) {
+    auto &names = Inputs(name);
+    SetDims(names, dims);
+  }
+  virtual framework::DDim GetOutputDim(const std::string &name) const = 0;
+  std::vector<framework::DDim> GetOutputsDim(const std::string &name) const {
+    const std::vector<std::string> &names = Outputs(name);
+    return GetDims(names);
+  }
+  virtual void SetOutputDim(const std::string &name, const DDim &dim) = 0;
+  void SetOutputsDim(const std::string &name,
+                     const std::vector<framework::DDim> &dims) {
+    auto &names = Outputs(name);
+    SetDims(names, dims);
+  }
+  virtual AttrReader Attrs() const = 0;
+  virtual const std::vector<std::string> &Inputs(
+      const std::string &name) const = 0;
+  virtual const std::vector<std::string> &Outputs(
+      const std::string &name) const = 0;
+  // TODO(qiao) implement this function
+  void ShareLoD(const std::string &in, const std::string &out, size_t i = 0,
+                size_t j = 0) const {}
+
+ protected:
+  virtual framework::DDim GetDim(const std::string &name) const = 0;
+  virtual void SetDim(const std::string &name, const framework::DDim &dim) = 0;
+  std::vector<framework::DDim> GetDims(
+      const std::vector<std::string> &names) const {
+    std::vector<framework::DDim> ret;
+    ret.reserve(names.size());
+    std::transform(
+        names.begin(), names.end(), std::back_inserter(ret),
+        [this](const std::string &name) { return this->GetDim(name); });
+    return ret;
+  }
+  void SetDims(const std::vector<std::string> &names,
+               const std::vector<framework::DDim> &dims) {
+    size_t length = names.size();
+    PADDLE_ENFORCE_EQ(length, dims.size());
+    for (size_t i = 0; i < length; ++i) {
+      SetDim(names[i], dims[i]);
+    }
+  }
+};
+
+}  // namespace framework
+}  // namespace paddle

paddle/framework/variable.md

@@ -7,7 +7,7 @@ Variable is also known as *blob* in MxNet and Caffe2.  It is the input and outpu
 
 For the flexibility of a DL system, a variable should be able to contain any typed value -- a tensor in most cases, but could also be some integer IDs or a scope of other variables in the case of RNN.
 
-To use the minimum amount of memory, we'd like that a variable to allocate memory when it has to, or, lazy memory allocation.  Let's take the following example:
+To use the minimum amount of memory, we would like that a variable allocates memory only when it has to, or, lazy memory allocation.  Let's take the following example:
 
 ```cpp
 Variable vr, v1, v2;
@@ -38,7 +38,7 @@ This syntax for lazy memory allocation when we call `Randomize` and `Mult`, thos
 
 To make memory allocation lazy, we cannot assume that we know the type held by a variable at definition time.  In other words, `class Variable` cannot be a template `template <T> class Variable`.
 
-Because we don't know the type `T`, we cannot save a `T*` as `Variable's` data member.  Instead, we save an interface object `Placeholder`, who can return the pointer to the saved object via `Placeholder::Ptr()` as `void*`.
+Because we don't know the type `T`, we cannot save a `T*` as `Variable's` data member.  Instead, we save an interface object `Placeholder`, which can return the pointer to the saved object via `Placeholder::Ptr()` as `void*`.
 
 But anyway, Variable needs to know `T` so could it `delete<T>(ptr)` and so could `Variable::Get` checks the expected type and the saved object's type.
 
@@ -49,4 +49,4 @@ Because `PlaceholderImpl` knows `T`, it can save and return `typeid(T)` for the
 
 ## Conclusion
 
-The technique type hiding utilizes C++ class templates, interface and derivation, and C++ RTTI (typeid).  This combination saves us from definition something like `caffe2::TypeMata`, which takes hundreds of lines of C++ code.
+The technique type hiding utilizes C++ class templates, interface and derivation, and C++ RTTI (typeid).  This combination saves us from defining something like `caffe2::TypeMeta`, which takes hundreds of lines of C++ code.

paddle/operators/accuracy_op.cc

@@ -22,25 +22,23 @@ class AccuracyOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(
-        ctx.InputVar("Inference"),
-        "Input(Inference) of AccuracyOp should not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"),
-                            "Input(Label) of AccuracyOp should not be null.");
-    PADDLE_ENFORCE_NOT_NULL(
-        ctx.OutputVar("Accuracy"),
-        "Output(Accuracy) of AccuracyOp should not be null.");
+  void InferShape(framework::InferShapeContextBase *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Inference"),
+                   "Input(Inference) of AccuracyOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Label"),
+                   "Input(Label) of AccuracyOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Accuracy"),
+                   "Output(Accuracy) of AccuracyOp should not be null.");
 
-    auto *inference = ctx.Input<framework::Tensor>("Inference");
-    auto *label = ctx.Input<framework::Tensor>("Label");
+    auto inference_dim = ctx->GetInputDim("Inference");
+    auto label_dim = ctx->GetInputDim("Label");
 
-    PADDLE_ENFORCE_EQ(label->dims().size(), 1, "label must be a vector");
-    PADDLE_ENFORCE_EQ(inference->dims()[0], label->dims()[0],
+    PADDLE_ENFORCE_EQ(label_dim.size(), 1, "label must be a vector");
+    PADDLE_ENFORCE_EQ(inference_dim[0], label_dim[0],
                       "inference size must be the same as label size");
 
-    ctx.Output<framework::Tensor>("Accuracy")->Resize({1});
-    ctx.ShareLoD("Inference", /*->*/ "Accuracy");
+    ctx->SetOutputDim("Accuracy", {1});
+    ctx->ShareLoD("Inference", /*->*/ "Accuracy");
   }
 };
 

paddle/operators/activation_op.cc

@@ -22,10 +22,9 @@ class ActivationOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    ctx.Output<framework::Tensor>("Y")->Resize(
-        ctx.Input<framework::Tensor>("X")->dims());
-    ctx.ShareLoD("X", /*->*/ "Y");
+  void InferShape(framework::InferShapeContextBase *ctx) const override {
+    ctx->SetOutputDim("Y", ctx->GetInputDim("X"));
+    ctx->ShareLoD("X", /*->*/ "Y");
   }
 };
 
@@ -34,9 +33,8 @@ class ActivationOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    ctx.Output<framework::Tensor>(framework::GradVarName("X"))
-        ->Resize(ctx.Input<framework::Tensor>("Y")->dims());
+  void InferShape(framework::InferShapeContextBase *ctx) const override {
+    ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Y"));
   }
 };
 

paddle/operators/add_op.cc

@@ -22,25 +22,23 @@ class AddOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input(X) of AddOp should not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
-                            "Input(Y) of AddOp should not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
-                            "Output(Out) of AddOp should not be null.");
+  void InferShape(framework::InferShapeContextBase* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of AddOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) of AddOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of AddOp should not be null.");
 
-    PADDLE_ENFORCE_EQ(ctx.Input<Tensor>("X")->dims(),
-                      ctx.Input<Tensor>("Y")->dims(),
+    auto x_dims = ctx->GetInputDim("X");
+    auto y_dims = ctx->GetInputDim("Y");
+    PADDLE_ENFORCE_EQ(x_dims, y_dims,
                       "Two input of Add Op's dimension must be same.");
-    ctx.Output<framework::Tensor>("Out")->Resize(
-        ctx.Input<Tensor>("X")->dims());
+    ctx->SetOutputDim("Out", x_dims);
   }
 };
 
 class AddOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  AddOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+  AddOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The first input of add op");
     AddInput("Y", "The second input of add op");
@@ -58,7 +56,7 @@ class AddOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {}
+  void InferShape(framework::InferShapeContextBase* ctx) const override {}
 };
 
 }  // namespace operators

paddle/operators/clip_op.cc

@@ -22,24 +22,24 @@ class ClipOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input(X) of ClipOp should not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
-                            "Output(Out) of ClipOp should not be null.");
-    auto x_dims = ctx.Input<Tensor>("X")->dims();
-    auto max = Attr<float>("max");
-    auto min = Attr<float>("min");
+  void InferShape(framework::InferShapeContextBase* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of ClipOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of ClipOp should not be null.");
+    auto x_dims = ctx->GetInputDim("X");
+    auto max = ctx->Attrs().Get<float>("max");
+    auto min = ctx->Attrs().Get<float>("min");
     PADDLE_ENFORCE_LT(min, max, "max should be greater than min.");
-    ctx.Output<Tensor>("Out")->Resize(x_dims);
-    ctx.ShareLoD("X", /*->*/ "Out");
+    ctx->SetOutputDim("Out", x_dims);
+    ctx->ShareLoD("X", /*->*/ "Out");
   }
 };
 
 template <typename AttrType>
 class ClipOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  ClipOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+  ClipOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X",
              "(Tensor)The input of clip op."
@@ -61,14 +61,13 @@ class ClipOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
-                            "Input(Out@GRAD) should not be null");
-    auto x_dims = ctx.Input<Tensor>("X")->dims();
-    auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
-    if (x_grad != nullptr) {
-      x_grad->Resize(x_dims);
+  void InferShape(framework::InferShapeContextBase* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
+                   "Input(Out@GRAD) should not be null");
+    auto x_dims = ctx->GetInputDim("X");
+    if (ctx->HasOutput(framework::GradVarName("X"))) {
+      ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
     }
   }
 };

paddle/operators/concat_op.cc

@@ -24,31 +24,30 @@ class ConcatOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
-                            "Output(Out) of ConcatOp should not be null.");
+  void InferShape(framework::InferShapeContextBase *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of ConcatOp should not be null.");
 
-    auto ins = ctx.MultiInput<framework::Tensor>("X");
-    auto *out = ctx.Output<framework::Tensor>("Out");
-    size_t axis = static_cast<size_t>(ctx.Attr<int>("axis"));
+    auto ins = ctx->GetInputsDim("X");
+    size_t axis = static_cast<size_t>(ctx->Attrs().Get<int>("axis"));
     size_t n = ins.size();
 
     PADDLE_ENFORCE_GT(n, 1, "Input tensors count should > 1.");
 
-    auto out_dims = ins[0]->dims();
+    auto out_dims = ins[0];
     size_t in_zero_dims_size = out_dims.size();
     for (size_t i = 1; i < n; i++) {
       for (size_t j = 0; j < in_zero_dims_size; j++) {
         if (j == axis) {
-          out_dims[axis] += ins[i]->dims()[j];
+          out_dims[axis] += ins[i][j];
           continue;
         }
-        PADDLE_ENFORCE_EQ(out_dims[j], ins[i]->dims()[j],
+        PADDLE_ENFORCE_EQ(out_dims[j], ins[i][j],
                           "Input tensors should have the same "
                           "elements except the specify axis.")
       }
     }
-    out->Resize(out_dims);
+    ctx->SetOutputDim("Out", out_dims);
   }
 };
 

paddle/operators/cond_op.cc

@@ -215,7 +215,7 @@ class CondOpProtoAndCheckerMaker : public framework::OpProtoAndCheckerMaker {
     AddComment(R"DOC(
 Sample dependent Cond Operator:
 Given Cond[i] as a 1/0 vector to indicate true/false
-The equation is: 
+The equation is:
 Out[i] = subnet_t[i], if Cond[i] == true
 Out[i] = subnet_t[i], if Cond[i] == false
 )DOC");

paddle/operators/conv2d_op.cc

@@ -27,27 +27,25 @@ class Conv2DOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Input"),
-                            "Input(Input) of Conv2DOp should not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Filter"),
-                            "Input(Filter) of Conv2DOp should not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Output"),
-                            "Output(Output) of Conv2DOp should not be null.");
-
-    auto in = ctx.Input<Tensor>("Input");
-    auto filter = ctx.Input<Tensor>("Filter");
-    auto out = ctx.Output<framework::Tensor>("Output");
-    std::vector<int> strides = Attr<std::vector<int>>("strides");
-    std::vector<int> paddings = Attr<std::vector<int>>("paddings");
-    int groups = Attr<int>("groups");
-    int input_channels = in->dims()[1];
-    int output_channels = filter->dims()[0];
-
-    PADDLE_ENFORCE_EQ(in->dims().size(), 4, "Conv2DOp input should be 4-D.");
-    PADDLE_ENFORCE_EQ(filter->dims().size(), 4,
-                      "Conv2DOp filter should be 4-D.");
-    PADDLE_ENFORCE_EQ(input_channels, filter->dims()[1] * groups,
+  void InferShape(framework::InferShapeContextBase* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("Input"),
+                   "Input(Input) of Conv2DOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Filter"),
+                   "Input(Filter) of Conv2DOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Output"),
+                   "Output(Output) of Conv2DOp should not be null.");
+
+    auto in_dims = ctx->GetInputDim("Input");
+    auto filter_dims = ctx->GetInputDim("Filter");
+    std::vector<int> strides = ctx->Attrs().Get<std::vector<int>>("strides");
+    std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
+    int groups = ctx->Attrs().Get<int>("groups");
+    int input_channels = in_dims[1];
+    int output_channels = filter_dims[0];
+
+    PADDLE_ENFORCE_EQ(in_dims.size(), 4, "Conv2DOp input should be 4-D.");
+    PADDLE_ENFORCE_EQ(filter_dims.size(), 4, "Conv2DOp filter should be 4-D.");
+    PADDLE_ENFORCE_EQ(input_channels, filter_dims[1] * groups,
                       "The number of input channels should be equal to filter "
                       "channels * groups.");
     PADDLE_ENFORCE_EQ(
@@ -55,17 +53,17 @@ class Conv2DOp : public framework::OperatorWithKernel {
         "The number of output channels should be divided by groups.");
 
     auto output_height =
-        outputSize(in->dims()[2], filter->dims()[2], paddings[0], strides[0]);
+        outputSize(in_dims[2], filter_dims[2], paddings[0], strides[0]);
     auto output_width =
-        outputSize(in->dims()[3], filter->dims()[3], paddings[1], strides[1]);
-    out->Resize(
-        {in->dims()[0], filter->dims()[0], output_height, output_width});
+        outputSize(in_dims[3], filter_dims[3], paddings[1], strides[1]);
+    ctx->SetOutputDim(
+        "Output", {in_dims[0], filter_dims[0], output_height, output_width});
   }
 };
 
 class Conv2DOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  Conv2DOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+  Conv2DOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput(
         "Input",
@@ -108,14 +106,15 @@ class Conv2DOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    auto in = ctx.Input<Tensor>("Input");
-    auto filter = ctx.Input<Tensor>("Filter");
-    auto d_in = ctx.Output<framework::Tensor>(framework::GradVarName("Input"));
-    auto d_filter =
-        ctx.Output<framework::Tensor>(framework::GradVarName("Filter"));
-    if (d_in) d_in->Resize(in->dims());
-    if (d_filter) d_filter->Resize(filter->dims());
+  void InferShape(framework::InferShapeContextBase* ctx) const override {
+    auto in_dims = ctx->GetInputDim("Input");
+    auto filter_dims = ctx->GetInputDim("Filter");
+    if (ctx->HasOutput(framework::GradVarName("Input"))) {
+      ctx->SetOutputDim(framework::GradVarName("Input"), in_dims);
+    }
+    if (ctx->HasOutput(framework::GradVarName("Filter"))) {
+      ctx->SetOutputDim(framework::GradVarName("Filter"), filter_dims);
+    }
   }
 };
 

paddle/operators/cos_sim_op.cc

@@ -24,22 +24,22 @@ class CosSimOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
+  void InferShape(framework::InferShapeContextBase* ctx) const override {
     // notnull check
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input(X) of CosSimOp should not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
-                            "Input(Y) of CosSimOp should not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
-                            "Output(Out) of CosSimOp should not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("XNorm"),
-                            "Output(XNorm) of CosSimOp should not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("YNorm"),
-                            "Output(YNorm) of CosSimOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of CosSimOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Y"),
+                   "Input(Y) of CosSimOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of CosSimOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("XNorm"),
+                   "Output(XNorm) of CosSimOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("YNorm"),
+                   "Output(YNorm) of CosSimOp should not be null.");
 
     // shape check
-    auto x_dims = ctx.Input<Tensor>("X")->dims();
-    auto y_dims = ctx.Input<Tensor>("Y")->dims();
+    auto x_dims = ctx->GetInputDim("X");
+    auto y_dims = ctx->GetInputDim("Y");
 
     PADDLE_ENFORCE_EQ(x_dims.size(), y_dims.size(),
                       "Ranks of Input(X) and Input(Y) must be equal.");
@@ -54,16 +54,16 @@ class CosSimOp : public framework::OperatorWithKernel {
                    " just 1 (which will be broadcasted to match Input(X)).");
 
     // resize tensor
-    ctx.Output<framework::Tensor>("Out")->Resize({x_dims[0], 1});
-    ctx.Output<framework::Tensor>("XNorm")->Resize({x_dims[0], 1});
-    ctx.Output<framework::Tensor>("YNorm")->Resize({y_dims[0], 1});
-    ctx.ShareLoD("X", /*->*/ "Out");
+    ctx->SetOutputDim("Out", {x_dims[0], 1});
+    ctx->SetOutputDim("XNorm", {x_dims[0], 1});
+    ctx->SetOutputDim("YNorm", {y_dims[0], 1});
+    ctx->ShareLoD("X", /*->*/ "Out");
   }
 };
 
 class CosSimOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  CosSimOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+  CosSimOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The 1st input of cos_sim op.");
     AddInput("Y", "The 2nd input of cos_sim op.");
@@ -98,27 +98,23 @@ class CosSimOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
+  void InferShape(framework::InferShapeContextBase* ctx) const override {
     // notnull check
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) must not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("XNorm"),
-                            "Input(XNorm) must not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("YNorm"),
-                            "Input(YNorm) must not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Out"),
-                            "Input(Out) must not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
-                            "Input(Out@GRAD) must not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) must not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("XNorm"), "Input(XNorm) must not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("YNorm"), "Input(YNorm) must not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Out"), "Input(Out) must not be null.");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
+                   "Input(Out@GRAD) must not be null.");
 
     // shape check
-    auto x_dims = ctx.Input<Tensor>("X")->dims();
-    auto y_dims = ctx.Input<Tensor>("Y")->dims();
-    auto xnorm_dims = ctx.Input<Tensor>("XNorm")->dims();
-    auto ynorm_dims = ctx.Input<Tensor>("YNorm")->dims();
-    auto out_dims = ctx.Input<Tensor>("Out")->dims();
-    auto out_grad_dims =
-        ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
+    auto x_dims = ctx->GetInputDim("X");
+    auto y_dims = ctx->GetInputDim("Y");
+    auto xnorm_dims = ctx->GetInputDim("XNorm");
+    auto ynorm_dims = ctx->GetInputDim("YNorm");
+    auto out_dims = ctx->GetInputDim("Out");
+    auto out_grad_dims = ctx->GetInputDim(framework::GradVarName("Out"));
 
     PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
                       "Ranks of Input(X) and Input(Y) must be equal.");
@@ -143,10 +139,14 @@ class CosSimOpGrad : public framework::OperatorWithKernel {
                       "Shape of Input(Out@Grad) must be [X.Dim(0), 1].");
 
     // resize tensor
-    auto *x_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
-    auto *y_grad = ctx.Output<framework::Tensor>(framework::GradVarName("Y"));
-    if (x_grad) x_grad->Resize(x_dims);
-    if (y_grad) y_grad->Resize(y_dims);
+    auto x_grad_name = framework::GradVarName("X");
+    auto y_grad_name = framework::GradVarName("Y");
+    if (ctx->HasOutput(x_grad_name)) {
+      ctx->SetOutputDim(x_grad_name, x_dims);
+    }
+    if (ctx->HasOutput(y_grad_name)) {
+      ctx->SetOutputDim(y_grad_name, y_dims);
+    }
   }
 };
 

paddle/operators/crop_op.cc

@@ -25,16 +25,14 @@ class CropOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input(X) of CropOp should not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
-                            "Output(Out) of CropOp should not be null.");
-    auto x_dim = ctx.Input<Tensor>("X")->dims();
-    auto *y = ctx.Input<Tensor>("Y");
-    auto *out = ctx.Output<Tensor>("Out");
-    if (y == nullptr) {
-      auto shape = Attr<std::vector<int>>("shape");
+  void InferShape(framework::InferShapeContextBase* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of CropOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of CropOp should not be null.");
+    auto x_dim = ctx->GetInputDim("X");
+    if (!ctx->HasInput("Y")) {
+      auto shape = ctx->Attrs().Get<std::vector<int>>("shape");
       PADDLE_ENFORCE_EQ(
           int64_t(shape.size()), x_dim.size(),
           "Shape size should be equal to dimention size of input tensor.");
@@ -42,19 +40,20 @@ class CropOp : public framework::OperatorWithKernel {
       for (size_t i = 0; i < shape.size(); ++i) {
         tensor_shape[i] = static_cast<int64_t>(shape[i]);
       }
-      out->Resize(framework::make_ddim(tensor_shape));
+      ctx->SetOutputDim("Out", framework::make_ddim(tensor_shape));
     } else {
-      PADDLE_ENFORCE_EQ(framework::arity(x_dim), framework::arity(y->dims()),
+      auto y_dim = ctx->GetInputDim("Y");
+      PADDLE_ENFORCE_EQ(framework::arity(x_dim), framework::arity(y_dim),
                         "Tensor rank of both CropOp's "
                         "inputs must be same.");
-      out->Resize(y->dims());
+      ctx->SetOutputDim("Out", y_dim);
     }
   }
 };
 
 class CropOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  CropOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+  CropOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X",
              "The input of pad op. "
@@ -78,12 +77,12 @@ class CropOpMaker : public framework::OpProtoAndCheckerMaker {
 Crop Operator.
 Crop input into output, as specified by offsets and shape.
 
-There are two ways to set shape: 
+There are two ways to set shape:
 1. referenc input: crop input X as shape as reference input.
-                    The dimension of reference input should 
+                    The dimension of reference input should
                     be as same as input X.
 2. shape list: crop input X by shape described by a list<int>.
-               The size of shape list should be as same as 
+               The size of shape list should be as same as
                dimension size of  input X.
 
 The input should be a k-D tensor(k > 0 and k < 7). As an example:
@@ -94,15 +93,15 @@ Given:
          [0, 3, 4, 0, 0]
          [0, 0, 0, 0, 0]]
 
-and 
+and
 
     offsets = [0, 1]
 
 and
- 
+
     shape = [2, 2]
 
-then we get 
+then we get
 
     Out = [[1, 2],
            [3, 4]]
@@ -116,14 +115,14 @@ class CropOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
-                            "Input(Out@GRAD) should not be null");
-    auto x_dims = ctx.Input<Tensor>("X")->dims();
-    auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
-    if (x_grad != nullptr) {
-      x_grad->Resize(x_dims);
+  void InferShape(framework::InferShapeContextBase* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
+                   "Input(Out@GRAD) should not be null");
+    auto x_dims = ctx->GetInputDim("X");
+    auto x_grad_name = framework::GradVarName("X");
+    if (ctx->HasOutput(x_grad_name)) {
+      ctx->SetOutputDim(x_grad_name, x_dims);
     }
   }
 };

paddle/operators/cross_entropy_op.cc

@@ -22,33 +22,30 @@ class CrossEntropyOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should be not null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"),
-                            "Input(Label) should be not null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"),
-                            "Output(Y) should be not null.");
-
-    auto x = ctx.Input<Tensor>("X");
-    auto label = ctx.Input<Tensor>("Label");
-    PADDLE_ENFORCE_EQ(x->dims().size(), 2, "Input(X)'s rank should be 2.");
-    PADDLE_ENFORCE_EQ(label->dims().size(), 2,
-                      "Input(Label)'s rank should be 2.");
-    PADDLE_ENFORCE_EQ(x->dims()[0], label->dims()[0],
+  void InferShape(framework::InferShapeContextBase* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
+    PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Y"), "Output(Y) should be not null.");
+
+    auto x_dims = ctx->GetInputDim("X");
+    auto label_dims = ctx->GetInputDim("Label");
+    PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2.");
+    PADDLE_ENFORCE_EQ(label_dims.size(), 2, "Input(Label)'s rank should be 2.");
+    PADDLE_ENFORCE_EQ(x_dims[0], label_dims[0],
                       "The 1st dimension of Input(X) and Input(Label) should "
                       "be equal.");
-    if (ctx.Attr<bool>("softLabel")) {
-      PADDLE_ENFORCE_EQ(x->dims()[1], label->dims()[1],
+    if (ctx->Attrs().Get<bool>("softLabel")) {
+      PADDLE_ENFORCE_EQ(x_dims[1], label_dims[1],
                         "If Attr(softLabel) == true, the 2nd dimension of "
                         "Input(X) and Input(Label) should be equal.");
     } else {
-      PADDLE_ENFORCE_EQ(label->dims()[1], 1,
+      PADDLE_ENFORCE_EQ(label_dims[1], 1,
                         "If Attr(softLabel) == false, the 2nd dimension of "
                         "Input(Label) should be 1.");
     }
 
-    ctx.Output<Tensor>("Y")->Resize({x->dims()[0], 1});
-    ctx.ShareLoD("X", /*->*/ "Y");
+    ctx->SetOutputDim("Y", {x_dims[0], 1});
+    ctx->ShareLoD("X", /*->*/ "Y");
   }
 };
 
@@ -57,50 +54,45 @@ class CrossEntropyGradientOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should be not null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Label"),
-                            "Input(Label) should be not null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Y")),
-                            "Input(Y@GRAD) shoudl be not null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar(framework::GradVarName("X")),
-                            "Output(X@GRAD) should be not null.");
-
-    auto x = ctx.Input<Tensor>("X");
-    auto label = ctx.Input<Tensor>("Label");
-    auto dy = ctx.Input<Tensor>(framework::GradVarName("Y"));
-    PADDLE_ENFORCE_EQ(x->dims().size(), 2, "Input(X)'s rank should be 2.");
-    PADDLE_ENFORCE_EQ(dy->dims().size(), 2,
-                      "Input(Y@Grad)'s rank should be 2.");
-    PADDLE_ENFORCE_EQ(label->dims().size(), 2,
-                      "Input(Label)'s rank should be 2.");
-    PADDLE_ENFORCE_EQ(x->dims()[0], label->dims()[0],
+  void InferShape(framework::InferShapeContextBase* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
+    PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should be not null.");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")),
+                   "Input(Y@GRAD) shoudl be not null.");
+    PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")),
+                   "Output(X@GRAD) should be not null.");
+
+    auto x_dims = ctx->GetInputDim("X");
+    auto label_dims = ctx->GetInputDim("Label");
+    auto dy_dims = ctx->GetInputDim(framework::GradVarName("Y"));
+    PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2.");
+    PADDLE_ENFORCE_EQ(dy_dims.size(), 2, "Input(Y@Grad)'s rank should be 2.");
+    PADDLE_ENFORCE_EQ(label_dims.size(), 2, "Input(Label)'s rank should be 2.");
+    PADDLE_ENFORCE_EQ(x_dims[0], label_dims[0],
                       "The 1st dimension of Input(X) and Input(Label) should "
                       "be equal.");
-    PADDLE_ENFORCE_EQ(x->dims()[0], dy->dims()[0],
+    PADDLE_ENFORCE_EQ(x_dims[0], dy_dims[0],
                       "The 1st dimension of Input(X) and Input(Y@Grad) should "
                       "be equal.");
-    PADDLE_ENFORCE_EQ(dy->dims()[1], 1,
+    PADDLE_ENFORCE_EQ(dy_dims[1], 1,
                       "The 2nd dimension of Input(Y@Grad) should be 1.");
-    if (ctx.Attr<bool>("softLabel")) {
-      PADDLE_ENFORCE_EQ(x->dims()[1], label->dims()[1],
+    if (ctx->Attrs().Get<bool>("softLabel")) {
+      PADDLE_ENFORCE_EQ(x_dims[1], label_dims[1],
                         "When Attr(softLabel) == true, the 2nd dimension of "
                         "Input(X) and Input(Label) should be equal.");
     } else {
-      PADDLE_ENFORCE_EQ(label->dims()[1], 1,
+      PADDLE_ENFORCE_EQ(label_dims[1], 1,
                         "When Attr(softLabel) == false, the 2nd dimension of "
                         "Input(Label) should be 1.");
     }
-
-    auto dx = ctx.Output<Tensor>(framework::GradVarName("X"));
-    dx->Resize(x->dims());
+    ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
   }
 };
 
 class CrossEntropyOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  CrossEntropyOpMaker(framework::OpProto *proto,
-                      framework::OpAttrChecker *op_checker)
+  CrossEntropyOpMaker(framework::OpProto* proto,
+                      framework::OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X",
              "(Tensor, default Tensor<float>), a 2-D tensor with shape N x D, "

paddle/operators/dropout_op.cc

@@ -24,25 +24,25 @@ class DropoutOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
-    PADDLE_ENFORCE_GE(ctx.Attr<float>("dropout_prob"), 0);
-    PADDLE_ENFORCE_LE(ctx.Attr<float>("dropout_prob"), 1);
-
-    auto dims = ctx.Input<Tensor>("X")->dims();
-    ctx.Output<Tensor>("Out")->Resize(dims);
-    if (ctx.Attr<bool>("is_training")) {
-      ctx.Output<Tensor>("Mask")->Resize(dims);
+  void InferShape(framework::InferShapeContextBase* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null.");
+    PADDLE_ENFORCE_GE(ctx->Attrs().Get<float>("dropout_prob"), 0);
+    PADDLE_ENFORCE_LE(ctx->Attrs().Get<float>("dropout_prob"), 1);
+
+    auto x_dims = ctx->GetInputDim("X");
+    ctx->SetOutputDim("Out", x_dims);
+    if (ctx->Attrs().Get<bool>("is_training") == 1) {
+      ctx->SetOutputDim("Mask", x_dims);
     }
-    ctx.ShareLoD("X", /*->*/ "Out");
+    ctx->ShareLoD("X", /*->*/ "Out");
   }
 };
 
 template <typename AttrType>
 class DropoutOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  DropoutOpMaker(framework::OpProto *proto,
-                 framework::OpAttrChecker *op_checker)
+  DropoutOpMaker(framework::OpProto* proto,
+                 framework::OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddAttr<AttrType>("dropout_prob", "Probability of setting units to zero.")
         .SetDefault(.5f);
@@ -70,27 +70,26 @@ class DropoutOpGrad : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE(ctx.Attr<bool>("is_training"),
-                   "GradOp is only callable when is_training is true");
-
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) must not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Mask"), "Mask must not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
-                            "Input(Out@GRAD) must not be null.");
-
-    PADDLE_ENFORCE_GE(ctx.Attr<AttrType>("dropout_prob"), 0);
-    PADDLE_ENFORCE_LE(ctx.Attr<AttrType>("dropout_prob"), 1);
-    auto x_dims = ctx.Input<Tensor>("X")->dims();
-    auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
+  void InferShape(framework::InferShapeContextBase* ctx) const override {
+    PADDLE_ENFORCE_EQ(ctx->Attrs().Get<bool>("is_training"), 1,
+                      "GradOp is only callable when is_training is true");
+
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Mask"), "Mask must not be null.");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
+                   "Input(Out@GRAD) must not be null.");
+
+    PADDLE_ENFORCE_GE(ctx->Attrs().Get<AttrType>("dropout_prob"), 0);
+    PADDLE_ENFORCE_LE(ctx->Attrs().Get<AttrType>("dropout_prob"), 1);
+    auto x_dims = ctx->GetInputDim("X");
+    auto out_dims = ctx->GetInputDim(framework::GradVarName("Out"));
     PADDLE_ENFORCE_EQ(x_dims, out_dims,
                       "Dimensions of Input(X) and Out@Grad must be the same.");
-    auto mask_dims = ctx.Input<Tensor>("Mask")->dims();
+    auto mask_dims = ctx->GetInputDim("Mask");
     PADDLE_ENFORCE_EQ(x_dims, mask_dims,
                       "Dimensions of Input(X) and Mask must be the same.");
 
-    auto *x_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
-    x_grad->Resize(x_dims);
+    ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
   }
 };
 

paddle/operators/elementwise_op.h

@@ -202,21 +202,20 @@ class ElementwiseOp : public framework::OperatorWithKernel {
 
  protected:
   using Tensor = framework::Tensor;
-  void InferShape(const framework::InferShapeContext& ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input(X) of elementwise op should not be null");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"),
-                            "Input(Y) of elementwise op should not be null");
-    PADDLE_ENFORCE_NOT_NULL(
-        ctx.OutputVar("Out"),
-        "Output(Out) of elementwise op should not be null.");
-
-    auto x_dim = ctx.Input<Tensor>("X")->dims();
-    auto y_dim = ctx.Input<Tensor>("Y")->dims();
+  void InferShape(framework::InferShapeContextBase* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of elementwise op should not be null");
+    PADDLE_ENFORCE(ctx->HasInput("Y"),
+                   "Input(Y) of elementwise op should not be null");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of elementwise op should not be null.");
+
+    auto x_dim = ctx->GetInputDim("X");
+    auto y_dim = ctx->GetInputDim("Y");
     PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(),
                       "Rank of first input must >= rank of second input.")
-    ctx.Output<framework::Tensor>("Out")->Resize(x_dim);
-    ctx.ShareLoD("X", /*->*/ "Out");
+    ctx->SetOutputDim("Out", x_dim);
+    ctx->ShareLoD("X", /*->*/ "Out");
   }
 };
 
@@ -234,7 +233,7 @@ must be small or equal to X's dimensions.
 )DOC");
     AddAttr<int>("axis",
                  R"DOC(
-When the shape(Y) does not equal the shape(X),Y will be broadcasted 
+When the shape(Y) does not equal the shape(X),Y will be broadcasted
 to match the shape of X and axis should be dimension index Y in X
         )DOC")
         .SetDefault(-1)
@@ -244,7 +243,7 @@ to match the shape of X and axis should be dimension index Y in X
     comment_ = R"DOC(
 Limited elementwise {name} operator.The equation is: Out = {equation}.
 1. The shape of Y should be same with X or
-2. Y's shape is a subset of X. 
+2. Y's shape is a subset of X.
    Y will be broadcasted to match the shape of X and axis should be dimension index Y in X.
 
    example:
@@ -284,27 +283,26 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel {
   using Tensor = framework::Tensor;
 
  protected:
-  void InferShape(const framework::InferShapeContext& ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"), "Input(X) should not be null");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Y"), "Input(Y) should not be null");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar(framework::GradVarName("Out")),
-                            "Input(Out@GRAD) should not be null");
-
-    auto x_dims = ctx.Input<Tensor>("X")->dims();
-    auto y_dims = ctx.Input<Tensor>("Y")->dims();
-    auto out_dims = ctx.Input<Tensor>(framework::GradVarName("Out"))->dims();
-    auto* x_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
-    auto* y_grad = ctx.Output<framework::Tensor>(framework::GradVarName("Y"));
+  void InferShape(framework::InferShapeContextBase* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null");
+    PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should not be null");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
+                   "Input(Out@GRAD) should not be null");
+
+    auto x_dims = ctx->GetInputDim("X");
+    auto y_dims = ctx->GetInputDim("Y");
+    auto out_dims = ctx->GetInputDim(framework::GradVarName("Out"));
 
     PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
                       "Rank of first input must >= rank of second input.")
 
-    if (x_grad) {
-      x_grad->Resize(x_dims);
+    auto x_grad_name = framework::GradVarName("X");
+    auto y_grad_name = framework::GradVarName("Y");
+    if (ctx->HasOutput(x_grad_name)) {
+      ctx->SetOutputDim(x_grad_name, x_dims);
     }
-
-    if (y_grad) {
-      y_grad->Resize(y_dims);
+    if (ctx->HasOutput(y_grad_name)) {
+      ctx->SetOutputDim(y_grad_name, y_dims);
     }
   }
 };

paddle/operators/fill_zeros_like_op.cc

@@ -22,15 +22,13 @@ class FillZerosLikeOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input(X) of FillZerosLikeOp should not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Y"),
-                            "Output(Y) of FillZerosLikeOp should not be null.");
-
-    ctx.Output<framework::Tensor>("Y")->Resize(
-        ctx.Input<framework::Tensor>("X")->dims());
-    ctx.ShareLoD("X", /*->*/ "Y");
+  void InferShape(framework::InferShapeContextBase *ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of FillZerosLikeOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Y"),
+                   "Output(Y) of FillZerosLikeOp should not be null.");
+    ctx->SetOutputDim("Y", ctx->GetInputDim("X"));
+    ctx->ShareLoD("X", /*->*/ "Y");
   }
 };
 

paddle/operators/gather_op.cc

@@ -23,19 +23,19 @@ class GatherOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("X"),
-                            "Input(X) of GatherOp should not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.InputVar("Index"),
-                            "Input(Index) of GatherOp should not be null.");
-    PADDLE_ENFORCE_NOT_NULL(ctx.OutputVar("Out"),
-                            "Output(Out) of GatherOp should not be null.");
+  void InferShape(framework::InferShapeContextBase* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"),
+                   "Input(X) of GatherOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Index"),
+                   "Input(Index) of GatherOp should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of GatherOp should not be null.");
 
-    int batch_size = ctx.Input<Tensor>("Index")->dims()[0];
+    int batch_size = ctx->GetInputDim("Index")[0];
     PADDLE_ENFORCE_GE(batch_size, 0, "Batch size must be >0");
-    framework::DDim output_dims(ctx.Input<Tensor>("X")->dims());
+    framework::DDim output_dims(ctx->GetInputDim("X"));
     output_dims[0] = batch_size;
-    ctx.Output<framework::Tensor>("Out")->Resize(output_dims);
+    ctx->SetOutputDim("Out", output_dims);
   }
 };
 
@@ -44,23 +44,20 @@ class GatherGradOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext &ctx) const override {
-    auto X_grad = ctx.Output<framework::Tensor>(framework::GradVarName("X"));
-    auto X = ctx.Input<Tensor>("X");
-
-    X_grad->Resize(X->dims());
+  void InferShape(framework::InferShapeContextBase* ctx) const override {
+    ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
   }
 };
 
 class GatherOpMaker : public framework::OpProtoAndCheckerMaker {
  public:
-  GatherOpMaker(framework::OpProto *proto, framework::OpAttrChecker *op_checker)
+  GatherOpMaker(framework::OpProto* proto, framework::OpAttrChecker* op_checker)
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The source input of gather op");
     AddInput("Index", "The index input of gather op");
     AddOutput("Out", "The output of add op");
     AddComment(R"DOC(
-Gather Operator by selecting from the first axis, 
+Gather Operator by selecting from the first axis,
 
 Out = X[Index]
 )DOC");

paddle/operators/gaussian_random_op.cc

@@ -43,13 +43,10 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
   using framework::OperatorWithKernel::OperatorWithKernel;
 
  protected:
-  void InferShape(const framework::InferShapeContext& ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(
-        ctx.OutputVar("Out"),
-        "Output(Out) of GaussianRandomOp should not be null.");
-
-    auto* tensor = ctx.Output<framework::Tensor>("Out");
-    auto dims = Attr<std::vector<int>>("dims");
+  void InferShape(framework::InferShapeContextBase* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasOutput("Out"),
+                   "Output(Out) of GaussianRandomOp should not be null.");
+    auto dims = ctx->Attrs().Get<std::vector<int>>("dims");
     std::vector<int64_t> temp;
     temp.reserve(dims.size());
     for (auto dim : dims) {
@@ -57,7 +54,7 @@ class GaussianRandomOp : public framework::OperatorWithKernel {
     }
     PADDLE_ENFORCE(dims.size() > 0UL,
                    "dims can be one int or array. dims must be set.");
-    tensor->Resize(framework::make_ddim(temp));
+    ctx->SetOutputDim("Out", framework::make_ddim(temp));
   }
 };