PaddlePaddle · luotao1 · Aug 31, 2017 · Aug 31, 2017 · Aug 31, 2017 · Aug 31, 2017
diff --git a/doc/api/v2/config/layer.rst b/doc/api/v2/config/layer.rst
@@ -434,9 +434,9 @@ lambda_cost
 ..  autoclass:: paddle.v2.layer.lambda_cost
     :noindex:
 
-mse_cost
+square_error_cost
 --------
-..  autoclass:: paddle.v2.layer.mse_cost
+..  autoclass:: paddle.v2.layer.square_error_cost
     :noindex:
 
 rank_cost

diff --git a/doc/design/graph.md b/doc/design/graph.md
@@ -0,0 +1,51 @@
+# Design Doc: Computations as Graphs
+
+A primary goal of the refactorization of PaddlePaddle is a more flexible representation of deep learning computation, in particular, a graph of operators and variables, instead of sequences of layers as before.
+
+This document explains that the construction of a graph as three steps:
+
+- construct the forward part
+- construct the backward part
+- construct the optimization part
+
+Let us take the problem of image classification as a simple example.  The application program that trains the model looks like:
+
+```python
+x = layer.data("images")
+l = layer.data("label")
+y = layer.fc(x)
+cost = layer.mse(y, l)
+optimize(cost)
+train(cost, reader=mnist.train())
+```
+
+### Forward Part
+
+The first four lines of above program build the forward part of the graph.
+
+![](images/graph_construction_example_forward_only.png)
+
+In particular, the first line `x = layer.data("images")` creates variable x and a Feed operator that copies a column from the minibatch to x.  `y = layer.fc(x)` creates not only the FC operator and output variable y, but also two parameters, W and b.
+
+In this example, all operators are created as `OpDesc` protobuf messages, and all variables are `VarDesc`.  These protobuf messages are saved in a `BlockDesc` protobuf message.
+
+### Backward Part
+
+The fifth line `optimize(cost)` calls two functions, `ConstructBackwardGraph` and `ConstructOptimizationGraph`.
+
+`ConstructBackwardGraph` traverses the forward graph in the `BlockDesc` protobuf message and builds the backward part.
+
+![](images/graph_construction_example_forward_backward.png)
+
+According to the chain rule of gradient computation, `ConstructBackwardGraph` would
+
+1. create a gradient operator G for each operator F,
+1. make all inputs, outputs, and outputs' gradient of F as inputs of G,
+1. create gradients for all inputs of F, except for those who don't have gradients, like x and l, and
+1. make all these gradients as outputs of G.
+
+### Optimization Part
+
+For each parameter, like W and b created by `layer.fc`, marked as double circles in above graphs, `ConstructOptimizationGraph` creates an optimization operator to apply its gradient.  Here results in the complete graph:
+
+![](images/graph_construction_example_all.png)
diff --git a/doc/design/images/graph_construction_example.bash b/doc/design/images/graph_construction_example.bash
@@ -0,0 +1,11 @@
+cat ./graph_construction_example.dot | \
+    sed 's/color=red/color=red, style=invis/g' | \
+    sed 's/color=green/color=green, style=invis/g' | \
+    dot -Tpng > graph_construction_example_forward_only.png
+
+cat ./graph_construction_example.dot | \
+    sed 's/color=green/color=green, style=invis/g' | \
+    dot -Tpng > graph_construction_example_forward_backward.png
+
+cat ./graph_construction_example.dot | \
+    dot -Tpng > graph_construction_example_all.png
diff --git a/doc/design/images/graph_construction_example.dot b/doc/design/images/graph_construction_example.dot
@@ -0,0 +1,65 @@
+digraph ImageClassificationGraph {
+        ///////// The forward part /////////
+        FeedX [label="Feed", color=blue, shape=box];
+        FeedY [label="Feed", color=blue, shape=box];
+        FC [label="FC", color=blue, shape=box];
+        MSE [label="MSE", color=blue, shape=box];
+
+        x [label="x", color=blue, shape=oval];
+        l [label="l", color=blue, shape=oval];
+        y [label="y", color=blue, shape=oval];
+        W [label="W", color=blue, shape=doublecircle];
+        b [label="b", color=blue, shape=doublecircle];
+        cost [label="cost", color=blue, shape=oval];
+
+        FeedX -> x -> FC -> y -> MSE -> cost [color=blue];
+        FeedY -> l [color=blue];
+        W -> FC [color=blue];
+        b -> FC [color=blue];
+        l -> MSE [color=blue];
+
+        ////////// The backward part /////////
+        MSE_Grad [label="MSE_grad", color=red, shape=box];
+        FC_Grad [label="FC_grad", color=red, shape=box];
+
+        d_cost [label="d cost", color=red, shape=oval];
+        d_y [label="d y", color=red, shape=oval];
+        d_b [label="d b", color=red, shape=oval];
+        d_W [label="d W", color=red, shape=oval];
+
+        cost -> MSE_Grad [color=red];
+        d_cost -> MSE_Grad [color=red];
+        x -> MSE_Grad [color=red];
+        l -> MSE_Grad [color=red];
+        y -> MSE_Grad -> d_y [color=red];
+
+        x -> FC_Grad [color=red];
+        y -> FC_Grad [color=red];
+        d_y -> FC_Grad [color=red];
+        W -> FC_Grad -> d_W [color=red];
+        b -> FC_Grad -> d_b [color=red];
+
+        ////////// The optimizaiton part //////////
+
+        OPT_W [label="SGD", color=green, shape=box];
+        OPT_b [label="SGD", color=green, shape=box];
+
+        W -> OPT_W [color=green];
+        b -> OPT_b [color=green];
+        d_W -> OPT_W -> W [color=green];
+        d_b -> OPT_b -> b [color=green];
+
+        ////////// Groupings //////////
+
+        subgraph clusterMSE {
+                style=invis;
+                MSE;
+                MSE_Grad;
+        }
+
+        subgraph clusterFC {
+                style=invis;
+                FC;
+                FC_Grad;
+        }
+}
diff --git a/doc/design/images/graph_construction_example_all.png b/doc/design/images/graph_construction_example_all.png
diff --git a/doc/design/images/graph_construction_example_forward_backward.png b/doc/design/images/graph_construction_example_forward_backward.png
diff --git a/doc/design/images/graph_construction_example_forward_only.png b/doc/design/images/graph_construction_example_forward_only.png
diff --git a/doc/getstarted/basic_usage/index_cn.rst b/doc/getstarted/basic_usage/index_cn.rst
@@ -55,7 +55,7 @@ PaddlePaddle是源于百度的一个深度学习平台。这份简短的介绍
     # 线性计算网络层: ȳ = wx + b
     ȳ = fc_layer(input=x, param_attr=ParamAttr(name='w'), size=1, act=LinearActivation(), bias_attr=ParamAttr(name='b'))
     # 计算误差函数，即  ȳ 和真实 y 之间的距离
-    cost = mse_cost(input= ȳ, label=y)
+    cost = square_error_cost(input= ȳ, label=y)
     outputs(cost)
 
 
@@ -69,7 +69,7 @@ PaddlePaddle是源于百度的一个深度学习平台。这份简短的介绍
 
     - **数据层**：数据层 `data_layer` 是神经网络的入口，它读入数据并将它们传输到接下来的网络层。这里数据层有两个，分别对应于变量 `x` 和 `y`。
     - **全连接层**：全连接层 `fc_layer` 是基础的计算单元，这里利用它建模变量之间的线性关系。计算单元是神经网络的核心，PaddlePaddle支持大量的计算单元和任意深度的网络连接，从而可以拟合任意的函数来学习复杂的数据关系。
-    - **回归误差代价层**：回归误差代价层 `mse_cost` 是众多误差代价函数层的一种，它们在训练过程作为网络的出口，用来计算模型的误差，是模型参数优化的目标函数。
+    - **回归误差代价层**：回归误差代价层 `square_error_cost` 是众多误差代价函数层的一种，它们在训练过程作为网络的出口，用来计算模型的误差，是模型参数优化的目标函数。
 
 定义了网络结构并保存为 `trainer_config.py` 之后，运行以下训练命令：
 

diff --git a/doc/getstarted/basic_usage/index_en.rst b/doc/getstarted/basic_usage/index_en.rst
@@ -49,7 +49,7 @@ To recover this relationship between ``X`` and ``Y``, we use a neural network wi
         x = data_layer(name='x', size=1)
         y = data_layer(name='y', size=1)
         y_predict = fc_layer(input=x, param_attr=ParamAttr(name='w'), size=1, act=LinearActivation(), bias_attr=ParamAttr(name='b'))
-        cost = mse_cost(input=y_predict, label=y)
+        cost = square_error_cost(input=y_predict, label=y)
         outputs(cost)
 
 Some of the most fundamental usages of PaddlePaddle are demonstrated:

diff --git a/doc/getstarted/concepts/src/train.py b/doc/getstarted/concepts/src/train.py
@@ -8,7 +8,7 @@
 x = paddle.layer.data(name='x', type=paddle.data_type.dense_vector(2))
 y_predict = paddle.layer.fc(input=x, size=1, act=paddle.activation.Linear())
 y = paddle.layer.data(name='y', type=paddle.data_type.dense_vector(1))
-cost = paddle.layer.mse_cost(input=y_predict, label=y)
+cost = paddle.layer.square_error_cost(input=y_predict, label=y)
 
 # create parameters
 parameters = paddle.parameters.create(cost)

diff --git a/doc/getstarted/concepts/use_concepts_cn.rst b/doc/getstarted/concepts/use_concepts_cn.rst
@@ -81,9 +81,9 @@ PaddlePaddle支持不同类型的输入数据，主要包括四种类型，和
 ..	code-block:: bash
 
     y_predict = paddle.layer.fc(input=x, size=1, act=paddle.activation.Linear())
-    cost = paddle.layer.mse_cost(input=y_predict, label=y)
+    cost = paddle.layer.square_error_cost(input=y_predict, label=y)
 
-其中，x与y为之前描述的输入层；而y_predict是接收x作为输入，接上一个全连接层；cost接收y_predict与y作为输入，接上均方误差层。
+其中，x与y为之前描述的输入层；而y_predict是接收x作为输入，接上一个全连接层；cost接收y_predict与y作为输入，接上平方误差层。
 
 最后一层cost中记录了神经网络的所有拓扑结构，通过组合不同的layer，我们即可完成神经网络的搭建。
 
@@ -147,4 +147,4 @@ PaddlePaddle支持不同类型的输入数据，主要包括四种类型，和
 ..  literalinclude:: src/train.py
     :linenos:
 
-有关线性回归的实际应用，可以参考PaddlePaddle book的 `第一章节 <http://book.paddlepaddle.org/index.html>`_。
+有关线性回归的实际应用，可以参考PaddlePaddle book的 `第一章节 <http://book.paddlepaddle.org/index.html>`_。
diff --git a/doc/howto/dev/new_op_cn.md b/doc/howto/dev/new_op_cn.md
@@ -178,13 +178,13 @@ class MulKernel : public framework::OpKernel {
 
 ```c++
 namespace ops = paddle::operators;
-REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, ops::MulOpGrad);
+REGISTER_OP(mul, ops::MulOp, ops::MulOpMaker, mul_grad, ops::MulOpGrad);
 REGISTER_OP_CPU_KERNEL(mul, ops::MulKernel<paddle::platform::CPUPlace, float>);
 REGISTER_OP_CPU_KERNEL(mul_grad,
               ops::MulGradKernel<paddle::platform::CPUPlace, float>);
 ```
 
-  - `REGISTER_OP` ： 注册`ops::MulOp`类，类型名为`mul`，该类的`ProtoMaker`为`ops::MulOpMaker`，并且注册`ops::MulOpGrad`为其反向Op。
+  - `REGISTER_OP` ： 注册`ops::MulOp`类，类型名为`mul`，该类的`ProtoMaker`为`ops::MulOpMaker`，注册`ops::MulOpGrad`，类型名为`mul_grad`，
   - `REGISTER_OP_WITHOUT_GRADIENT` ： 用于注册没有反向的Op。
   - `REGISTER_OP_CPU_KERNEL` ：注册`ops::MulKernel`类，并特化模板参数为`paddle::platform::CPUPlace`和`float`类型，同理，注册`ops::MulKernel`类。
 
@@ -227,6 +227,12 @@ make mul_op
     USE_CPU_ONLY_OP(gather);
     ```
 
+    如果OP不带Kernel，则使用`USE_NO_KENREL_OP`:
+
+    ```
+    USE_NO_KENREL_OP(recurrent);
+    ```
+
     使用`USE_OP`告知编译器需要链接该Op的目标文件，具体解释参考[代码注释](/~https://github.com/PaddlePaddle/Paddle/blob/develop/paddle/framework/op_registry.h#L81)。
 
 

diff --git a/doc/howto/usage/k8s/k8s_distributed_cn.md b/doc/howto/usage/k8s/k8s_distributed_cn.md
@@ -213,7 +213,7 @@ I1116 09:10:17.123440    50 Util.cpp:130] Calling runInitFunctions
 I1116 09:10:17.123764    50 Util.cpp:143] Call runInitFunctions done.
 [WARNING 2016-11-16 09:10:17,227 default_decorators.py:40] please use keyword arguments in paddle config.
 [INFO 2016-11-16 09:10:17,239 networks.py:1282] The input order is [movie_id, title, genres, user_id, gender, age, occupation, rating]
-[INFO 2016-11-16 09:10:17,239 networks.py:1289] The output order is [__mse_cost_0__]
+[INFO 2016-11-16 09:10:17,239 networks.py:1289] The output order is [__square_error_cost_0__]
 I1116 09:10:17.392917    50 Trainer.cpp:170] trainer mode: Normal
 I1116 09:10:17.613910    50 PyDataProvider2.cpp:257] loading dataprovider dataprovider::process
 I1116 09:10:17.680917    50 PyDataProvider2.cpp:257] loading dataprovider dataprovider::process

diff --git a/paddle/framework/backward.cc b/paddle/framework/backward.cc
@@ -182,7 +182,7 @@ static std::unique_ptr<OperatorBase> BackwardRecursive(
                    });
 
     // process recurrent gradient op as a special operator.
-    if (forwardOp.Type() == "recurrent_op") {
+    if (forwardOp.Type() == "recurrent") {
       // NOTE clean up cycle call somewhere (RNN's stepnet constains itself), or
       // this will result in infinite loop.
       const auto& rnnop =

diff --git a/paddle/framework/backward.md b/paddle/framework/backward.md
@@ -18,7 +18,7 @@ A backward network is built up with several backward operators. Backward operato
 For example, we have got a `mul_op`, and we can register it's information and corresponding backward operator by the following macro:
 
 ```cpp
-REGISTER_OP(mul, MulOp, MulOpMaker, MulOpGrad);
+REGISTER_OP(mul, MulOp, MulOpMaker, mul_grad, MulOpGrad);
 ```
 
 `mul` is the operator's type. `MulOp` and `MulOpMaker` are the operator class and the operator maker class respectively.

diff --git a/paddle/framework/backward_test.cc b/paddle/framework/backward_test.cc
@@ -148,14 +148,16 @@ class AddOpMaker : public OpProtoAndCheckerMaker {
 namespace f = paddle::framework;
 namespace ops = paddle::operators;
 using EnforceNotMet = paddle::platform::EnforceNotMet;
-REGISTER_OP(rowwise_add, f::NOP, f::RowWiseAddOpMaker, f::NOP);
-REGISTER_OP(mul, f::NOP, f::MulOpMaker, f::NOP);
-REGISTER_OP(sigmoid, f::NOP, f::SigmoidOpMaker, f::NOP);
+REGISTER_OP(rowwise_add, f::NOP, f::RowWiseAddOpMaker, rowwise_add_grad,
+            f::NOP);
+REGISTER_OP(mul, f::NOP, f::MulOpMaker, mul_grad, f::NOP);
+REGISTER_OP(sigmoid, f::NOP, f::SigmoidOpMaker, sigmoid_grad, f::NOP);
 REGISTER_OP_WITHOUT_GRADIENT(nograd, f::NOP, f::NoGradOpMaker);
 REGISTER_OP_WITHOUT_GRADIENT(fill_zeros_like, f::NOP, f::FillZeroOpMaker);
-REGISTER_OP(add, f::NOP, f::AddOpMaker, f::NOP);
+REGISTER_OP(add, f::NOP, f::AddOpMaker, add_grad, f::NOP);
 REGISTER_OP_WITHOUT_GRADIENT(fc, f::FcOp, f::FcOpMaker);
-REGISTER_OP(many_output_op, f::NOP, f::ManyOutputOpMaker, f::NOP);
+REGISTER_OP(many_output_op, f::NOP, f::ManyOutputOpMaker, many_output_op_grad,
+            f::NOP);
 
 TEST(Backward, simple_op_grad) {
   auto fwd = f::OpRegistry::CreateOp(

diff --git a/paddle/framework/grad_op_builder_test.cc b/paddle/framework/grad_op_builder_test.cc
@@ -3,7 +3,7 @@
 #include "paddle/framework/op_registry.h"
 #include "paddle/framework/operator.h"
 
-USE_OP(add_two);
+USE_OP(add);
 
 namespace paddle {
 namespace framework {
@@ -41,7 +41,7 @@ namespace f = paddle::framework;
 
 TEST(GradOpBuilder, AddTwo) {
   std::shared_ptr<f::OperatorBase> add_op(f::OpRegistry::CreateOp(
-      "add_two", {{"X", {"x"}}, {"Y", {"y"}}}, {{"Out", {"out"}}}, {}));
+      "add", {{"X", {"x"}}, {"Y", {"y"}}}, {{"Out", {"out"}}}, {}));
   std::shared_ptr<f::OperatorBase> grad_add_op =
       f::OpRegistry::CreateGradOp(*add_op);
   EXPECT_EQ(grad_add_op->Inputs().size(), 4UL);
@@ -54,8 +54,8 @@ TEST(GradOpBuilder, AddTwo) {
   EXPECT_EQ(grad_add_op->Output(f::GradVarName("Y")), f::GradVarName("y"));
 }
 
-REGISTER_OP(mult_io, f::NOP, f::MutiInOutOpMaker, f::NOP);
-REGISTER_OP(io_ignored, f::NOP, f::IOIgnoredOpMaker, f::NOP);
+REGISTER_OP(mult_io, f::NOP, f::MutiInOutOpMaker, mult_io_grad, f::NOP);
+REGISTER_OP(io_ignored, f::NOP, f::IOIgnoredOpMaker, io_ignored_grad, f::NOP);
 
 TEST(GradOpBuilder, MutiInOut) {
   std::shared_ptr<f::OperatorBase> test_op(f::OpRegistry::CreateOp(

diff --git a/paddle/framework/op_registry.h b/paddle/framework/op_registry.h
@@ -33,7 +33,8 @@ namespace framework {
 class OpRegistry {
  public:
   template <typename OpType, typename ProtoMakerType, typename GradOpType>
-  static void RegisterOp(const std::string& op_type) {
+  static void RegisterOp(const std::string& op_type,
+                         const std::string& grad_op_type) {
     PADDLE_ENFORCE(!OpInfoMap::Instance().Has(op_type),
                    "'%s' is registered more than once.", op_type);
     OpInfo op_info;
@@ -42,9 +43,9 @@ class OpRegistry {
         const VariableNameMap& outputs, const AttributeMap& attrs) {
       return new OpType(type, inputs, outputs, attrs);
     };
+    op_info.grad_op_type_ = grad_op_type;
     if (std::type_index(typeid(ProtoMakerType)) !=
         std::type_index(typeid(NOPMaker))) {
-      op_info.grad_op_type_ = op_type + "_grad";
       op_info.proto_ = new OpProto;
       op_info.checker_ = new OpAttrChecker;
       auto maker = ProtoMakerType(op_info.proto_, op_info.checker_);
@@ -54,14 +55,15 @@ class OpRegistry {
           op_info.proto_->IsInitialized(),
           "Fail to initialize %s's OpProto, because %s is not initialized",
           op_type, op_info.proto_->InitializationErrorString());
-      // register gradient op
-      RegisterOp<GradOpType, NOPMaker, NOP>(op_info.grad_op_type_);
     } else {
-      op_info.grad_op_type_ = "";
       op_info.proto_ = nullptr;
       op_info.checker_ = nullptr;
     }
     OpInfoMap::Instance().Insert(op_type, op_info);
+    // register gradient op
+    if (!grad_op_type.empty()) {
+      RegisterOp<GradOpType, NOPMaker, NOP>(grad_op_type, "");
+    }
   }
 
   static std::unique_ptr<OperatorBase> CreateOp(const std::string& type,
@@ -90,8 +92,10 @@ class Registrar {
 template <typename OpType, typename ProtoMakerType, typename GradOpType>
 class OpRegistrar : public Registrar {
  public:
-  explicit OpRegistrar(const char* op_type) {
-    OpRegistry::RegisterOp<OpType, ProtoMakerType, GradOpType>(op_type);
+  explicit OpRegistrar(const char* op_type) { OpRegistrar(op_type, ""); }
+  OpRegistrar(const char* op_type, const char* grad_op_type) {
+    OpRegistry::RegisterOp<OpType, ProtoMakerType, GradOpType>(op_type,
+                                                               grad_op_type);
   }
 };
 
@@ -117,7 +121,8 @@ class OpKernelRegistrar : public Registrar {
 /**
  * Macro to register Operator.
  */
-#define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_class)         \
+#define REGISTER_OP(op_type, op_class, op_maker_class, grad_op_type,          \
+                    grad_op_class)                                            \
   STATIC_ASSERT_GLOBAL_NAMESPACE(                                             \
       __reg_op__##op_type, "REGISTER_OP must be called in global namespace"); \
   class _OpClass_##op_type##_ : public op_class {                             \
@@ -132,14 +137,14 @@ class OpKernelRegistrar : public Registrar {
   };                                                                          \
   static ::paddle::framework::OpRegistrar<                                    \
       _OpClass_##op_type##_, op_maker_class, _OpGradClass_##op_type##_>       \
-      __op_registrar_##op_type##__(#op_type);                                 \
+      __op_registrar_##op_type##__(#op_type, #grad_op_type);                  \
   int TouchOpRegistrar_##op_type() {                                          \
     __op_registrar_##op_type##__.Touch();                                     \
     return 0;                                                                 \
   }
 
 #define REGISTER_OP_WITHOUT_GRADIENT(op_type, op_class, op_maker_class) \
-  REGISTER_OP(op_type, op_class, op_maker_class, ::paddle::framework::NOP)
+  REGISTER_OP(op_type, op_class, op_maker_class, , ::paddle::framework::NOP)
 
 /**
  * Macro to register OperatorKernel.
@@ -194,6 +199,8 @@ class OpKernelRegistrar : public Registrar {
   USE_OP_DEVICE_KERNEL(op_type, GPU)
 #endif
 
+#define USE_NO_KERNEL_OP(op_type) USE_OP_ITSELF(op_type);
+
 #define USE_CPU_ONLY_OP(op_type) \
   USE_OP_ITSELF(op_type);        \
   USE_OP_DEVICE_KERNEL(op_type, CPU);