Share LoD between input and output of each opeators.

paddle/framework/operator.h

@@ -336,6 +336,14 @@ class InferShapeContext {
     return &var->Get<Tensor>();
   }
 
+  void ShareLoD(const std::string& in, const std::string& out) const {
+    PADDLE_ENFORCE(InputVar(in)->IsType<LoDTensor>(),
+                   "The Input(%s) must be LoDTensor.", in);
+    PADDLE_ENFORCE(OutputVar(out)->IsType<LoDTensor>(),
+                   "The Output(%s) must be LoDTensor.", out);
+    Output<LoDTensor>(out)->set_lod(Input<LoDTensor>(in)->lod());
+  }
+
  private:
   const OperatorBase& op_;
   const Scope& scope_;

paddle/operators/accuracy_op.cc

@@ -40,6 +40,7 @@ class AccuracyOp : public framework::OperatorWithKernel {
                       "inference size must be the same as label size");
 
     ctx.Output<framework::LoDTensor>("Accuracy")->Resize({1});
+    ctx.ShareLoD("Inference", "Accuracy");
   }
 };
 
@@ -58,7 +59,11 @@ class AccuracyOpMaker : public framework::OpProtoAndCheckerMaker {
         R"DOC(Accuracy. It will print accuracy rate for classification.
 The accuracy is:
 ..  math::
-accuracy = \\frac{NumOfCorrectPredicts}{NumOfAllSamples})DOC");
+accuracy = \\frac{NumOfCorrectPredicts}{NumOfAllSamples})
+
+Both the input `Inference` and `Label` can carry the LoD (Level of Details)
+information, or not. But the output only shares the LoD with input `Inference`.
+DOC");
   }
 };
 

paddle/operators/cos_sim_op.cc

@@ -57,6 +57,7 @@ class CosSimOp : public framework::OperatorWithKernel {
     ctx.Output<framework::LoDTensor>("Out")->Resize({x_dims[0], 1});
     ctx.Output<framework::LoDTensor>("XNorm")->Resize({x_dims[0], 1});
     ctx.Output<framework::LoDTensor>("YNorm")->Resize({y_dims[0], 1});
+    ctx.ShareLoD("X", "Out");
   }
 };
 
@@ -81,10 +82,13 @@ Cosine Similarity Operator.
 
 The equation is: Out = X^T * Y / (sqrt(X^T * X) * sqrt(Y^T * Y)).
 
-Input(X) and Input(Y) must have the same shape, except that the 1st dimension
-of Input(Y) could be just 1 (different from Input(X)), which will be
-broadcasted to match the shape of Input(X) before computing their cosine
+The input `X` and `Y` must have the same shape, except that the 1st dimension
+of input `Y` could be just 1 (different from input `X`), which will be
+broadcasted to match the shape of input `X` before computing their cosine
 similarity.
+
+Both the input `X` and `Y` can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with input `X`.
 )DOC");
   }
 };

paddle/operators/elementwise_mul_op.cc

@@ -38,6 +38,7 @@ class ElementWiseMulOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_GE(x_dim.size(), y_dim.size(),
                       "Rank of first input must >= rank of second input.")
     ctx.Output<framework::LoDTensor>("Out")->Resize(x_dim);
+    ctx.ShareLoD("X", "Out");
   }
 };
 
@@ -63,11 +64,15 @@ Limited elementwise multiple operator.The equation is: Out = X ⊙ Y.
 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:
+
       shape(X) = (2, 3, 4, 5), shape(Y) = (,)
       shape(X) = (2, 3, 4, 5), shape(Y) = (5,)
       shape(X) = (2, 3, 4, 5), shape(Y) = (4, 5)
       shape(X) = (2, 3, 4, 5), shape(Y) = (3, 4), with axis=1
       shape(X) = (2, 3, 4, 5), shape(Y) = (2), with axis=0
+
+Both the input X and Y can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with input X.
 )DOC");
   }
 };

paddle/operators/fc_op.cc

@@ -186,6 +186,10 @@ W_i is a 2-D matrix of size (K x N), where N means the number of neurons
 in the fully connected layer. B is a 1-D vector of size N.
 Thus, the output Out is a 2-D matrix of size (M x N).
 Activation type can be set to `identity` (default), `sigmoid` or `softmax`.
+
+All the inputs can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with first input (`X[0]`).
+)DOC");
 )DOC");
   }
 };

paddle/operators/fill_zeros_like_op.cc

@@ -23,15 +23,14 @@ class FillZerosLikeOp : public framework::OperatorWithKernel {
 
  protected:
   void InferShape(const framework::InferShapeContext &ctx) const override {
-    PADDLE_ENFORCE_NOT_NULL(
-        ctx.InputVar("Src"),
-        "Input(Src) of FillZerosLikeOp should not be null.");
-    PADDLE_ENFORCE_NOT_NULL(
-        ctx.OutputVar("Dst"),
-        "Output(Dst) of FillZerosLikeOp should not be null.");
-
-    ctx.Output<framework::LoDTensor>("Dst")->Resize(
-        ctx.Input<framework::Tensor>("Src")->dims());
+    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::LoDTensor>("Y")->Resize(
+        ctx.Input<framework::Tensor>("X")->dims());
+    ctx.ShareLoD("X", "Y");
   }
 };
 
@@ -40,8 +39,8 @@ class FillZerosLikeOpMaker : public framework::OpProtoAndCheckerMaker {
   FillZerosLikeOpMaker(framework::OpProto *proto,
                        framework::OpAttrChecker *op_checker)
       : framework::OpProtoAndCheckerMaker(proto, op_checker) {
-    AddInput("Src", "The input of fill-zeros-like op.");
-    AddOutput("Dst", "The varibale will be filled up with zeros.");
+    AddInput("X", "The input of fill-zeros-like op.");
+    AddOutput("Y", "The varibale will be filled up with zeros.");
     AddComment(R"DOC(
 Fill up a vriable with zeros.
 

paddle/operators/fill_zeros_like_op.h

@@ -23,7 +23,7 @@ template <typename Place, typename T>
 class FillZerosLikeKernel : public framework::OpKernel {
  public:
   void Compute(const framework::ExecutionContext& context) const override {
-    auto* output = context.Output<framework::Tensor>("Dst");
+    auto* output = context.Output<framework::Tensor>("Y");
     output->mutable_data<T>(context.GetPlace());
     auto t = framework::EigenVector<T>::Flatten(*output);
     t.device(context.GetEigenDevice<Place>()) = t.constant(static_cast<T>(0));

paddle/operators/lookup_table_op.cc

@@ -35,6 +35,7 @@ class LookupTableOp : public framework::OperatorWithKernel {
     auto output_t = ctx.Output<framework::LoDTensor>("Out");
 
     output_t->Resize({ids_t->dims()[0], table_t->dims()[1]});
+    ctx.ShareLoD("Ids", "Out");
   }
 };
 
@@ -50,9 +51,13 @@ class LookupTableOpMaker : public framework::OpProtoAndCheckerMaker {
              "An input with type int32 or int64"
              "contains the ids to be looked up in W.");
     AddOutput("Out", "The lookup results, which have the same type with W.");
-    AddComment(
-        "This operator is used to perform lookups on the parameter W,"
-        "then concatenated into a dense tensor.");
+    AddComment(R"DOC(
+This operator is used to perform lookups on the parameter W,
+then concatenated into a dense tensor.
+
+The input `Ids` can carry the LoD (Level of Details) information,
+or not. And the output only shares the LoD with input `Ids`.
+)DOC");
   }
 };
 

paddle/operators/mean_op.cc

@@ -37,7 +37,8 @@ class MeanOpMaker : public framework::OpProtoAndCheckerMaker {
       : OpProtoAndCheckerMaker(proto, op_checker) {
     AddInput("X", "The input of mean op");
     AddOutput("Out", "The output of mean op").NotInGradient();
-    AddComment("Mean Operator");
+    AddComment(R"DOC( Mean Operator
+)DOC");
   }
 };
 

paddle/operators/minus_op.cc

@@ -41,6 +41,7 @@ class MinusOp : public framework::OperatorWithKernel {
         left_tensor->numel(), right_tensor->numel(),
         "Minus operator must take two tensor with same num of elements");
     ctx.Output<framework::LoDTensor>("Out")->Resize(left_tensor->dims());
+    ctx.ShareLoD("X", "Out");
   }
 };
 
@@ -54,7 +55,12 @@ class MinusOpMaker : public framework::OpProtoAndCheckerMaker {
 
     AddComment(R"DOC(Minus Operator
 
-Equation: Out = X - Y
+Equation:
+
+    Out = X - Y
+
+Both the input `X` and `Y` can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with input `X`.
 )DOC");
   }
 };

paddle/operators/mul_op.cc

@@ -55,6 +55,7 @@ class MulOp : public framework::OperatorWithKernel {
         "First matrix's width must be equal with second matrix's height.");
     ctx.Output<framework::LoDTensor>("Out")->Resize(
         {x_mat_dims[0], y_mat_dims[1]});
+    ctx.ShareLoD("X", "Out");
   }
 };
 
@@ -83,9 +84,14 @@ class MulOpMaker : public framework::OpProtoAndCheckerMaker {
         .SetDefault(1)
         .EqualGreaterThan(1);
     AddComment(R"DOC(
-Two Element Mul Operator.
+Mul operator is used to perform matrix multiplication for input X and Y.
 
-The equation is: Out = X * Y
+The equation is:
+
+    Out = X * Y
+
+Both the input `X` and `Y` can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with input `X`.
 )DOC");
   }
 };

paddle/operators/onehot_cross_entropy_op.cc

@@ -40,6 +40,7 @@ class OnehotCrossEntropyOp : public framework::OperatorWithKernel {
     PADDLE_ENFORCE_EQ(label->dims().size(), 1, "label's dimension must be 1.");
     PADDLE_ENFORCE_EQ(X->dims()[0], label->dims()[0]);
     ctx.Output<framework::LoDTensor>("Y")->Resize({X->dims()[0], 1});
+    ctx.ShareLoD("X", "Y");
   }
 };
 
@@ -69,6 +70,8 @@ OnehotCrossEntropy Operator.
 
                 Y[i] = -log(X[i][j])
 
+Both the input `X` and `Label` can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with input `X`.
 )DOC");
   }
 };

paddle/operators/prelu_op.cc

@@ -38,6 +38,7 @@ class PReluOp : public framework::OperatorWithKernel {
                             "Output(Out) should not be null");
     auto *out = ctx.Output<framework::LoDTensor>("Out");
     out->Resize(in->dims());
+    ctx.ShareLoD("X", "Out");
   }
 };
 
@@ -55,6 +56,8 @@ The equation is:
   f(x) = alpha * x , for x < 0
   f(x) = x         , for x >= 0
 
+The input `X` can carry the LoD (Level of Details) information,
+or not. And the output shares the LoD with input `X`.
 )DOC");
   }
 };

paddle/operators/rowwise_add_op.cc

@@ -45,6 +45,7 @@ class RowwiseAddOp : public framework::OperatorWithKernel {
         "The width of two operands must be same");
     PADDLE_ENFORCE_EQ(ctx.OutputSize("Out"), 1, "The output size must be 1");
     ctx.Output<framework::LoDTensor>("Out")->Resize(x_dims);
+    ctx.ShareLoD("X", "Out");
   }
 };
 

paddle/operators/scale_op.cc

@@ -35,6 +35,7 @@ class ScaleOp : public framework::OperatorWithKernel {
     auto *in = ctx.Input<framework::Tensor>("X");
     auto *out = ctx.Output<framework::LoDTensor>("Out");
     out->Resize(in->dims());
+    ctx.ShareLoD("X", "Out");
   }
 };
 

paddle/operators/sigmoid_op.cc

@@ -30,6 +30,7 @@ class SigmoidOp : public framework::OperatorWithKernel {
 
     ctx.Output<framework::LoDTensor>("Y")->Resize(
         ctx.Input<Tensor>("X")->dims());
+    ctx.ShareLoD("X", "Y");
   }
 };
 

paddle/operators/squared_l2_distance_op.cc

@@ -57,6 +57,7 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel {
     ctx.Output<framework::LoDTensor>("sub_result")
         ->Resize({x_dims[0], x->numel() / x_dims[0]});
     ctx.Output<framework::LoDTensor>("Out")->Resize({x_dims[0], 1});
+    ctx.ShareLoD("X", "Out");
   }
 };
 
@@ -79,6 +80,9 @@ class SquaredL2DistanceOpMaker : public framework::OpProtoAndCheckerMaker {
     input or to 1. If the first dimension of target is 1, SquaredL2DistanceOp
     will broadcast target's first dimension to input's first dimension.
     You can decide whether calculate the gradient of input and target.
+
+    Both the input X and Y can carry the LoD (Level of Details) information,
+    or not. But the output only shares the LoD with input X.
     )DOC");
   }
 };

paddle/operators/sum_op.cc

@@ -39,6 +39,7 @@ class SumOp : public framework::OperatorWithKernel {
       PADDLE_ENFORCE(in_dim == dim, "Input tensors must have same shape");
     }
     out->Resize(in_dim);
+    ctx.ShareLoD(ctx.op().Inputs("X")[0], "Out");
   }
 };
 
@@ -49,8 +50,11 @@ class SumOpMaker : public framework::OpProtoAndCheckerMaker {
     AddInput("X", "the input tensors of sum operator.").AsDuplicable();
     AddOutput("Out", "the output tensor of sum operator.");
     AddComment(R"DOC(
-            Sum the input tensors.
-        )DOC");
+Sum the input tensors.
+
+All the inputs can carry the LoD (Level of Details) information,
+or not. But the output only shares the LoD with the first input.
+)DOC");
   }
 };
 

python/paddle/v2/framework/tests/test_fill_zeros_like_op.py

@@ -6,8 +6,8 @@ from op_test import OpTest
 class TestFillZerosLikeOp(OpTest):
     def setUp(self):
         self.op_type = "fill_zeros_like"
-        self.inputs = {'Src': np.random.random((219, 232)).astype("float32")}
-        self.outputs = {'Dst': np.zeros_like(self.inputs["Src"])}
+        self.inputs = {'X': np.random.random((219, 232)).astype("float32")}
+        self.outputs = {'Y': np.zeros_like(self.inputs["X"])}
 
     def test_check_output(self):
         self.check_output()