From b65709e4039f338d90391b0fed9b8f6118b23380 Mon Sep 17 00:00:00 2001 From: dangqingqing Date: Tue, 19 Sep 2017 16:44:28 +0800 Subject: [PATCH] Share LoD between input and output of each opeators. --- paddle/framework/operator.h | 8 +++++++ paddle/operators/accuracy_op.cc | 7 ++++++- paddle/operators/cos_sim_op.cc | 10 ++++++--- paddle/operators/elementwise_mul_op.cc | 5 +++++ paddle/operators/fc_op.cc | 4 ++++ paddle/operators/fill_zeros_like_op.cc | 21 +++++++++---------- paddle/operators/fill_zeros_like_op.h | 2 +- paddle/operators/lookup_table_op.cc | 11 +++++++--- paddle/operators/mean_op.cc | 3 ++- paddle/operators/minus_op.cc | 8 ++++++- paddle/operators/mul_op.cc | 10 +++++++-- paddle/operators/onehot_cross_entropy_op.cc | 3 +++ paddle/operators/prelu_op.cc | 3 +++ paddle/operators/rowwise_add_op.cc | 1 + paddle/operators/scale_op.cc | 1 + paddle/operators/sigmoid_op.cc | 1 + paddle/operators/squared_l2_distance_op.cc | 4 ++++ paddle/operators/sum_op.cc | 8 +++++-- .../tests/test_fill_zeros_like_op.py | 4 ++-- 19 files changed, 87 insertions(+), 27 deletions(-) diff --git a/paddle/framework/operator.h b/paddle/framework/operator.h index b7c9c39402..28a253ec0b 100644 --- a/paddle/framework/operator.h +++ b/paddle/framework/operator.h @@ -336,6 +336,14 @@ class InferShapeContext { return &var->Get(); } + void ShareLoD(const std::string& in, const std::string& out) const { + PADDLE_ENFORCE(InputVar(in)->IsType(), + "The Input(%s) must be LoDTensor.", in); + PADDLE_ENFORCE(OutputVar(out)->IsType(), + "The Output(%s) must be LoDTensor.", out); + Output(out)->set_lod(Input(in)->lod()); + } + private: const OperatorBase& op_; const Scope& scope_; diff --git a/paddle/operators/accuracy_op.cc b/paddle/operators/accuracy_op.cc index 0c813748b2..32479ae5a3 100644 --- a/paddle/operators/accuracy_op.cc +++ b/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("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"); } }; diff --git a/paddle/operators/cos_sim_op.cc b/paddle/operators/cos_sim_op.cc index 72c4464936..840848fa08 100644 --- a/paddle/operators/cos_sim_op.cc +++ b/paddle/operators/cos_sim_op.cc @@ -57,6 +57,7 @@ class CosSimOp : public framework::OperatorWithKernel { ctx.Output("Out")->Resize({x_dims[0], 1}); ctx.Output("XNorm")->Resize({x_dims[0], 1}); ctx.Output("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"); } }; diff --git a/paddle/operators/elementwise_mul_op.cc b/paddle/operators/elementwise_mul_op.cc index ee6e975b44..304e45fa5b 100644 --- a/paddle/operators/elementwise_mul_op.cc +++ b/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("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"); } }; diff --git a/paddle/operators/fc_op.cc b/paddle/operators/fc_op.cc index e5d0f3c372..56fe654d1e 100644 --- a/paddle/operators/fc_op.cc +++ b/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"); } }; diff --git a/paddle/operators/fill_zeros_like_op.cc b/paddle/operators/fill_zeros_like_op.cc index ba7857cc65..a238b59b78 100644 --- a/paddle/operators/fill_zeros_like_op.cc +++ b/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("Dst")->Resize( - ctx.Input("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("Y")->Resize( + ctx.Input("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. diff --git a/paddle/operators/fill_zeros_like_op.h b/paddle/operators/fill_zeros_like_op.h index 969998ce2e..4474581784 100644 --- a/paddle/operators/fill_zeros_like_op.h +++ b/paddle/operators/fill_zeros_like_op.h @@ -23,7 +23,7 @@ template class FillZerosLikeKernel : public framework::OpKernel { public: void Compute(const framework::ExecutionContext& context) const override { - auto* output = context.Output("Dst"); + auto* output = context.Output("Y"); output->mutable_data(context.GetPlace()); auto t = framework::EigenVector::Flatten(*output); t.device(context.GetEigenDevice()) = t.constant(static_cast(0)); diff --git a/paddle/operators/lookup_table_op.cc b/paddle/operators/lookup_table_op.cc index 07f6dfabca..8f533f1cc3 100644 --- a/paddle/operators/lookup_table_op.cc +++ b/paddle/operators/lookup_table_op.cc @@ -35,6 +35,7 @@ class LookupTableOp : public framework::OperatorWithKernel { auto output_t = ctx.Output("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"); } }; diff --git a/paddle/operators/mean_op.cc b/paddle/operators/mean_op.cc index 7d7eeb59a2..96540ff454 100644 --- a/paddle/operators/mean_op.cc +++ b/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"); } }; diff --git a/paddle/operators/minus_op.cc b/paddle/operators/minus_op.cc index a97bbecdca..5036f9f98a 100644 --- a/paddle/operators/minus_op.cc +++ b/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("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"); } }; diff --git a/paddle/operators/mul_op.cc b/paddle/operators/mul_op.cc index b6d320b415..b2409a1870 100644 --- a/paddle/operators/mul_op.cc +++ b/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("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"); } }; diff --git a/paddle/operators/onehot_cross_entropy_op.cc b/paddle/operators/onehot_cross_entropy_op.cc index f38be3549f..1d87032d27 100644 --- a/paddle/operators/onehot_cross_entropy_op.cc +++ b/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("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"); } }; diff --git a/paddle/operators/prelu_op.cc b/paddle/operators/prelu_op.cc index 7ae80b2968..2b7b82a3e1 100644 --- a/paddle/operators/prelu_op.cc +++ b/paddle/operators/prelu_op.cc @@ -38,6 +38,7 @@ class PReluOp : public framework::OperatorWithKernel { "Output(Out) should not be null"); auto *out = ctx.Output("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"); } }; diff --git a/paddle/operators/rowwise_add_op.cc b/paddle/operators/rowwise_add_op.cc index 2a3fd3be94..90cdb2558b 100644 --- a/paddle/operators/rowwise_add_op.cc +++ b/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("Out")->Resize(x_dims); + ctx.ShareLoD("X", "Out"); } }; diff --git a/paddle/operators/scale_op.cc b/paddle/operators/scale_op.cc index d1f42e8662..ca1bc4ac80 100644 --- a/paddle/operators/scale_op.cc +++ b/paddle/operators/scale_op.cc @@ -35,6 +35,7 @@ class ScaleOp : public framework::OperatorWithKernel { auto *in = ctx.Input("X"); auto *out = ctx.Output("Out"); out->Resize(in->dims()); + ctx.ShareLoD("X", "Out"); } }; diff --git a/paddle/operators/sigmoid_op.cc b/paddle/operators/sigmoid_op.cc index 992b19965e..42befa22d0 100644 --- a/paddle/operators/sigmoid_op.cc +++ b/paddle/operators/sigmoid_op.cc @@ -30,6 +30,7 @@ class SigmoidOp : public framework::OperatorWithKernel { ctx.Output("Y")->Resize( ctx.Input("X")->dims()); + ctx.ShareLoD("X", "Y"); } }; diff --git a/paddle/operators/squared_l2_distance_op.cc b/paddle/operators/squared_l2_distance_op.cc index 39f4305877..dfe8e6decd 100644 --- a/paddle/operators/squared_l2_distance_op.cc +++ b/paddle/operators/squared_l2_distance_op.cc @@ -57,6 +57,7 @@ class SquaredL2DistanceOp : public framework::OperatorWithKernel { ctx.Output("sub_result") ->Resize({x_dims[0], x->numel() / x_dims[0]}); ctx.Output("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"); } }; diff --git a/paddle/operators/sum_op.cc b/paddle/operators/sum_op.cc index 41e05c27f9..ebc57d6b7b 100644 --- a/paddle/operators/sum_op.cc +++ b/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"); } }; diff --git a/python/paddle/v2/framework/tests/test_fill_zeros_like_op.py b/python/paddle/v2/framework/tests/test_fill_zeros_like_op.py index 2473daaba2..eff8fa87d9 100644 --- a/python/paddle/v2/framework/tests/test_fill_zeros_like_op.py +++ b/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() -- GitLab