From e1761709f80aca97391d785125fa66dc6cc1bad5 Mon Sep 17 00:00:00 2001 From: chengduo Date: Wed, 10 Oct 2018 10:00:18 +0800 Subject: [PATCH] Set the right shape of selected_rows (#13723) * set the right shape of selected_rows test=develop * enhance check * fix activation_op * remove cast * use ShareDimInfo replace SetDim and ShareLod * use ShareDimAndLod test=develop * follow comment test=develop * check whether the input has lod test=develop * Split ShareDimAndLod test=develop * checkout clip.py test=develop --- .gitignore | 1 + paddle/fluid/framework/op_desc.cc | 21 ++++++++ paddle/fluid/framework/operator.cc | 30 +++++++++++ paddle/fluid/framework/shape_inference.h | 3 ++ paddle/fluid/operators/activation_op.cc | 20 ++++++- paddle/fluid/operators/argsort_op.cc | 4 +- paddle/fluid/operators/conv_shift_op.cc | 2 +- paddle/fluid/operators/elementwise_op.h | 19 ++++--- paddle/fluid/operators/fake_dequantize_op.cc | 3 +- paddle/fluid/operators/lookup_table_op.cc | 1 + paddle/fluid/operators/prelu_op.cc | 2 +- .../fluid/operators/rnn_memory_helper_op.cc | 2 +- paddle/fluid/operators/sequence_conv_op.cc | 4 +- paddle/fluid/operators/sequence_pool_op.cc | 5 +- paddle/fluid/operators/sequence_reshape_op.cc | 2 +- paddle/fluid/operators/sequence_softmax_op.cc | 3 +- .../fluid/operators/shrink_rnn_memory_op.cc | 6 +-- .../sigmoid_cross_entropy_with_logits_op.cc | 2 +- .../unittests/test_elementwise_mul_op.py | 53 +++++++++++++++++++ 19 files changed, 158 insertions(+), 25 deletions(-) diff --git a/.gitignore b/.gitignore index b92bb9cc12..90138f996c 100644 --- a/.gitignore +++ b/.gitignore @@ -25,5 +25,6 @@ third_party/ bazel-* third_party/ +build_* # clion workspace. cmake-build-* diff --git a/paddle/fluid/framework/op_desc.cc b/paddle/fluid/framework/op_desc.cc index 17f942571d..b29ac44699 100644 --- a/paddle/fluid/framework/op_desc.cc +++ b/paddle/fluid/framework/op_desc.cc @@ -50,6 +50,27 @@ class CompileTimeInferShapeContext : public InferShapeContext { const std::vector &Outputs( const std::string &name) const override; + void ShareDim(const std::string &in, const std::string &out, size_t i = 0, + size_t j = 0) override { + PADDLE_ENFORCE_LT(i, Inputs(in).size()); + PADDLE_ENFORCE_LT(j, Outputs(out).size()); + const std::string &input_n = Inputs(in)[i]; + const std::string &output_n = Outputs(out)[j]; + + PADDLE_ENFORCE(input_n != framework::kEmptyVarName, "The %s[%d] is @EMPTY@", + in, i); + PADDLE_ENFORCE(output_n != framework::kEmptyVarName, + "The %s[%d] is @EMPTY@", out, j); + + auto *in_var = block_.FindVarRecursive(input_n); + auto *out_var = block_.FindVarRecursive(output_n); + + PADDLE_ENFORCE(in_var->GetType() == out_var->GetType(), + "The type of %s and %s is not the same.", input_n, output_n); + + SetDim(output_n, GetDim(input_n)); + } + void ShareLoD(const std::string &in, const std::string &out, size_t i = 0, size_t j = 0) const override { PADDLE_ENFORCE_LT(i, Inputs(in).size()); diff --git a/paddle/fluid/framework/operator.cc b/paddle/fluid/framework/operator.cc index 6666dd8e60..9f93006532 100644 --- a/paddle/fluid/framework/operator.cc +++ b/paddle/fluid/framework/operator.cc @@ -542,6 +542,36 @@ class RuntimeInferShapeContext : public InferShapeContext { return op_.Outputs(name); } + void ShareDim(const std::string& in, const std::string& out, size_t i = 0, + size_t j = 0) override { + PADDLE_ENFORCE_LT(i, Inputs(in).size()); + PADDLE_ENFORCE_LT(j, Outputs(out).size()); + const std::string& input_n = Inputs(in)[i]; + const std::string& output_n = Outputs(out)[j]; + + Variable* in_var = scope_.FindVar(input_n); + Variable* out_var = scope_.FindVar(output_n); + PADDLE_ENFORCE(in_var->Type() == out_var->Type(), + "The type of %s and %s is not the same.", output_n, + GetDim(input_n)); + + if (in_var->IsType()) { + auto& in_sele_rows = in_var->Get(); + auto out_sele_rows = out_var->GetMutable(); + out_sele_rows->mutable_value()->Resize(in_sele_rows.value().dims()); + out_sele_rows->set_rows(in_sele_rows.rows()); + out_sele_rows->set_height(in_sele_rows.height()); + } else if (in_var->IsType()) { + auto& in_lod_tensor = in_var->Get(); + auto* out_lod_tensor = out_var->GetMutable(); + out_lod_tensor->Resize(in_lod_tensor.dims()); + } else { + PADDLE_THROW( + "Currently, the input type of ShareDim only can be LoDTensor " + "or SelectedRows."); + } + } + void ShareLoD(const std::string& in, const std::string& out, size_t i = 0, size_t j = 0) const override { const std::vector& inputs = Inputs(in); diff --git a/paddle/fluid/framework/shape_inference.h b/paddle/fluid/framework/shape_inference.h index 5f497cafa0..280bc19dce 100644 --- a/paddle/fluid/framework/shape_inference.h +++ b/paddle/fluid/framework/shape_inference.h @@ -56,6 +56,9 @@ class InferShapeContext { virtual const std::vector &Outputs( const std::string &name) const = 0; + virtual void ShareDim(const std::string &in, const std::string &out, + size_t i = 0, size_t j = 0) = 0; + virtual void ShareLoD(const std::string &in, const std::string &out, size_t i = 0, size_t j = 0) const = 0; diff --git a/paddle/fluid/operators/activation_op.cc b/paddle/fluid/operators/activation_op.cc index c091476d6d..bbf52bea13 100644 --- a/paddle/fluid/operators/activation_op.cc +++ b/paddle/fluid/operators/activation_op.cc @@ -80,7 +80,7 @@ class ActivationOp : public framework::OperatorWithKernel { using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext* ctx) const override { - ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + ctx->ShareDim("X", /*->*/ "Out"); ctx->ShareLoD("X", /*->*/ "Out"); } @@ -91,12 +91,26 @@ class ActivationOp : public framework::OperatorWithKernel { } }; +class ActivationOpInferVarType : public framework::VarTypeInference { + public: + void operator()(const framework::OpDesc& op_desc, + framework::BlockDesc* block) const override { + auto x_name = op_desc.Input("X")[0]; + auto out_name = op_desc.Output("Out")[0]; + auto& x = block->FindRecursiveOrCreateVar(x_name); + auto& out = block->FindRecursiveOrCreateVar(out_name); + out.SetType(x.GetType()); + out.SetDataType(x.GetDataType()); + } +}; + class ActivationOpGrad : public framework::OperatorWithKernel { public: using framework::OperatorWithKernel::OperatorWithKernel; void InferShape(framework::InferShapeContext* ctx) const override { - ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("Out")); + ctx->ShareDim("Out", framework::GradVarName("X")); + ctx->ShareLoD("Out", framework::GradVarName("X")); } protected: @@ -525,12 +539,14 @@ namespace ops = paddle::operators; #define REGISTER_INPLACE_ACTIVATION_OP(OP_NAME, KERNEL_TYPE) \ REGISTER_OPERATOR(KERNEL_TYPE, ::paddle::operators::ActivationOp, \ ::paddle::operators::OP_NAME##OpMaker, \ + ::paddle::operators::ActivationOpInferVarType, \ ::paddle::operators::OP_NAME##GradMaker); \ REGISTER_OPERATOR(KERNEL_TYPE##_grad, ::paddle::operators::ActivationOpGrad) #define REGISTER_ACTIVATION_OP(OP_NAME, KERNEL_TYPE) \ REGISTER_OPERATOR(KERNEL_TYPE, ::paddle::operators::ActivationOp, \ ::paddle::operators::OP_NAME##OpMaker, \ + ::paddle::operators::ActivationOpInferVarType, \ ::paddle::framework::DefaultGradOpDescMaker); \ REGISTER_OPERATOR(KERNEL_TYPE##_grad, ::paddle::operators::ActivationOpGrad) diff --git a/paddle/fluid/operators/argsort_op.cc b/paddle/fluid/operators/argsort_op.cc index a2f5a25457..d25160f423 100644 --- a/paddle/fluid/operators/argsort_op.cc +++ b/paddle/fluid/operators/argsort_op.cc @@ -42,8 +42,8 @@ class ArgsortOp : public framework::OperatorWithKernel { "-rank(Input(X)) (%d).", axis, num_dims); - ctx->SetOutputDim("Out", in_dims); - ctx->SetOutputDim("Indices", in_dims); + ctx->ShareDim("X", "Out"); + ctx->ShareDim("X", "Indices"); ctx->ShareLoD("X", "Out"); ctx->ShareLoD("X", "Indices"); } diff --git a/paddle/fluid/operators/conv_shift_op.cc b/paddle/fluid/operators/conv_shift_op.cc index f2549e814d..08506ddd18 100644 --- a/paddle/fluid/operators/conv_shift_op.cc +++ b/paddle/fluid/operators/conv_shift_op.cc @@ -44,7 +44,7 @@ class ConvShiftOp : public framework::OperatorWithKernel { PADDLE_ENFORCE_LE(y_dims[1], x_dims[1], "The 2nd dimension of Input(Y) should be less than or " "equal to the 2nd dimension of Input(X)."); - ctx->SetOutputDim("Out", x_dims); + ctx->ShareDim("X", /*->*/ "Out"); ctx->ShareLoD("X", /*->*/ "Out"); } }; diff --git a/paddle/fluid/operators/elementwise_op.h b/paddle/fluid/operators/elementwise_op.h index 94df11bee7..7e5975ead6 100644 --- a/paddle/fluid/operators/elementwise_op.h +++ b/paddle/fluid/operators/elementwise_op.h @@ -41,7 +41,8 @@ class ElementwiseOp : public framework::OperatorWithKernel { 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->SetOutputDim("Out", x_dim); + + ctx->ShareDim("X", /*->*/ "Out"); ctx->ShareLoD("X", /*->*/ "Out"); } @@ -70,6 +71,7 @@ class ElementwiseOpInferVarType : public framework::VarTypeInference { auto& x = block->FindRecursiveOrCreateVar(x_name); auto& out = block->FindRecursiveOrCreateVar(out_name); out.SetType(x.GetType()); + out.SetDataType(x.GetDataType()); } }; @@ -157,10 +159,12 @@ class ElementwiseOpGrad : public framework::OperatorWithKernel { 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); + ctx->ShareDim("X", /*->*/ x_grad_name); + ctx->ShareLoD("X", /*->*/ x_grad_name); } if (ctx->HasOutput(y_grad_name)) { - ctx->SetOutputDim(y_grad_name, y_dims); + ctx->ShareDim("Y", /*->*/ y_grad_name); + ctx->ShareLoD("Y", /*->*/ y_grad_name); } } @@ -193,14 +197,15 @@ class ElementwiseOpExplicitGrad : public ElementwiseOpGrad { auto x_grad_name = framework::GradVarName("X"); if (ctx->HasOutput(x_grad_name)) { - auto out_dims = ctx->GetInputDim(framework::GradVarName("Out")); - ctx->SetOutputDim(x_grad_name, out_dims); + ctx->ShareDim(framework::GradVarName("Out"), /*->*/ x_grad_name); + ctx->ShareLoD(framework::GradVarName("Out"), /*->*/ x_grad_name); } auto y_grad_name = framework::GradVarName("Y"); if (ctx->HasOutput(y_grad_name)) { PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should not be null"); - auto y_dims = ctx->GetInputDim("Y"); - ctx->SetOutputDim(y_grad_name, y_dims); + + ctx->ShareDim("Y", /*->*/ y_grad_name); + ctx->ShareLoD("Y", /*->*/ y_grad_name); } } }; diff --git a/paddle/fluid/operators/fake_dequantize_op.cc b/paddle/fluid/operators/fake_dequantize_op.cc index 2008e70275..5d6488c67e 100644 --- a/paddle/fluid/operators/fake_dequantize_op.cc +++ b/paddle/fluid/operators/fake_dequantize_op.cc @@ -48,7 +48,8 @@ class FakeDequantizeMaxAbsOp : public framework::OperatorWithKernel { "Input(X) of FakeDequantizeMaxAbsOp should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of FakeDequantizeMaxAbsOp should not be null."); - ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + + ctx->ShareDim("X", /*->*/ "Out"); ctx->ShareLoD("X", /*->*/ "Out"); } }; diff --git a/paddle/fluid/operators/lookup_table_op.cc b/paddle/fluid/operators/lookup_table_op.cc index d77b095c5d..b9ac54e446 100644 --- a/paddle/fluid/operators/lookup_table_op.cc +++ b/paddle/fluid/operators/lookup_table_op.cc @@ -137,6 +137,7 @@ class LookupTableOpGradVarTypeInference : public framework::VarTypeInference { << " is set to LoDTensor"; block->Var(out_var_name)->SetType(framework::proto::VarType::LOD_TENSOR); } + block->Var(out_var_name)->SetDataType(block->Var("W")->GetDataType()); } }; diff --git a/paddle/fluid/operators/prelu_op.cc b/paddle/fluid/operators/prelu_op.cc index e0c4c81bdd..58cfbb76e9 100644 --- a/paddle/fluid/operators/prelu_op.cc +++ b/paddle/fluid/operators/prelu_op.cc @@ -49,7 +49,7 @@ class PReluOp : public framework::OperatorWithKernel { } else { PADDLE_THROW("Unkown mode %s", mode); } - ctx->SetOutputDim("Out", x_dim); + ctx->ShareDim("X", /*->*/ "Out"); ctx->ShareLoD("X", /*->*/ "Out"); } diff --git a/paddle/fluid/operators/rnn_memory_helper_op.cc b/paddle/fluid/operators/rnn_memory_helper_op.cc index 13df1d4b4b..0fb7776fd9 100644 --- a/paddle/fluid/operators/rnn_memory_helper_op.cc +++ b/paddle/fluid/operators/rnn_memory_helper_op.cc @@ -54,7 +54,7 @@ class RNNMemoryHelperOpShapeInference : public framework::InferShapeBase { "Input(X) of rnn_memory_helper op should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output of rnn_memory_helper op should not be null."); - ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + ctx->ShareDim("X", /*->*/ "Out"); ctx->ShareLoD("X", /*->*/ "Out"); } }; diff --git a/paddle/fluid/operators/sequence_conv_op.cc b/paddle/fluid/operators/sequence_conv_op.cc index ec6cb24350..95a21a5d3e 100644 --- a/paddle/fluid/operators/sequence_conv_op.cc +++ b/paddle/fluid/operators/sequence_conv_op.cc @@ -90,8 +90,8 @@ class SequenceConvGradOp : public framework::OperatorWithKernel { ctx->GetInputDim("PaddingData")); } if (ctx->HasOutput(framework::GradVarName("X"))) { - ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); - ctx->ShareLoD("X", framework::GradVarName("X")); + ctx->ShareDim("X", /*->*/ framework::GradVarName("X")); + ctx->ShareLoD("X", /*->*/ framework::GradVarName("X")); } if (ctx->HasOutput(framework::GradVarName("Filter"))) { ctx->SetOutputDim(framework::GradVarName("Filter"), diff --git a/paddle/fluid/operators/sequence_pool_op.cc b/paddle/fluid/operators/sequence_pool_op.cc index 5c6fd13d42..15d3f064eb 100644 --- a/paddle/fluid/operators/sequence_pool_op.cc +++ b/paddle/fluid/operators/sequence_pool_op.cc @@ -102,8 +102,9 @@ class SequencePoolGradOp : public framework::OperatorWithKernel { for (int64_t i = 1; i < og_dims.size(); ++i) { PADDLE_ENFORCE_EQ(og_dims[i], x_dims[i], "The dimension mismatch."); } - ctx->SetOutputDim(framework::GradVarName("X"), x_dims); - ctx->ShareLoD("X", framework::GradVarName("X")); + + ctx->ShareDim("X", /*->*/ framework::GradVarName("X")); + ctx->ShareLoD("X", /*->*/ framework::GradVarName("X")); } protected: diff --git a/paddle/fluid/operators/sequence_reshape_op.cc b/paddle/fluid/operators/sequence_reshape_op.cc index ef5e6f3210..31d28d7234 100644 --- a/paddle/fluid/operators/sequence_reshape_op.cc +++ b/paddle/fluid/operators/sequence_reshape_op.cc @@ -92,7 +92,7 @@ class SequenceReshapeGradOp : public framework::OperatorWithKernel { PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) of SequenceReshapeGradOp should not be null."); - ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X")); + ctx->ShareDim("X", /*->*/ framework::GradVarName("X")); ctx->ShareLoD("X", /*->*/ framework::GradVarName("X")); } }; diff --git a/paddle/fluid/operators/sequence_softmax_op.cc b/paddle/fluid/operators/sequence_softmax_op.cc index c44f8206eb..ada3e0c8db 100644 --- a/paddle/fluid/operators/sequence_softmax_op.cc +++ b/paddle/fluid/operators/sequence_softmax_op.cc @@ -27,7 +27,8 @@ class SequenceSoftmaxOp : public framework::OperatorWithKernel { "Input(X) of SequenceSoftmaxOp should not be null."); PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) of SequenceSoftmaxOp should not be null."); - ctx->SetOutputDim("Out", ctx->GetInputDim("X")); + + ctx->ShareDim("X", /*->*/ "Out"); ctx->ShareLoD("X", /*->*/ "Out"); } diff --git a/paddle/fluid/operators/shrink_rnn_memory_op.cc b/paddle/fluid/operators/shrink_rnn_memory_op.cc index 29d2fb9897..e1c74c3a2f 100644 --- a/paddle/fluid/operators/shrink_rnn_memory_op.cc +++ b/paddle/fluid/operators/shrink_rnn_memory_op.cc @@ -151,9 +151,9 @@ class ShrinkRNNMemoryGradInferShape : public framework::InferShapeBase { void operator()(framework::InferShapeContext *context) const override { PADDLE_ENFORCE(context->HasInput("X")); PADDLE_ENFORCE(context->HasOutput(framework::GradVarName("X"))); - context->SetOutputDim(framework::GradVarName("X"), - context->GetInputDim("X")); - context->ShareLoD("X", framework::GradVarName("X")); + + context->ShareDim("X", /*->*/ framework::GradVarName("X")); + context->ShareLoD("X", /*->*/ framework::GradVarName("X")); } }; diff --git a/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc b/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc index c3b0fe3209..193de05422 100644 --- a/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc +++ b/paddle/fluid/operators/sigmoid_cross_entropy_with_logits_op.cc @@ -40,7 +40,7 @@ class SigmoidCrossEntropyWithLogitsOp : public framework::OperatorWithKernel { "The 2nd dimension of Input(X) and Input(Label) should " "be equal."); - ctx->SetOutputDim("Out", x_dims); + ctx->ShareDim("X", /*->*/ "Out"); ctx->ShareLoD("X", /*->*/ "Out"); } }; diff --git a/python/paddle/fluid/tests/unittests/test_elementwise_mul_op.py b/python/paddle/fluid/tests/unittests/test_elementwise_mul_op.py index 775c2253ab..6a129b6df9 100644 --- a/python/paddle/fluid/tests/unittests/test_elementwise_mul_op.py +++ b/python/paddle/fluid/tests/unittests/test_elementwise_mul_op.py @@ -16,6 +16,8 @@ from __future__ import print_function import unittest import numpy as np from op_test import OpTest +import paddle.fluid.core as core +from paddle.fluid.op import Operator class ElementwiseMulOp(OpTest): @@ -115,5 +117,56 @@ class TestElementwiseMulOp_broadcast_3(ElementwiseMulOp): } +class TestElementWiseMulSelectedRows(OpTest): + def setUp(self): + self.rows = [0, 1, 2, 3, 4, 5, 6] + self.feature = 12 + self.height = 100 + self.input_shape = (len(self.rows), self.feature) + + def prepare_input(self, scope, place): + self.input = { + "X": np.random.random(self.input_shape).astype("float32"), + "Y": np.random.random(self.input_shape).astype("float32") + } + + def init_input(in_name): + x_selected_rows = scope.var(in_name).get_selected_rows() + x_selected_rows.set_height(self.height) + x_selected_rows.set_rows(self.rows) + x_array = self.input[in_name] + x_tensor = x_selected_rows.get_tensor() + x_tensor.set(x_array, place) + + init_input("X") + init_input("Y") + + def create_out_selected_row(self, scope): + return scope.var('Out').get_selected_rows() + + def check_result(self, out_selected_rows): + assert out_selected_rows.height() == self.height + assert out_selected_rows.rows() == self.rows + out_tensor = np.array(out_selected_rows.get_tensor()) + assert out_tensor.shape == self.input_shape + + def check_with_place(self, place): + scope = core.Scope() + self.prepare_input(scope, place) + + out_selected_rows = self.create_out_selected_row(scope) + out_selected_rows.set_height(0) + out_selected_rows.set_rows([]) + + elementwise_mul = Operator("elementwise_mul", X='X', Y='Y', Out='Out') + elementwise_mul.run(scope, place) + self.check_result(out_selected_rows) + + def test_elewisemul_with_selected_rows_input(self): + places = [core.CPUPlace()] + for place in places: + self.check_with_place(place) + + if __name__ == '__main__': unittest.main() -- GitLab