Skip to content

P
Paddle

PaddlePaddle / Paddle
大约 1 年 前同步成功

通知 2298
Star 20931
Fork 5422
P
Paddle
未验证 提交 de8f2748 编写于 作者: F From00 提交者： GitHub
浏览文件
操作

Move BilinearTensorProduct OP to phi (#39903) 

* Move BilinearTensorProduct OP to phi

* Set dtype for Infermeta
上级 a456dda6
隐藏空白更改
内联 并排
Showing with 569 addition and 347 deletion
paddle/fluid/operators/bilinear_tensor_product_op.cc
浏览文件 @ de8f2748
...@@ -12,84 +12,18 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. ...@@ -12,84 +12,18 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and See the License for the specific language governing permissions and
limitations under the License. */ limitations under the License. */
#include "paddle/fluid/operators/bilinear_tensor_product_op.h" #include "paddle/fluid/framework/infershape_utils.h"
#include <memory> #include "paddle/fluid/framework/op_registry.h"
#include <string> #include "paddle/phi/core/infermeta_utils.h"
#include <vector> #include "paddle/phi/infermeta/backward.h"
#include "paddle/phi/infermeta/multiary.h"
namespace paddle { namespace paddle {
namespace operators { namespace operators {
using framework::Tensor;
class BilinearTensorProductOp : public framework::OperatorWithKernel { class BilinearTensorProductOp : public framework::OperatorWithKernel {
public: public:
using framework::OperatorWithKernel::OperatorWithKernel; using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE_EQ(
ctx->HasInput("X"), true,
platform::errors::InvalidArgument("Input(X) should not be null."));
PADDLE_ENFORCE_EQ(
ctx->HasInput("Y"), true,
platform::errors::InvalidArgument("Input(Y) should not be null."));
PADDLE_ENFORCE_EQ(
ctx->HasInput("Weight"), true,
platform::errors::InvalidArgument("Input(Weight) should not be null."));
PADDLE_ENFORCE_EQ(
ctx->HasOutput("Out"), true,
platform::errors::InvalidArgument("Output(Out) should not be null."));
auto x_dims = ctx->GetInputDim("X");
auto y_dims = ctx->GetInputDim("Y");
auto weight_dims = ctx->GetInputDim("Weight");
PADDLE_ENFORCE_EQ(
x_dims.size(), 2UL,
platform::errors::InvalidArgument("The input(X) must be a 2D Tensor."));
PADDLE_ENFORCE_EQ(
y_dims.size(), 2UL,
platform::errors::InvalidArgument("The input(Y) must be a 2D Tensor."));
PADDLE_ENFORCE_EQ(
weight_dims.size(), 3UL,
platform::errors::InvalidArgument("Expected the input(Weight) is a 3D "
"tensor. But received %dD tensor.",
weight_dims.size()));
if (ctx->IsRuntime() || (x_dims[0] > 0 && y_dims[0] > 0)) {
PADDLE_ENFORCE_EQ(
x_dims[0], y_dims[0],
platform::errors::InvalidArgument(
"The first dimension(batch_size) of input(X) must be "
"equal to the first dimension of the input(Y)."));
}
PADDLE_ENFORCE_EQ(x_dims[1], weight_dims[1],
platform::errors::InvalidArgument(
"The second dimension of input(X) must be equal to "
"the second dimension of the input(Weight)."));
PADDLE_ENFORCE_EQ(y_dims[1], weight_dims[2],
platform::errors::InvalidArgument(
"The second dimension of input(Y) must be equal to "
"the third dimension of the input(Weight)."));
if (ctx->HasInput("Bias")) {
auto bias_dims = ctx->GetInputDim("Bias");
PADDLE_ENFORCE_EQ(bias_dims.size(), 2UL,
platform::errors::InvalidArgument(
"The Input(Bias) must be a 2-D tensor with "
"the 2nd dimension fixed to 1 (a row vector)."));
PADDLE_ENFORCE_EQ(bias_dims[0], 1UL,
platform::errors::InvalidArgument(
"The Input(Bias) must be a 2-D tensor with "
"the 2nd dimension fixed to 1 (a row vector)."));
PADDLE_ENFORCE_EQ(bias_dims[1], weight_dims[0],
platform::errors::InvalidArgument(
"The second dimension of input(Bias) must be equal "
"to the first dimension of the input(Weight)."));
}
ctx->SetOutputDim("Out", {x_dims[0], weight_dims[0]});
ctx->ShareLoD("X", /*->*/ "Out");
}
}; };
class BilinearTensorProductOpMaker : public framework::OpProtoAndCheckerMaker { class BilinearTensorProductOpMaker : public framework::OpProtoAndCheckerMaker {
...@@ -125,59 +59,6 @@ Where $W_i$ is the $i$-th slice of Input(Weight); ...@@ -125,59 +59,6 @@ Where $W_i$ is the $i$-th slice of Input(Weight);
class BilinearTensorProductOpGrad : public framework::OperatorWithKernel { class BilinearTensorProductOpGrad : public framework::OperatorWithKernel {
public: public:
using framework::OperatorWithKernel::OperatorWithKernel; using framework::OperatorWithKernel::OperatorWithKernel;
protected:
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE_EQ(
ctx->HasInput("X"), true,
platform::errors::InvalidArgument("Input(X) should not be null."));
PADDLE_ENFORCE_EQ(
ctx->HasInput("Y"), true,
platform::errors::InvalidArgument("Input(Y) should not be null."));
PADDLE_ENFORCE_EQ(
ctx->HasInput("Weight"), true,
platform::errors::InvalidArgument("Input(Weight) should not be null."));
PADDLE_ENFORCE_EQ(ctx->HasInput(framework::GradVarName("Out")), true,
platform::errors::InvalidArgument(
"Input(Out@GRAD) should not be null."));
auto x_dims = ctx->GetInputDim("X");
auto y_dims = ctx->GetInputDim("Y");
auto weight_dims = ctx->GetInputDim("Weight");
auto out_dims = ctx->GetInputDim(framework::GradVarName("Out"));
PADDLE_ENFORCE_EQ(out_dims.size(), 2UL,
platform::errors::InvalidArgument(
"The input(Out@GRAD) must be a 2D Tensor."));
PADDLE_ENFORCE_EQ(
x_dims[0], out_dims[0],
platform::errors::InvalidArgument(
"The first dimension(batch_size) of input(Out@GRAD) must be "
"equal to the first dimension of the Input(X)."));
PADDLE_ENFORCE_EQ(
weight_dims[0], out_dims[1],
platform::errors::InvalidArgument(
"The second dimension of input(Out@GRAD) must be equal to "
"the third dimension of the Input(Weight)."));
auto bias_grad_name = framework::GradVarName("Bias");
if (ctx->HasOutput(bias_grad_name)) {
ctx->SetOutputDim(bias_grad_name, {1, out_dims[1]});
}
auto x_grad_name = framework::GradVarName("X");
auto y_grad_name = framework::GradVarName("Y");
auto weight_grad_name = framework::GradVarName("Weight");
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);
}
if (ctx->HasOutput(weight_grad_name)) {
ctx->SetOutputDim(weight_grad_name, weight_dims);
}
}
}; };
template <typename T> template <typename T>
...@@ -208,21 +89,20 @@ class BilinearTensorProductGradOpMaker ...@@ -208,21 +89,20 @@ class BilinearTensorProductGradOpMaker
} // namespace paddle } // namespace paddle
namespace ops = paddle::operators; namespace ops = paddle::operators;
DELCARE_INFER_SHAPE_FUNCTOR(bilinear_tensor_product,
BilinearTensorProductInferShapeFunctor,
PT_INFER_META(phi::BilinearTensorProductInferMeta));
DELCARE_INFER_SHAPE_FUNCTOR(
bilinear_tensor_product_grad, BilinearTensorProductGradInferShapeFunctor,
PT_INFER_META(phi::BilinearTensorProductGradInferMeta));
REGISTER_OPERATOR( REGISTER_OPERATOR(
bilinear_tensor_product, ops::BilinearTensorProductOp, bilinear_tensor_product, ops::BilinearTensorProductOp,
ops::BilinearTensorProductOpMaker, ops::BilinearTensorProductOpMaker,
ops::BilinearTensorProductGradOpMaker<paddle::framework::OpDesc>, ops::BilinearTensorProductGradOpMaker<paddle::framework::OpDesc>,
ops::BilinearTensorProductGradOpMaker<paddle::imperative::OpBase>); ops::BilinearTensorProductGradOpMaker<paddle::imperative::OpBase>,
BilinearTensorProductInferShapeFunctor);
REGISTER_OPERATOR(bilinear_tensor_product_grad, REGISTER_OPERATOR(bilinear_tensor_product_grad,
ops::BilinearTensorProductOpGrad); ops::BilinearTensorProductOpGrad,
REGISTER_OP_CPU_KERNEL( BilinearTensorProductGradInferShapeFunctor);
bilinear_tensor_product,
ops::BilinearTensorProductKernel<paddle::platform::CPUDeviceContext, float>,
ops::BilinearTensorProductKernel<paddle::platform::CPUDeviceContext,
double>);
REGISTER_OP_CPU_KERNEL(
bilinear_tensor_product_grad,
ops::BilinearTensorProductGradKernel<paddle::platform::CPUDeviceContext,
float>,
ops::BilinearTensorProductGradKernel<paddle::platform::CPUDeviceContext,
double>);
paddle/fluid/operators/bilinear_tensor_product_op.cu 已删除 100644 → 0
浏览文件 @ a456dda6
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
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. */
#include "paddle/fluid/operators/bilinear_tensor_product_op.h"
namespace ops = paddle::operators;
REGISTER_OP_CUDA_KERNEL(
bilinear_tensor_product,
ops::BilinearTensorProductKernel<paddle::platform::CUDADeviceContext,
float>,
ops::BilinearTensorProductKernel<paddle::platform::CUDADeviceContext,
double>);
REGISTER_OP_CUDA_KERNEL(
bilinear_tensor_product_grad,
ops::BilinearTensorProductGradKernel<paddle::platform::CUDADeviceContext,
float>,
ops::BilinearTensorProductGradKernel<paddle::platform::CUDADeviceContext,
double>);
paddle/fluid/operators/bilinear_tensor_product_op.h 已删除 100644 → 0
浏览文件 @ a456dda6
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
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/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/phi/kernels/funcs/blas/blas.h"
namespace paddle {
namespace operators {
using framework::Tensor;
template <typename T, int MajorType = Eigen::RowMajor,
typename IndexType = Eigen::DenseIndex>
using EigenMatrix = framework::EigenMatrix<T, MajorType, IndexType>;
template <typename DeviceContext, typename T>
class BilinearTensorProductKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
auto* x = ctx.Input<Tensor>("X");
auto* y = ctx.Input<Tensor>("Y");
auto* weight = ctx.Input<Tensor>("Weight");
auto* bias = ctx.Input<Tensor>("Bias");
auto* out = ctx.Output<Tensor>("Out");
out->mutable_data<T>(ctx.GetPlace());
auto y_mat = EigenMatrix<T>::From(*y);
auto output_mat = EigenMatrix<T>::From(*out);
auto batch_size = x->dims()[0];
auto weight_dims = weight->dims();
int out_dim = weight_dims[0];
auto x_dim = weight_dims[1];
auto y_dim = weight_dims[2];
auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
auto& dev_ctx = ctx.template device_context<DeviceContext>();
// Create the intermediate variable to calculate the result of
// Input(X) multiplied by Input(Weight_i), the formula is:
// left_mul = X Weight_i.
Tensor left_mul;
left_mul.mutable_data<T>(phi::make_ddim({batch_size, y_dim}),
ctx.GetPlace());
auto left_mul_mat = EigenMatrix<T>::From(left_mul);
for (int i = 0; i < out_dim; ++i) {
auto output_col_vec = output_mat.chip(i, 1);
Tensor weight_mat =
weight->Slice(i, i + 1).Resize(phi::make_ddim({x_dim, y_dim}));
phi::funcs::GetBlas<DeviceContext, T>(dev_ctx).GEMM(
CblasNoTrans, CblasNoTrans, batch_size, y_dim, x_dim, 1, x->data<T>(),
weight_mat.data<T>(), 0, left_mul.data<T>());
output_col_vec.device(place) =
(left_mul_mat * y_mat).sum(Eigen::DSizes<int, 1>(1));
}
if (bias) {
auto bias_vec = EigenMatrix<T>::From(*bias);
Eigen::DSizes<int, 2> bcast(batch_size, 1);
output_mat.device(place) = bias_vec.broadcast(bcast) + output_mat;
}
}
};
template <typename DeviceContext, typename T>
class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& ctx) const override {
const Tensor* x = ctx.Input<Tensor>("X");
const Tensor* y = ctx.Input<Tensor>("Y");
const Tensor* weight = ctx.Input<Tensor>("Weight");
Tensor* d_x = ctx.Output<Tensor>(framework::GradVarName("X"));
Tensor* d_y = ctx.Output<Tensor>(framework::GradVarName("Y"));
Tensor* d_weight = ctx.Output<Tensor>(framework::GradVarName("Weight"));
Tensor* d_bias = ctx.Output<Tensor>(framework::GradVarName("Bias"));
const Tensor* d_out = ctx.Input<Tensor>(framework::GradVarName("Out"));
auto batch_size = x->dims()[0];
auto weight_dims = weight->dims();
int out_dim = weight_dims[0];
auto x_dim = weight_dims[1];
auto y_dim = weight_dims[2];
auto x_mat = EigenMatrix<T>::From(*x);
auto y_mat = EigenMatrix<T>::From(*y);
auto d_out_mat = EigenMatrix<T>::From(*d_out);
auto& place = *ctx.template device_context<DeviceContext>().eigen_device();
auto& dev_ctx = ctx.template device_context<DeviceContext>();
// Create the intermediate variable to calculate the Output(Y@Grad).
Tensor x_scale;
x_scale.mutable_data<T>(phi::make_ddim({batch_size, x_dim}),
ctx.GetPlace());
auto x_scale_mat = EigenMatrix<T>::From(x_scale);
// Create the intermediate variable to calculate the Output(X@Grad).
Tensor y_scale;
y_scale.mutable_data<T>(phi::make_ddim({batch_size, y_dim}),
ctx.GetPlace());
auto y_scale_mat = EigenMatrix<T>::From(y_scale);
phi::funcs::SetConstant<DeviceContext, T> set_zero;
if (d_x) {
d_x->mutable_data<T>(ctx.GetPlace());
set_zero(dev_ctx, d_x, static_cast<T>(0));
}
if (d_y) {
d_y->mutable_data<T>(ctx.GetPlace());
set_zero(dev_ctx, d_y, static_cast<T>(0));
}
if (d_weight) {
d_weight->mutable_data<T>(ctx.GetPlace());
}
auto blas = phi::funcs::GetBlas<DeviceContext, T>(ctx);
// Caculate the Output(X@Grad) and Output(Y@Grad).
if (d_x || d_y || d_weight) {
Eigen::DSizes<int, 2> bcast_for_x(1, y_dim);
Eigen::DSizes<int, 2> bcast_for_y(1, x_dim);
Eigen::DSizes<int, 2> bcast_for_weight(1, x_dim);
for (int i = 0; i < out_dim; ++i) {
Tensor weight_i =
weight->Slice(i, i + 1).Resize(phi::make_ddim({x_dim, y_dim}));
auto output_vec = d_out_mat.chip(i, 1);
if (d_x) {
y_scale_mat.device(place) =
output_vec.reshape(Eigen::DSizes<int, 2>(batch_size, 1))
.broadcast(bcast_for_x) *
y_mat;
blas.GEMM(CblasNoTrans, CblasTrans, batch_size, x_dim, y_dim, 1,
y_scale.data<T>(), weight_i.data<T>(), 1, d_x->data<T>());
}
if (d_y || d_weight) {
auto output_vec_y =
output_vec.reshape(Eigen::DSizes<int, 2>(batch_size, 1))
.broadcast(bcast_for_y);
x_scale_mat.device(place) = output_vec_y * x_mat;
if (d_y) {
blas.GEMM(CblasNoTrans, CblasNoTrans, batch_size, y_dim, x_dim, 1,
x_scale.data<T>(), weight_i.data<T>(), 1, d_y->data<T>());
}
if (d_weight) {
Tensor d_weight_i = d_weight->Slice(i, i + 1).Resize(
phi::make_ddim({x_dim, y_dim}));
blas.GEMM(CblasTrans, CblasNoTrans, x_dim, y_dim, batch_size, 1,
x_scale.data<T>(), y->data<T>(), 0, d_weight_i.data<T>());
}
}
}
}
// calculate the gradient of Input(Bias).
if (d_bias) {
d_bias->mutable_data<T>(ctx.GetPlace());
auto d_bias_mat = framework::EigenVector<T>::Flatten(*d_bias);
d_bias_mat.device(place) = d_out_mat.sum(Eigen::DSizes<int, 1>(0));
}
}
};
} // namespace operators
} // namespace paddle
paddle/phi/infermeta/backward.cc
浏览文件 @ de8f2748
...@@ -16,6 +16,54 @@ limitations under the License. */ ...@@ -16,6 +16,54 @@ limitations under the License. */
namespace phi { namespace phi {
void BilinearTensorProductGradInferMeta(const MetaTensor& x,
const MetaTensor& y,
const MetaTensor& weight,
const MetaTensor& dout,
MetaTensor* dx,
MetaTensor* dy,
MetaTensor* dweight,
MetaTensor* dbias) {
auto x_dims = x.dims();
auto y_dims = y.dims();
auto weight_dims = weight.dims();
auto out_dims = dout.dims();
PADDLE_ENFORCE_EQ(
out_dims.size(),
2UL,
errors::InvalidArgument("The input(Out@GRAD) must be a 2D Tensor."));
PADDLE_ENFORCE_EQ(
x_dims[0],
out_dims[0],
errors::InvalidArgument(
"The first dimension(batch_size) of input(Out@GRAD) must be "
"equal to the first dimension of the Input(X)."));
PADDLE_ENFORCE_EQ(
weight_dims[0],
out_dims[1],
errors::InvalidArgument(
"The second dimension of input(Out@GRAD) must be equal to "
"the third dimension of the Input(Weight)."));
if (dx) {
dx->set_dims(x_dims);
dx->set_dtype(x.dtype());
}
if (dy) {
dy->set_dims(y_dims);
dy->set_dtype(y.dtype());
}
if (dweight) {
dweight->set_dims(weight_dims);
dweight->set_dtype(weight.dtype());
}
if (dbias) {
dbias->set_dims({1, out_dims[1]});
dbias->set_dtype(dout.dtype());
}
}
void GeneralBinaryGradInferMeta(const MetaTensor& x, void GeneralBinaryGradInferMeta(const MetaTensor& x,
const MetaTensor& y, const MetaTensor& y,
MetaTensor* dx, MetaTensor* dx,
......
paddle/phi/infermeta/backward.h
浏览文件 @ de8f2748
...@@ -20,6 +20,15 @@ limitations under the License. */ ...@@ -20,6 +20,15 @@ limitations under the License. */
namespace phi { namespace phi {
void BilinearTensorProductGradInferMeta(const MetaTensor& x,
const MetaTensor& y,
const MetaTensor& weight,
const MetaTensor& dout,
MetaTensor* dx,
MetaTensor* dy,
MetaTensor* dweight,
MetaTensor* dbias);
void GeneralBinaryGradInferMeta(const MetaTensor& x, void GeneralBinaryGradInferMeta(const MetaTensor& x,
const MetaTensor& y, const MetaTensor& y,
MetaTensor* dx, MetaTensor* dx,
......
paddle/phi/infermeta/multiary.cc
浏览文件 @ de8f2748
...@@ -18,6 +18,72 @@ limitations under the License. */ ...@@ -18,6 +18,72 @@ limitations under the License. */
#include "paddle/phi/kernels/funcs/concat_funcs.h" #include "paddle/phi/kernels/funcs/concat_funcs.h"
namespace phi { namespace phi {
void BilinearTensorProductInferMeta(const MetaTensor& x,
const MetaTensor& y,
const MetaTensor& weight,
paddle::optional<const MetaTensor&> bias,
MetaTensor* out,
MetaConfig config) {
auto x_dims = x.dims();
auto y_dims = y.dims();
auto weight_dims = weight.dims();
PADDLE_ENFORCE_EQ(
x_dims.size(),
2UL,
errors::InvalidArgument("The input(X) must be a 2D Tensor."));
PADDLE_ENFORCE_EQ(
y_dims.size(),
2UL,
errors::InvalidArgument("The input(Y) must be a 2D Tensor."));
PADDLE_ENFORCE_EQ(
weight_dims.size(),
3UL,
errors::InvalidArgument(
"Expected the input(Weight) is a 3D tensor. But received %dD tensor.",
weight_dims.size()));
if (config.is_runtime || (x_dims[0] > 0 && y_dims[0] > 0)) {
PADDLE_ENFORCE_EQ(x_dims[0],
y_dims[0],
errors::InvalidArgument(
"The first dimension(batch_size) of input(X) must be "
"equal to the first dimension of the input(Y)."));
}
PADDLE_ENFORCE_EQ(x_dims[1],
weight_dims[1],
errors::InvalidArgument(
"The second dimension of input(X) must be equal to "
"the second dimension of the input(Weight)."));
PADDLE_ENFORCE_EQ(y_dims[1],
weight_dims[2],
errors::InvalidArgument(
"The second dimension of input(Y) must be equal to "
"the third dimension of the input(Weight)."));
if (bias.get_ptr()) {
auto bias_dims = bias->dims();
PADDLE_ENFORCE_EQ(bias_dims.size(),
2UL,
errors::InvalidArgument(
"The Input(Bias) must be a 2-D tensor with "
"the 2nd dimension fixed to 1 (a row vector)."));
PADDLE_ENFORCE_EQ(bias_dims[0],
1UL,
errors::InvalidArgument(
"The Input(Bias) must be a 2-D tensor with "
"the 2nd dimension fixed to 1 (a row vector)."));
PADDLE_ENFORCE_EQ(bias_dims[1],
weight_dims[0],
errors::InvalidArgument(
"The second dimension of input(Bias) must be equal "
"to the first dimension of the input(Weight)."));
}
out->set_dims({x_dims[0], weight_dims[0]});
out->share_lod(x);
out->set_dtype(x.dtype());
}
void ConcatInferMeta(const std::vector<MetaTensor>& x, void ConcatInferMeta(const std::vector<MetaTensor>& x,
const Scalar& axis_scalar, const Scalar& axis_scalar,
MetaTensor* out, MetaTensor* out,
......
paddle/phi/infermeta/multiary.h
浏览文件 @ de8f2748
...@@ -18,6 +18,13 @@ limitations under the License. */ ...@@ -18,6 +18,13 @@ limitations under the License. */
#include "paddle/phi/core/meta_tensor.h" #include "paddle/phi/core/meta_tensor.h"
namespace phi { namespace phi {
void BilinearTensorProductInferMeta(const MetaTensor& x,
const MetaTensor& y,
const MetaTensor& weight,
paddle::optional<const MetaTensor&> bias,
MetaTensor* out,
MetaConfig config = MetaConfig());
void ConcatInferMeta(const std::vector<MetaTensor>& x, void ConcatInferMeta(const std::vector<MetaTensor>& x,
const Scalar& axis_scalar, const Scalar& axis_scalar,
MetaTensor* out, MetaTensor* out,
......
paddle/phi/kernels/bilinear_tensor_product_grad_kernel.h 0 → 100644
浏览文件 @ de8f2748
// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
//
// 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/phi/core/dense_tensor.h"
namespace phi {
template <typename T, typename Context>
void BilinearTensorProductGradKernel(const Context& dev_ctx,
const DenseTensor& x,
const DenseTensor& y,
const DenseTensor& weight,
const DenseTensor& dout,
DenseTensor* dx,
DenseTensor* dy,
DenseTensor* dweight,
DenseTensor* dbias);
} // namespace phi
paddle/phi/kernels/bilinear_tensor_product_kernel.h 0 → 100644
浏览文件 @ de8f2748
// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
//
// 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/phi/core/dense_tensor.h"
#include "paddle/utils/optional.h"
namespace phi {
template <typename T, typename Context>
void BilinearTensorProductKernel(const Context& dev_ctx,
const DenseTensor& x,
const DenseTensor& y,
const DenseTensor& weight,
paddle::optional<const DenseTensor&> bias,
DenseTensor* out);
} // namespace phi
paddle/phi/kernels/cpu/bilinear_tensor_product_grad_kernel.cc 0 → 100644
浏览文件 @ de8f2748
// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
//
// 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.
#include "paddle/phi/kernels/bilinear_tensor_product_grad_kernel.h"
#include "paddle/phi/kernels/impl/bilinear_tensor_product_grad_kernel_impl.h"
#include "paddle/phi/core/kernel_registry.h"
PD_REGISTER_KERNEL(bilinear_tensor_product_grad,
CPU,
ALL_LAYOUT,
phi::BilinearTensorProductGradKernel,
float,
double) {}
paddle/phi/kernels/cpu/bilinear_tensor_product_kernel.cc 0 → 100644
浏览文件 @ de8f2748
// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
//
// 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.
#include "paddle/phi/kernels/bilinear_tensor_product_kernel.h"
#include "paddle/phi/kernels/impl/bilinear_tensor_product_kernel_impl.h"
#include "paddle/phi/core/kernel_registry.h"
PD_REGISTER_KERNEL(bilinear_tensor_product,
CPU,
ALL_LAYOUT,
phi::BilinearTensorProductKernel,
float,
double) {}
paddle/phi/kernels/gpu/bilinear_tensor_product_grad_kernel.cu 0 → 100644
浏览文件 @ de8f2748
// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
//
// 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.
#include "paddle/phi/kernels/bilinear_tensor_product_grad_kernel.h"
#include "paddle/phi/kernels/impl/bilinear_tensor_product_grad_kernel_impl.h"
#include "paddle/phi/core/kernel_registry.h"
PD_REGISTER_KERNEL(bilinear_tensor_product_grad,
GPU,
ALL_LAYOUT,
phi::BilinearTensorProductGradKernel,
float,
double) {}
paddle/phi/kernels/gpu/bilinear_tensor_product_kernel.cu 0 → 100644
浏览文件 @ de8f2748
// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
//
// 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.
#include "paddle/phi/kernels/bilinear_tensor_product_kernel.h"
#include "paddle/phi/kernels/impl/bilinear_tensor_product_kernel_impl.h"
#include "paddle/phi/core/kernel_registry.h"
PD_REGISTER_KERNEL(bilinear_tensor_product,
GPU,
ALL_LAYOUT,
phi::BilinearTensorProductKernel,
float,
double) {}
paddle/phi/kernels/impl/bilinear_tensor_product_grad_kernel_impl.h 0 → 100644
浏览文件 @ de8f2748
// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
//
// 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/phi/core/dense_tensor.h"
#include "paddle/phi/kernels/funcs/blas/blas.h"
#include "paddle/phi/kernels/funcs/eigen/common.h"
namespace phi {
template <typename T, typename Context>
void BilinearTensorProductGradKernel(const Context& ctx,
const DenseTensor& x,
const DenseTensor& y,
const DenseTensor& weight,
const DenseTensor& dout,
DenseTensor* dx,
DenseTensor* dy,
DenseTensor* dweight,
DenseTensor* dbias) {
auto batch_size = x.dims()[0];
auto weight_dims = weight.dims();
int out_dim = weight_dims[0];
auto x_dim = weight_dims[1];
auto y_dim = weight_dims[2];
auto x_mat = EigenMatrix<T>::From(x);
auto y_mat = EigenMatrix<T>::From(y);
auto dout_mat = EigenMatrix<T>::From(dout);
auto& place = *ctx.eigen_device();
// Create the intermediate variable to calculate the Output(Y@Grad).
DenseTensor x_scale;
x_scale.Resize(make_ddim({batch_size, x_dim}));
ctx.template Alloc<T>(&x_scale);
auto x_scale_mat = EigenMatrix<T>::From(x_scale);
// Create the intermediate variable to calculate the Output(X@Grad).
DenseTensor y_scale;
y_scale.Resize(make_ddim({batch_size, y_dim}));
ctx.template Alloc<T>(&y_scale);
auto y_scale_mat = EigenMatrix<T>::From(y_scale);
funcs::SetConstant<Context, T> set_zero;
if (dx) {
ctx.template Alloc<T>(dx);
set_zero(ctx, dx, static_cast<T>(0));
}
if (dy) {
ctx.template Alloc<T>(dy);
set_zero(ctx, dy, static_cast<T>(0));
}
if (dweight) {
ctx.template Alloc<T>(dweight);
}
auto blas = funcs::GetBlas<Context, T>(ctx);
// Caculate the Output(X@Grad) and Output(Y@Grad).
if (dx || dy || dweight) {
Eigen::DSizes<int, 2> bcast_for_x(1, y_dim);
Eigen::DSizes<int, 2> bcast_for_y(1, x_dim);
Eigen::DSizes<int, 2> bcast_for_weight(1, x_dim);
for (int i = 0; i < out_dim; ++i) {
DenseTensor weight_i =
weight.Slice(i, i + 1).Resize(make_ddim({x_dim, y_dim}));
auto output_vec = dout_mat.chip(i, 1);
if (dx) {
y_scale_mat.device(place) =
output_vec.reshape(Eigen::DSizes<int, 2>(batch_size, 1))
.broadcast(bcast_for_x) *
y_mat;
blas.GEMM(CblasNoTrans,
CblasTrans,
batch_size,
x_dim,
y_dim,
1,
y_scale.data<T>(),
weight_i.data<T>(),
1,
dx->data<T>());
}
if (dy || dweight) {
auto output_vec_y =
output_vec.reshape(Eigen::DSizes<int, 2>(batch_size, 1))
.broadcast(bcast_for_y);
x_scale_mat.device(place) = output_vec_y * x_mat;
if (dy) {
blas.GEMM(CblasNoTrans,
CblasNoTrans,
batch_size,
y_dim,
x_dim,
1,
x_scale.data<T>(),
weight_i.data<T>(),
1,
dy->data<T>());
}
if (dweight) {
DenseTensor dweight_i =
dweight->Slice(i, i + 1).Resize(make_ddim({x_dim, y_dim}));
blas.GEMM(CblasTrans,
CblasNoTrans,
x_dim,
y_dim,
batch_size,
1,
x_scale.data<T>(),
y.data<T>(),
0,
dweight_i.data<T>());
}
}
}
}
// calculate the gradient of Input(Bias).
if (dbias) {
ctx.template Alloc<T>(dbias);
auto dbias_mat = EigenVector<T>::Flatten(*dbias);
dbias_mat.device(place) = dout_mat.sum(Eigen::DSizes<int, 1>(0));
}
}
} // namespace phi
paddle/phi/kernels/impl/bilinear_tensor_product_kernel_impl.h 0 → 100644
浏览文件 @ de8f2748
// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
//
// 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/phi/core/dense_tensor.h"
#include "paddle/phi/kernels/funcs/blas/blas.h"
#include "paddle/phi/kernels/funcs/eigen/common.h"
#include "paddle/utils/optional.h"
namespace phi {
template <typename T, typename Context>
void BilinearTensorProductKernel(const Context& ctx,
const DenseTensor& x,
const DenseTensor& y,
const DenseTensor& weight,
paddle::optional<const DenseTensor&> bias,
DenseTensor* out) {
ctx.template Alloc<T>(out);
auto y_mat = EigenMatrix<T>::From(y);
auto output_mat = EigenMatrix<T>::From(*out);
auto batch_size = x.dims()[0];
auto weight_dims = weight.dims();
int out_dim = weight_dims[0];
auto x_dim = weight_dims[1];
auto y_dim = weight_dims[2];
auto& place = *ctx.eigen_device();
// Create the intermediate variable to calculate the result of
// Input(X) multiplied by Input(Weight_i), the formula is:
// left_mul = X Weight_i.
DenseTensor left_mul;
left_mul.Resize(phi::make_ddim({batch_size, y_dim}));
ctx.template Alloc<T>(&left_mul);
auto left_mul_mat = EigenMatrix<T>::From(left_mul);
for (int i = 0; i < out_dim; ++i) {
auto output_col_vec = output_mat.chip(i, 1);
DenseTensor weight_mat =
weight.Slice(i, i + 1).Resize(phi::make_ddim({x_dim, y_dim}));
phi::funcs::GetBlas<Context, T>(ctx).GEMM(CblasNoTrans,
CblasNoTrans,
batch_size,
y_dim,
x_dim,
1,
x.data<T>(),
weight_mat.data<T>(),
0,
left_mul.data<T>());
output_col_vec.device(place) =
(left_mul_mat * y_mat).sum(Eigen::DSizes<int, 1>(1));
}
if (bias.get_ptr()) {
auto bias_vec = EigenMatrix<T>::From(*(bias.get_ptr()));
Eigen::DSizes<int, 2> bcast(batch_size, 1);
output_mat.device(place) = bias_vec.broadcast(bcast) + output_mat;
}
}
} // namespace phi
paddle/phi/ops/compat/bilinear_tensor_product_sig.cc 0 → 100644
浏览文件 @ de8f2748
// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
//
// 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.
#include "paddle/phi/core/compat/op_utils.h"
namespace phi {
KernelSignature BilinearTensorProductOpArgumentMapping(
const ArgumentMappingContext& ctx) {
return KernelSignature(
"bilinear_tensor_product", {"X", "Y", "Weight", "Bias"}, {}, {"Out"});
}
KernelSignature BilinearTensorProductGradOpArgumentMapping(
const ArgumentMappingContext& ctx) {
return KernelSignature("bilinear_tensor_product_grad",
{"X", "Y", "Weight", GradVarName("Out")},
{},
{GradVarName("X"),
GradVarName("Y"),
GradVarName("Weight"),
GradVarName("Bias")});
}
} // namespace phi
PD_REGISTER_ARG_MAPPING_FN(bilinear_tensor_product,
phi::BilinearTensorProductOpArgumentMapping);
PD_REGISTER_ARG_MAPPING_FN(bilinear_tensor_product_grad,
phi::BilinearTensorProductGradOpArgumentMapping);
Markdown is supported
0% 或 .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑！
想要评论请 注册 或