Add Bilinear Tensor Product operator.

paddle/operators/bilinear_tensor_product_op.cc

@@ -0,0 +1,161 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+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/operators/bilinear_tensor_product_op.h"
+
+namespace paddle {
+namespace operators {
+
+using framework::Tensor;
+
+class BilinearTensorProductOp : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Weight"),
+                   "Input(Weight) should not be null.");
+    PADDLE_ENFORCE(ctx->HasOutput("Out"), "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, "The input X must be a 2D Tensor.");
+    PADDLE_ENFORCE_EQ(y_dims.size(), 2UL, "The input Y must be a 2D Tensor.");
+    PADDLE_ENFORCE_EQ(weight_dims.size(), 3UL,
+                      "The input Weight must be a 3D tensor.");
+    PADDLE_ENFORCE(weight_dims[0],
+                   "The first dimension of Weight must be larger than 0.");
+    PADDLE_ENFORCE(weight_dims[1],
+                   "The second dimension of Weight must be larger than 0.");
+    PADDLE_ENFORCE(weight_dims[2],
+                   "The third dimension of Weight must be larger than 0.");
+    PADDLE_ENFORCE_EQ(x_dims[0], y_dims[0],
+                      "The first dimension(batch_size) of X must be "
+                      "equal to the first dimension of the Y.");
+    PADDLE_ENFORCE_EQ(x_dims[1], weight_dims[1],
+                      "The second dimension of X must be equal to the second "
+                      "dimension of the Weight.");
+    PADDLE_ENFORCE_EQ(y_dims[1], weight_dims[2],
+                      "The second dimension of Y must be equal to the third "
+                      "dimension of the Weight.");
+
+    if (ctx->HasInput("Bias")) {
+      auto bias_dims = ctx->GetInputDim("Bias");
+      PADDLE_ENFORCE_EQ(bias_dims.size(), 2UL,
+                        "The input Bias must have 2 dimensions.");
+      PADDLE_ENFORCE_EQ(bias_dims[0], 1UL,
+                        "The first dimention of input Bias must be 1.");
+      PADDLE_ENFORCE_EQ(bias_dims[1], weight_dims[0],
+                        "The second dimension of Bias must be equal to the  "
+                        "first dimension of the Weight.");
+    }
+
+    ctx->SetOutputDim("Out", {x_dims[0], weight_dims[0]});
+    ctx->ShareLoD("X", /*->*/ "Out");
+  }
+};
+
+class BilinearTensorProductOpMaker : public framework::OpProtoAndCheckerMaker {
+ public:
+  BilinearTensorProductOpMaker(framework::OpProto* proto,
+                               framework::OpAttrChecker* op_checker)
+      : OpProtoAndCheckerMaker(proto, op_checker) {
+    AddInput("X", "The first input of BilinearTensorProduct op.");
+    AddInput("Y", "The second input of BilinearTensorProduct op.");
+    AddInput("Weight", "The input weight of BilinearTensorProduct op.");
+    AddInput("Bias", "The input bias of BilinearTensorProduct op.")
+        .AsDispensable();
+    AddOutput("Out", "The output of BilinearTensorProduct op.");
+    AddComment(R"DOC(
+Bilinear Tensor Product operator.
+Given input X and Y, a 3D tensor weight, and bias. Each column of the
+output is computed by one slice i = 1, . . . , k of the tensor:
+
+    M =  (X W_i) \cdot Y
+    Out_i = \sum_i {M_i} + Bias_i
+
+)DOC");
+  }
+};
+
+class BilinearTensorProductOpGrad : public framework::OperatorWithKernel {
+ public:
+  using framework::OperatorWithKernel::OperatorWithKernel;
+
+ protected:
+  void InferShape(framework::InferShapeContext* ctx) const override {
+    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput("Weight"),
+                   "Input(Weight) should not be null.");
+    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
+                   "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,
+                      "The Out@GRAD must be a 2D Tensor.");
+    PADDLE_ENFORCE_EQ(
+        x_dims[0], out_dims[0],
+        "The first dimension(batch_size) of Out@GRAD must be equal to "
+        "the first dimension of the Input(X).");
+    PADDLE_ENFORCE_EQ(weight_dims[0], out_dims[1],
+                      "The second dimension of Out@GRAD 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[1], out_dims[1],
+                        "The second dimension of Out@GRAD must be equal to "
+                        "the second dimension of the Input(Bias).");
+      auto bias_grad_name = framework::GradVarName("Bias");
+      if (ctx->HasOutput(bias_grad_name))
+        ctx->SetOutputDim(bias_grad_name, bias_dims);
+    }
+
+    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);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP(bilinear_tensor_product, ops::BilinearTensorProductOp,
+            ops::BilinearTensorProductOpMaker, bilinear_tensor_product_grad,
+            ops::BilinearTensorProductOpGrad);
+REGISTER_OP_CPU_KERNEL(
+    bilinear_tensor_product,
+    ops::BilinearTensorProductKernel<paddle::platform::CPUPlace, float>);
+REGISTER_OP_CPU_KERNEL(
+    bilinear_tensor_product_grad,
+    ops::BilinearTensorProductGradKernel<paddle::platform::CPUPlace, float>);

paddle/operators/bilinear_tensor_product_op.cu

@@ -0,0 +1,24 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+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. */
+
+#define EIGEN_USE_GPU
+#include "paddle/operators/bilinear_tensor_product_op.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_GPU_KERNEL(
+    bilinear_tensor_product,
+    ops::BilinearTensorProductKernel<paddle::platform::GPUPlace, float>);
+REGISTER_OP_GPU_KERNEL(
+    bilinear_tensor_product_grad,
+    ops::BilinearTensorProductGradKernel<paddle::platform::GPUPlace, float>);

paddle/operators/bilinear_tensor_product_op.h

@@ -0,0 +1,182 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+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/framework/eigen.h"
+#include "paddle/framework/op_registry.h"
+#include "paddle/operators/math/math_function.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 T, int MajorType = Eigen::RowMajor,
+          typename IndexType = Eigen::DenseIndex>
+using EigenVector = framework::EigenVector<T, MajorType, IndexType>;
+
+template <typename Place, 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();
+    auto place = ctx.GetEigenDevice<Place>();
+
+    // Create the intermediate variables.
+    Tensor left_mul;
+    left_mul.mutable_data<T>(framework::make_ddim({batch_size, weight_dims[2]}),
+                             ctx.GetPlace());
+    auto left_mul_mat = EigenMatrix<T>::From(left_mul);
+
+    for (size_t i = 0; i < weight_dims[0]; ++i) {
+      auto output_col_vec = output_mat.chip(i, 1);
+      Tensor weight_mat = weight->Slice(i, i + 1).Resize(
+          framework::make_ddim({weight_dims[1], weight_dims[2]}));
+      math::gemm<Place, T>(ctx.device_context(), CblasNoTrans, CblasNoTrans,
+                           batch_size, weight_dims[2], weight_dims[1], 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 Place, 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();
+
+    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.GetEigenDevice<Place>();
+
+    // Create the intermediate variables for gradient.
+    Tensor x_scale;
+    x_scale.mutable_data<T>(framework::make_ddim({batch_size, weight_dims[1]}),
+                            ctx.GetPlace());
+    auto x_scale_mat = EigenMatrix<T>::From(x_scale);
+    Tensor y_scale;
+    y_scale.mutable_data<T>(framework::make_ddim({batch_size, weight_dims[2]}),
+                            ctx.GetPlace());
+    auto y_scale_mat = EigenMatrix<T>::From(y_scale);
+
+    math::SetConstant<Place, T> set_zero;
+
+    // Set X@Grad be zero at first.
+    if (d_x) {
+      d_x->mutable_data<T>(ctx.GetPlace());
+      set_zero(ctx.device_context(), d_x, static_cast<T>(0));
+    }
+
+    // Set Y@Grad be zero at first.
+    if (d_y) {
+      d_y->mutable_data<T>(ctx.GetPlace());
+      set_zero(ctx.device_context(), d_y, static_cast<T>(0));
+    }
+
+    // Caculate the X@Grad and Y@Grad.
+    if (d_x || d_y) {
+      Eigen::DSizes<int, 2> bcast_for_x(1, weight_dims[2]);
+      Eigen::DSizes<int, 2> bcast_for_y(1, weight_dims[1]);
+      for (int i = 0; i < weight_dims[0]; ++i) {
+        Tensor weight_i = weight->Slice(i, i + 1).Resize(
+            framework::make_ddim({weight_dims[1], weight_dims[2]}));
+        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;
+          math::gemm<Place, T>(ctx.device_context(), CblasNoTrans, CblasTrans,
+                               batch_size, weight_dims[1], weight_dims[2], 1,
+                               y_scale.data<T>(), weight_i.data<T>(), 1,
+                               d_x->data<T>());
+        }
+        if (d_y) {
+          x_scale_mat.device(place) =
+              output_vec.reshape(Eigen::DSizes<int, 2>(batch_size, 1))
+                  .broadcast(bcast_for_y) *
+              x_mat;
+          math::gemm<Place, T>(ctx.device_context(), CblasNoTrans, CblasNoTrans,
+                               batch_size, weight_dims[2], weight_dims[1], 1,
+                               x_scale.data<T>(), weight_i.data<T>(), 1,
+                               d_y->data<T>());
+        }
+      }
+    }
+
+    // Caculate the gradient of Weight.
+    if (d_weight) {
+      d_weight->mutable_data<T>(ctx.GetPlace());
+      Eigen::DSizes<int, 2> bcast_for_weight(1, weight_dims[1]);
+      for (int i = 0; i < weight_dims[0]; ++i) {
+        Tensor d_weight_i = d_weight->Slice(i, i + 1).Resize(
+            framework::make_ddim({weight_dims[1], weight_dims[2]}));
+        auto output_vec = d_out_mat.chip(i, 1);
+        x_scale_mat.device(place) =
+            output_vec.reshape(Eigen::DSizes<int, 2>(batch_size, 1))
+                .broadcast(bcast_for_weight) *
+            x_mat;
+        math::gemm<Place, T>(ctx.device_context(), CblasTrans, CblasNoTrans,
+                             weight_dims[1], weight_dims[2], batch_size, 1,
+                             x_scale.data<T>(), y->data<T>(), 0,
+                             d_weight_i.data<T>());
+      }
+    }
+
+    // Caculate the gradient of Bias.
+    if (d_bias) {
+      d_bias->mutable_data<T>(ctx.GetPlace());
+      auto d_bias_mat = EigenMatrix<T>::From(*d_bias);
+      d_bias_mat.device(place) = d_out_mat.sum(Eigen::DSizes<int, 1>(0));
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle