cos_sim_op.cc 6.6 KB
Newer Older
X
Xinghai Sun 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
/* 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/cos_sim_op.h"

namespace paddle {
namespace operators {

using framework::Tensor;

class CosSimOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

26
  void InferShape(framework::InferShapeContext* ctx) const override {
27
    // notnull check
Q
Qiao Longfei 已提交
28 29 30 31 32 33 34 35 36 37
    PADDLE_ENFORCE(ctx->HasInput("X"),
                   "Input(X) of CosSimOp should not be null.");
    PADDLE_ENFORCE(ctx->HasInput("Y"),
                   "Input(Y) of CosSimOp should not be null.");
    PADDLE_ENFORCE(ctx->HasOutput("Out"),
                   "Output(Out) of CosSimOp should not be null.");
    PADDLE_ENFORCE(ctx->HasOutput("XNorm"),
                   "Output(XNorm) of CosSimOp should not be null.");
    PADDLE_ENFORCE(ctx->HasOutput("YNorm"),
                   "Output(YNorm) of CosSimOp should not be null.");
38 39

    // shape check
Q
Qiao Longfei 已提交
40 41
    auto x_dims = ctx->GetInputDim("X");
    auto y_dims = ctx->GetInputDim("Y");
42 43

    PADDLE_ENFORCE_EQ(x_dims.size(), y_dims.size(),
44
                      "Ranks of Input(X) and Input(Y) must be equal.");
45
    PADDLE_ENFORCE_GE(x_dims.size(), 2,
46
                      "Rank of Input(X) must not be less than 2.");
47 48 49 50
    PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 1, x_dims.size()),
                      framework::slice_ddim(y_dims, 1, y_dims.size()),
                      "All dimensions except the 1st of Input(X) and Input(Y) "
                      "must be equal.");
51
    PADDLE_ENFORCE(x_dims[0] == y_dims[0] || y_dims[0] == 1,
52 53
                   "The 1st dimension of Input(Y) must be equal to Input(X) or"
                   " just 1 (which will be broadcasted to match Input(X)).");
54 55

    // resize tensor
Q
Qiao Longfei 已提交
56 57 58 59
    ctx->SetOutputDim("Out", {x_dims[0], 1});
    ctx->SetOutputDim("XNorm", {x_dims[0], 1});
    ctx->SetOutputDim("YNorm", {y_dims[0], 1});
    ctx->ShareLoD("X", /*->*/ "Out");
X
Xinghai Sun 已提交
60 61 62 63 64
  }
};

class CosSimOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
65
  CosSimOpMaker(OpProto* proto, OpAttrChecker* op_checker)
X
Xinghai Sun 已提交
66
      : OpProtoAndCheckerMaker(proto, op_checker) {
67 68
    AddInput("X", "The 1st input of cos_sim op.");
    AddInput("Y", "The 2nd input of cos_sim op.");
X
Xinghai Sun 已提交
69
    AddOutput("Out", "The output of cos_sim op.");
70 71 72 73 74 75 76 77
    AddOutput("XNorm",
              "Norm of the first input, reduced along the 1st "
              "dimension.")
        .AsIntermediate();
    AddOutput("YNorm",
              "Norm of the second input, reduced along the 1st "
              "dimension.")
        .AsIntermediate();
78

X
Xinghai Sun 已提交
79 80 81
    AddComment(R"DOC(
Cosine Similarity Operator.

K
Kexin Zhao 已提交
82
$Out = X^T * Y / (\sqrt{X^T * X} * \sqrt{Y^T * Y})$
83

K
Kexin Zhao 已提交
84 85 86
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
87
similarity.
88

K
Kexin Zhao 已提交
89 90 91
Both the input X and Y can carry the LoD (Level of Details) information,
or not. But the output only shares the LoD information with input X.

X
Xinghai Sun 已提交
92 93 94 95 96 97 98 99
)DOC");
  }
};

class CosSimOpGrad : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

100
  void InferShape(framework::InferShapeContext* ctx) const override {
101
    // notnull check
Q
Qiao Longfei 已提交
102 103 104 105 106 107 108
    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) must not be null.");
    PADDLE_ENFORCE(ctx->HasInput("Y"), "Input(Y) must not be null.");
    PADDLE_ENFORCE(ctx->HasInput("XNorm"), "Input(XNorm) must not be null.");
    PADDLE_ENFORCE(ctx->HasInput("YNorm"), "Input(YNorm) must not be null.");
    PADDLE_ENFORCE(ctx->HasInput("Out"), "Input(Out) must not be null.");
    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
                   "Input(Out@GRAD) must not be null.");
X
Xinghai Sun 已提交
109

110
    // shape check
Q
Qiao Longfei 已提交
111 112 113 114 115 116
    auto x_dims = ctx->GetInputDim("X");
    auto y_dims = ctx->GetInputDim("Y");
    auto xnorm_dims = ctx->GetInputDim("XNorm");
    auto ynorm_dims = ctx->GetInputDim("YNorm");
    auto out_dims = ctx->GetInputDim("Out");
    auto out_grad_dims = ctx->GetInputDim(framework::GradVarName("Out"));
117 118 119 120 121 122 123 124 125 126 127 128

    PADDLE_ENFORCE_GE(x_dims.size(), y_dims.size(),
                      "Ranks of Input(X) and Input(Y) must be equal.");
    PADDLE_ENFORCE_GE(x_dims.size(), 2,
                      "Rank of Input(X) must not be less than 2.");
    PADDLE_ENFORCE_EQ(framework::slice_ddim(x_dims, 1, x_dims.size()),
                      framework::slice_ddim(y_dims, 1, y_dims.size()),
                      "All dimensions except the 1st of Input(X) and Input(Y) "
                      "must be equal.");
    PADDLE_ENFORCE(x_dims[0] == y_dims[0] || y_dims[0] == 1,
                   "The 1st dimension of Input(Y) must be equal to Input(X) or"
                   " just 1 (which will be broadcasted to match Input(X)).");
129 130 131 132
    auto target_xnorm_dims = framework::make_ddim({x_dims[0], 1});
    auto target_ynorm_dims = framework::make_ddim({y_dims[0], 1});
    PADDLE_ENFORCE_EQ(xnorm_dims, target_xnorm_dims,
                      "Shape of Input(XNorm) must be [X.Dim(0), 1].");
133 134 135 136 137
    PADDLE_ENFORCE_EQ(ynorm_dims, target_ynorm_dims,
                      "Shape of Input(YNorm) must be [Y.Dim(0), 1].");
    PADDLE_ENFORCE_EQ(out_dims, target_xnorm_dims,
                      "Shape of Input(Out) must be [X.Dim(0), 1].");
    PADDLE_ENFORCE_EQ(out_grad_dims, target_xnorm_dims,
138 139 140
                      "Shape of Input(Out@Grad) must be [X.Dim(0), 1].");

    // resize tensor
Q
Qiao Longfei 已提交
141 142 143 144 145 146 147 148
    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);
    }
    if (ctx->HasOutput(y_grad_name)) {
      ctx->SetOutputDim(y_grad_name, y_dims);
    }
X
Xinghai Sun 已提交
149 150 151 152 153 154 155 156 157 158
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
REGISTER_OP(cos_sim, ops::CosSimOp, ops::CosSimOpMaker, cos_sim_grad,
            ops::CosSimOpGrad);
REGISTER_OP_CPU_KERNEL(
Q
QI JUN 已提交
159 160 161 162
    cos_sim, ops::CosSimKernel<paddle::platform::CPUDeviceContext, float>);
REGISTER_OP_CPU_KERNEL(
    cos_sim_grad,
    ops::CosSimGradKernel<paddle::platform::CPUDeviceContext, float>);