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
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
/* Copyright (c) 2019 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/fluid/operators/cvm_op.h"
#include <memory>
#include "paddle/fluid/operators/math/math_function.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
class CVMOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
PADDLE_ENFORCE(ctx->HasInput("CVM"), "Input(CVM) should be not null.");
PADDLE_ENFORCE(ctx->HasOutput("Y"), "Output(Y) should be not null.");
auto x_dims = ctx->GetInputDim("X");
auto cvm_dims = ctx->GetInputDim("CVM");
PADDLE_ENFORCE_EQ(x_dims.size(), 2UL, "Input(X)'s rank should be 2.");
PADDLE_ENFORCE_EQ(cvm_dims.size(), 2UL, "Input(CVM)'s rank should be 2.");
PADDLE_ENFORCE_EQ(cvm_dims[1], 2UL,
"The 2nd dimension of "
"Input(CVM) should be 2.");
if (ctx->Attrs().Get<bool>("use_cvm")) {
ctx->SetOutputDim("Y", {x_dims[0], x_dims[1]});
} else {
ctx->SetOutputDim("Y", {x_dims[0], x_dims[1] - 2});
}
ctx->ShareLoD("X", /*->*/ "Y");
}
protected:
// Explicitly set that the data type of computation kernel of
// cvm
// is determined by its input "X".
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
OperatorWithKernel::IndicateVarDataType(ctx, "X"),
ctx.device_context());
}
};
class CVMGradientOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override {
PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should be not null.");
PADDLE_ENFORCE(ctx->HasInput("CVM"), "Input(CVM) should be not null.");
PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")),
"Input(Y@GRAD) should be not null.");
PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")),
"Output(X@GRAD) should be not null.");
auto x_dims = ctx->GetInputDim("X");
auto cvm_dims = ctx->GetInputDim("CVM");
auto dy_dims = ctx->GetInputDim(framework::GradVarName("Y"));
PADDLE_ENFORCE_EQ(x_dims.size(), 2, "Input(X)'s rank should be 2.");
PADDLE_ENFORCE_EQ(dy_dims.size(), 2, "Input(Y@Grad)'s rank should be 2.");
PADDLE_ENFORCE_EQ(cvm_dims.size(), 2, "Input(CVM)'s rank should be 2.");
PADDLE_ENFORCE_EQ(x_dims[0], dy_dims[0],
"The 1st dimension of Input(X) and Input(Y@Grad) should "
"be equal.");
PADDLE_ENFORCE_EQ(cvm_dims[1], 2,
"When Attr(soft_label) == false, the 2nd dimension of "
"Input(CVM) should be 2.");
ctx->SetOutputDim(framework::GradVarName("X"), x_dims);
ctx->ShareLoD("X", framework::GradVarName("X"));
}
protected:
// Explicitly set that the data type of computation kernel of
// cvm
// is determined by its input "X".
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override {
return framework::OpKernelType(
OperatorWithKernel::IndicateVarDataType(ctx, "X"),
ctx.device_context());
}
};
class CVMOpMaker : public framework::OpProtoAndCheckerMaker {
public:
void Make() override {
AddInput("X",
"(LodTensor, default LodTensor<float>), a 2-D tensor with shape "
"[N x D],"
" where N is the batch size and D is the emebdding dim. ");
AddInput("CVM",
"(Tensor), a 2-D Tensor with shape [N x 2], where N is the batch "
"size, 2 is show and click.");
AddOutput("Y",
"(LodTensor, default LodTensor<float>), a 2-D tensor with shape "
"[N x K].");
AddAttr<bool>("use_cvm", "bool, use cvm or not").SetDefault(true);
AddComment(R"DOC(
CVM Operator.
We assume that input X is a embedding vector with cvm_feature(show and click), which shape is [N * D] (D is 2(cvm_feature) + embedding dim, N is batch_size)
if use_cvm is True, we will log(cvm_feature), and output shape is [N * D].
if use_cvm is False, we will remove cvm_feature from input, and output shape is [N * (D - 2)].
)DOC");
}
};
template <typename T>
class CVMGradOpMaker : public framework::SingleGradOpMaker<T> {
public:
using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
protected:
std::unique_ptr<T> Apply() const override {
std::unique_ptr<T> op(new T());
op->SetType("cvm_grad");
op->SetInput("X", this->Input("X"));
op->SetInput("CVM", this->Input("CVM"));
op->SetInput(framework::GradVarName("Y"), this->OutputGrad("Y"));
op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
op->SetAttrMap(this->Attrs());
return op;
}
};
} // namespace operators
} // namespace paddle
namespace ops = paddle::operators;
REGISTER_OPERATOR(cvm, ops::CVMOp, ops::CVMOpMaker,
ops::CVMGradOpMaker<paddle::framework::OpDesc>,
ops::CVMGradOpMaker<paddle::imperative::OpBase>);
REGISTER_OPERATOR(cvm_grad, ops::CVMGradientOp);
REGISTER_OP_CPU_KERNEL(cvm, ops::CVMOpKernel<float>, ops::CVMOpKernel<double>);
REGISTER_OP_CPU_KERNEL(cvm_grad, ops::CVMGradOpKernel<float>,
ops::CVMGradOpKernel<double>);