hierarchical_sigmoid_op.cc 10.2 KB
Newer Older
Y
Yancey1989 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* 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. */

W
weixing02 已提交
15
#include "paddle/fluid/operators/hierarchical_sigmoid_op.h"
16
#include <string>
W
weixing02 已提交
17
#include <vector>
Y
Yancey1989 已提交
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
namespace paddle {
namespace operators {

/**
 * Organize the classes into a binary tree. At each node, a sigmoid function
 * is used to calculate the probability of belonging to the right branch.
 * This idea is from "F. Morin, Y. Bengio (AISTATS 05):
 * Hierarchical Probabilistic Neural Network Language Model."
 *
 * Here we uses a simple way of making the binary tree.
 * Assuming the number of classes C = 6,
 * The classes are organized as a binary tree in the following way:
 *
 * @code{.py}
 * *-*-*- 2
 * | | |- 3
 * | |
 * | |-*- 4
 * |   |- 5
 * |
 * |-*- 0
 *   |- 1
 * @endcode
 *
 * where * indicates an internal node, and each leaf node represents a class.
 * - Node 0 ... C-2 are internal nodes.
 * - Node C-1 ... 2C-2 are leaf nodes.
 * - Class c is represented by leaf node \f$c+C-1\f$.
 *
 * We assign an id for each node:
 * - the id of root be 0.
 * - the left child of a node i is 2*i+1.
 * - the right child of a node i is 2*i+2.
 *
 * It's easy to see that:
 * - the parent of node i is \f$\left\lfloor(i-1)/2\right\rfloor\f$.
 * - the j-th level ancestor of node i is
 * \f$\left\lfloor(i+1)/2^{j+1}\right\rfloor - 1\f$.
 * - A node i is a left child of its parent if \f$(i-1)\%2==0\f$.
 *
 */

class HierarchicalSigmoidOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;
  void InferShape(framework::InferShapeContext* ctx) const override {
Y
Yancey1989 已提交
64
    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null.");
W
weixing02 已提交
65
    PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should not be null.");
Y
Yancey1989 已提交
66
    PADDLE_ENFORCE(ctx->HasInput("W"), "Input(W) should not be null.");
Y
Yancey1989 已提交
67
    PADDLE_ENFORCE(ctx->HasOutput("Out"), "Output(Out) should not be null.");
W
weixing02 已提交
68 69
    PADDLE_ENFORCE(ctx->HasOutput("PreOut"),
                   "Output(PreOut) should not be null.");
70 71 72 73 74
    auto with_prefetch = ctx->Attrs().Get<bool>("remote_prefetch");
    if (with_prefetch) {
      PADDLE_ENFORCE(ctx->HasOutput("W_Out"),
                     "Output(W_Out) should not be null.");
    }
Y
Yancey1989 已提交
75
    const int64_t batch_size = ctx->GetInputDim("X")[0];
Y
Yancey1989 已提交
76
    std::vector<int64_t> output_shape({batch_size, 1});
Y
Yancey1989 已提交
77
    ctx->SetOutputDim("Out", framework::make_ddim(output_shape));
J
JiabinYang 已提交
78
    ctx->ShareLoD("X", /*->*/ "Out");
Y
Yancey1989 已提交
79
  }
Y
Yancey1989 已提交
80 81

 protected:
W
weixing02 已提交
82
  framework::OpKernelType GetExpectedKernelType(
Y
Yancey1989 已提交
83
      const framework::ExecutionContext& ctx) const override {
Y
Yu Yang 已提交
84 85
    return framework::OpKernelType(ctx.Input<framework::LoDTensor>("X")->type(),
                                   ctx.GetPlace());
Y
Yancey1989 已提交
86
  }
Y
Yancey1989 已提交
87 88
};

W
weixing02 已提交
89
template <typename AttrType>
Y
Yancey1989 已提交
90 91
class HierarchicalSigmoidOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
W
weixing02 已提交
92
  void Make() override {
Y
Yancey1989 已提交
93
    AddInput("X",
J
JiabinYang 已提交
94
             "(LoDTensor, required) The input tensor with shape [N, D], "
G
guosheng 已提交
95
             "where N is the size of mini-batch, and D is the feature size.");
Y
Yancey1989 已提交
96
    AddInput("W",
J
JiabinYang 已提交
97
             "(LoDTensor, required), The parameters of hierarchical "
G
guosheng 已提交
98
             "sigmoid operator, each of them is a 2-D tensor, the shape is"
99
             "[K, D]. Which K is the num of non-leaf node in Path Tree");
W
weixing02 已提交
100
    AddInput("Label",
J
JiabinYang 已提交
101
             "(LoDTensor, required), The labels of training data. It's a"
G
guosheng 已提交
102
             "tensor with shape [N, 1].");
103
    AddInput("PathTable",
J
JiabinYang 已提交
104
             "(LoDTensor, optional), The Path Table from root to current word"
105 106
             "it should have shape like [N, L], L is the length of the Path")
        .AsDispensable();
J
JiabinYang 已提交
107
    AddInput(
J
JiabinYang 已提交
108
        "PathCode",
J
JiabinYang 已提交
109 110 111
        "(LoDTensor, optional), The Code on each Node of the Path from root "
        "to current word"
        "it should have shape like [N, L], L is the length of the Path")
112
        .AsDispensable();
Y
Yancey1989 已提交
113
    AddInput("Bias",
J
JiabinYang 已提交
114
             "(LoDTensor, optional), The bias is a tensor with shape or "
115
             "[num_classes, 1]"
116 117
             "[num_classes - 1, 1].")
        .AsDispensable();
J
JiabinYang 已提交
118 119 120 121
    AddOutput(
        "Out",
        "(LoDTensor, required) The output of hierarchical sigmoid operator."
        "The shape is [N, 1].");
W
weixing02 已提交
122
    AddOutput("PreOut",
J
JiabinYang 已提交
123
              "(LoDTensor, required) A intermedia 2-D tensor with shape "
G
guosheng 已提交
124 125
              "[batch_size, code_length], where code_length represents the "
              "maximum path length from root to leaf nodes.")
W
weixing02 已提交
126
        .AsIntermediate();
127 128 129 130 131
    AddOutput(
        "W_Out",
        "(LoDTensor, optinal) using input 'W' as Output to make it mutable"
        "When we are using prefetch")
        .AsIntermediate();
J
JiabinYang 已提交
132
    AddAttr<AttrType>("num_classes", "(int, optional), The number of classes")
Y
Yancey1989 已提交
133
        .SetDefault(2);
134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150
    // for parameter prefetch
    AddAttr<bool>("remote_prefetch", "").SetDefault(false);
    AddAttr<int>("trainer_id", "trainer id from 0 ~ worker_num.").SetDefault(0);
    AddAttr<std::vector<int>>("height_sections",
                              "Height for each output SelectedRows.")
        .SetDefault(std::vector<int>({}));
    AddAttr<std::vector<std::string>>(
        "epmap",
        "(string vector, default 127.0.0.1:6164)"
        "Server endpoints in the order of input variables for mapping")
        .SetDefault({});
    AddAttr<std::vector<std::string>>(
        "table_names",
        "(string vector, the splited table names that will be fetched from "
        "parameter server)"
        "in the order of input variables for mapping")
        .SetDefault({});
Y
Yancey1989 已提交
151 152
    AddComment(R"DOC(
The hierarchical sigmoid operator organize the classes into a binary tree.
W
weixing02 已提交
153
At each node, a sigmoid function is used to calculate the probability of
W
weixing02 已提交
154 155
belonging to the right branch. This idea is from
"F. Morin, Y. Bengio (AISTATS 05):
Y
Yancey1989 已提交
156 157
Hierarchical Probabilistic Neural Network Language Model."
      )DOC");
J
JiabinYang 已提交
158 159 160 161
    AddAttr<bool>("is_sparse",
                  "(boolean, default false) "
                  "Sparse update.")
        .SetDefault(false);
Y
Yancey1989 已提交
162 163 164
  }
};

W
weixing02 已提交
165 166 167 168 169
class HierarchicalSigmoidGradOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;
  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE(ctx->HasInput("W"), "Input(W) should not be null.");
W
weixing02 已提交
170
    PADDLE_ENFORCE(ctx->HasInput("Label"), "Input(Label) should not be null.");
J
JiabinYang 已提交
171 172
    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
                   "Input(Out@Grad) should not be null");
W
weixing02 已提交
173 174 175
    PADDLE_ENFORCE(ctx->HasInput("PreOut"),
                   "Input(Preout) should not be null.");
    PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("W")),
J
JiabinYang 已提交
176 177 178
                   "Output(W@Grad should not be null.");
    PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")),
                   "Output(X@Grad should not be null.");
179 180 181 182

    if (ctx->HasOutput(framework::GradVarName("Bias"))) {
      ctx->SetOutputDim(framework::GradVarName("Bias"),
                        ctx->GetInputDim("Bias"));
J
JiabinYang 已提交
183
    }
184
    ctx->SetOutputDim(framework::GradVarName("W"), ctx->GetInputDim("W"));
W
weixing02 已提交
185
    ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
J
JiabinYang 已提交
186
    ctx->ShareLoD("X", /*->*/ framework::GradVarName("X"));
W
weixing02 已提交
187 188 189 190 191
  }

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
Y
Yu Yang 已提交
192 193
    return framework::OpKernelType(ctx.Input<framework::LoDTensor>("X")->type(),
                                   ctx.GetPlace());
W
weixing02 已提交
194 195 196
  }
};

J
JiabinYang 已提交
197 198 199 200 201
class HierarchicalSigmoidGradOpGradVarTypeInference
    : public framework::VarTypeInference {
 public:
  void operator()(const framework::OpDesc& op_desc,
                  framework::BlockDesc* block) const override {
202 203 204 205 206 207 208 209 210 211
    auto w_grad_var_name = op_desc.Output(framework::GradVarName("W")).front();
    auto bias_grad_var_name_vec =
        op_desc.Output(framework::GradVarName("Bias"));
    std::string bias_grad_var_name;
    bool hasBias = false;
    if (bias_grad_var_name_vec.size()) {
      hasBias = true;
      bias_grad_var_name =
          op_desc.Output(framework::GradVarName("Bias")).front();
    }
J
JiabinYang 已提交
212 213 214
    auto attr = op_desc.GetAttr("is_sparse");
    bool is_sparse = boost::get<bool>(attr);
    if (is_sparse) {
215 216 217
      VLOG(30) << "hierarchical_sigmoid_grad op " << framework::GradVarName("W")
               << " is set to SelectedRows";
      block->Var(w_grad_var_name)
J
JiabinYang 已提交
218 219
          ->SetType(framework::proto::VarType::SELECTED_ROWS);
    } else {
220 221 222
      VLOG(30) << "hierarchical_sigmoid_grad op " << framework::GradVarName("W")
               << " is set to LoDTensor";
      block->Var(w_grad_var_name)
J
JiabinYang 已提交
223
          ->SetType(framework::proto::VarType::LOD_TENSOR);
224 225 226 227 228 229
    }
    if (hasBias) {
      VLOG(30) << "hierarchical_sigmoid_grad op "
               << framework::GradVarName("Bias") << " is set to LoDTensor";
      block->Var(bias_grad_var_name)
          ->SetType(framework::proto::VarType::LOD_TENSOR);
J
JiabinYang 已提交
230
    }
231
    block->Var(w_grad_var_name)->SetDataType(block->Var("W")->GetDataType());
J
JiabinYang 已提交
232 233 234
  }
};

Y
Yancey1989 已提交
235 236 237 238
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
W
weixing02 已提交
239 240 241
REGISTER_OPERATOR(hierarchical_sigmoid, ops::HierarchicalSigmoidOp,
                  ops::HierarchicalSigmoidOpMaker<int>,
                  paddle::framework::DefaultGradOpDescMaker<true>);
J
JiabinYang 已提交
242 243
REGISTER_OPERATOR(hierarchical_sigmoid_grad, ops::HierarchicalSigmoidGradOp,
                  ops::HierarchicalSigmoidGradOpGradVarTypeInference);
W
weixing02 已提交
244 245 246 247 248 249 250 251 252 253 254
REGISTER_OP_CPU_KERNEL(
    hierarchical_sigmoid,
    ops::HierarchicalSigmoidOpKernel<paddle::platform::CPUDeviceContext, float>,
    ops::HierarchicalSigmoidOpKernel<paddle::platform::CPUDeviceContext,
                                     double>);
REGISTER_OP_CPU_KERNEL(
    hierarchical_sigmoid_grad,
    ops::HierarchicalSigmoidGradOpKernel<paddle::platform::CPUDeviceContext,
                                         float>,
    ops::HierarchicalSigmoidGradOpKernel<paddle::platform::CPUDeviceContext,
                                         double>);