gru_unit_op.cc 9.3 KB
Newer Older
G
guosheng 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/gru_unit_op.h"

namespace paddle {
namespace operators {

using framework::Tensor;

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

26 27 28 29 30 31 32 33 34 35
  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE(ctx->HasInput("Input"),
                   "Input(%s) of GRUUnitOp should not be null.", "Input");
    PADDLE_ENFORCE(ctx->HasInput("HiddenPrev"),
                   "Input(%s) of GRUUnitOp should not be null.", "HiddenPrev");
    PADDLE_ENFORCE(ctx->HasInput("Weight"),
                   "Input(%s) of GRUUnitOp should not be null.", "Weight");
    PADDLE_ENFORCE(ctx->HasOutput("Gate"),
                   "Output(%s) of GRUUnitOp should not be null.", "Gate");
    PADDLE_ENFORCE(ctx->HasOutput("ResetHiddenPrev"),
G
guosheng 已提交
36
                   "Output(%s) of GRUUnitOp should not be null.",
37 38 39 40 41 42
                   "ResetHiddenPrev");
    PADDLE_ENFORCE(ctx->HasOutput("Hidden"),
                   "Output(%s) of GRUUnitOp should not be null.", "Hidden");
    auto input_dims = ctx->GetInputDim("Input");
    auto hidden_prev_dims = ctx->GetInputDim("HiddenPrev");
    auto weight_dims = ctx->GetInputDim("Weight");
G
guosheng 已提交
43 44 45 46 47 48 49
    int batch_size = input_dims[0];
    int input_size = input_dims[1];
    int frame_size = hidden_prev_dims[1];
    int weight_height = weight_dims[0];
    int weight_width = weight_dims[1];
    PADDLE_ENFORCE_EQ(
        input_size, frame_size * 3,
50
        "The input_size must be 3 times of frame_size in GRUUnitOp.");
G
guosheng 已提交
51 52
    PADDLE_ENFORCE_EQ(
        weight_height, frame_size,
53
        "The shape of Weight matrix must be [frame_size, frame_size * 3].");
G
guosheng 已提交
54 55
    PADDLE_ENFORCE_EQ(
        weight_width, frame_size * 3,
56
        "The shape of Weight matrix must be [frame_size, frame_size * 3].");
Y
Yang Yang(Tony) 已提交
57
    if (ctx->HasInput("Bias")) {
G
guosheng 已提交
58 59 60 61 62 63 64 65
      auto bias_dims = ctx->GetInputDim("Bias");
      int bias_height = bias_dims[0];
      int bias_width = bias_dims[1];
      PADDLE_ENFORCE_EQ(bias_height, 1,
                        "The shape of Bias must be [1, frame_size * 3].");
      PADDLE_ENFORCE_EQ(bias_width, frame_size * 3,
                        "The shape of Bias must be [1, frame_size * 3].");
    }
66 67 68
    ctx->SetOutputDim("Gate", {batch_size, frame_size * 3});
    ctx->SetOutputDim("ResetHiddenPrev", {batch_size, frame_size});
    ctx->SetOutputDim("Hidden", {batch_size, frame_size});
G
guosheng 已提交
69 70 71 72 73
  }
};

class GRUUnitOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
74 75
  GRUUnitOpMaker(framework::OpProto* proto,
                 framework::OpAttrChecker* op_checker)
G
guosheng 已提交
76
      : OpProtoAndCheckerMaker(proto, op_checker) {
77
    AddInput("Input",
G
guosheng 已提交
78 79
             "(Tensor) Matrix with shape [batch_size, frame_size * 3] for the "
             "input.");
80
    AddInput("HiddenPrev",
G
guosheng 已提交
81 82
             "(Tensor) Matrix with shape [batch_size, frame_size] for the "
             "states of previous time step.");
83
    AddInput("Weight",
G
guosheng 已提交
84 85 86 87 88
             "(Tensor) Weight matrix with shape [frame_size, frame_size * 3]. "
             "The elements continuous in memory can be divided into two parts. "
             "The first part are weights of the update gate and reset gate "
             "with shape [frame_size, frame_size * 2], and the second part are "
             "weights of output candidate with shape [frame_size, frame_size]");
89
    AddInput("Bias",
G
guosheng 已提交
90
             "(Tensor) Bias vector with shape [1, frame_size * 3] concating "
Y
Yang Yang(Tony) 已提交
91 92
             "bias of the update gate, reset gate and output candidate.")
        .AsDispensable();
93
    AddOutput("Gate",
G
guosheng 已提交
94 95 96
              "(Tensor) Matrix with shape [batch_size, frame_size * 3] for the "
              "output of update gate, reset gate and output candidate")
        .AsIntermediate();
97
    AddOutput("ResetHiddenPrev",
G
guosheng 已提交
98 99 100
              "(Tensor) Matrix with shape [batch_size, frame_size] for the "
              "reseted hidden state of previous time step.")
        .AsIntermediate();
101
    AddOutput("Hidden",
G
guosheng 已提交
102 103
              "(Tensor) The GRU hidden state of the current time step "
              "with shape [batch_size, frame_size].");
104 105 106 107 108 109 110 111 112 113
    AddAttr<int>("activation",
                 "(enum int, default tanh) "
                 "The activation type used for output candidate {h}_t.")
        .SetDefault(tanh)
        .InEnum({identity, sigmoid, tanh, relu});
    AddAttr<int>("gate_activation",
                 "(enum int, default sigmoid) "
                 "The activation type used in update gate and reset gate.")
        .SetDefault(sigmoid)
        .InEnum({identity, sigmoid, tanh, relu});
G
guosheng 已提交
114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132
    AddComment(R"DOC(
GRUUnitOp implements part calculations of the GRU unit as following:

\f[
update \ gate: u_t = actGate(xu_t + W_u * hidden_prev + bias_u) \\
reset \ gate: r_t = actGate(xr_t + W_r * hidden_prev + bias_r)  \\
output \ candidate: {h}_t = actNode(xc_t + W_c * dot(r_t, hidden_prev) + bias_c) \\
output: h_t = dot((1-u_t), {h}_t) + dot(u_t, hidden_prev)
\f]

The rest of GRU unit can be completed by using FCOp's output as the input of GRUUnitOp.
)DOC");
  }
};

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

133 134 135 136
  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE(ctx->HasInput("Input"),
                   "Input(%s) of GRUUnitGradOp should not be null.", "Input");
    PADDLE_ENFORCE(ctx->HasInput("HiddenPrev"),
G
guosheng 已提交
137
                   "Input(%s) of GRUUnitGradOp should not be null.",
138 139 140 141 142 143
                   "HiddenPrev");
    PADDLE_ENFORCE(ctx->HasInput("Weight"),
                   "Input(%s) of GRUUnitGradOp should not be null.", "Weight");
    PADDLE_ENFORCE(ctx->HasInput("Gate"),
                   "Input(%s) of GRUUnitGradOp should not be null.", "Gate");
    PADDLE_ENFORCE(ctx->HasInput("ResetHiddenPrev"),
G
guosheng 已提交
144
                   "Input(%s) of GRUUnitGradOp should not be null.",
145 146 147 148
                   "ResetHiddenPrev");
    PADDLE_ENFORCE(ctx->HasInput("Hidden"),
                   "Input(%s) of GRUUnitGradOp should not be null.", "Hidden");
    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Gate")),
G
guosheng 已提交
149
                   "Input(%s@GRAD) of GRUUnitGradOp should not be null.",
150 151
                   "Gate");
    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("ResetHiddenPrev")),
G
guosheng 已提交
152
                   "Input(%s@GRAD) of GRUUnitGradOp should not be null.",
153 154
                   "ResetHiddenPrev");
    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Hidden")),
G
guosheng 已提交
155
                   "Input(%s@GRAD) of GRUUnitGradOp should not be null.",
156 157 158 159
                   "Hidden");
    auto input_dims = ctx->GetInputDim("Input");
    auto hidden_prev_dims = ctx->GetInputDim("HiddenPrev");
    auto weight_dims = ctx->GetInputDim("Weight");
G
guosheng 已提交
160 161 162 163 164 165 166
    // int batch_size = input_dims[0];
    int input_size = input_dims[1];
    int frame_size = hidden_prev_dims[1];
    int weight_height = weight_dims[0];
    int weight_width = weight_dims[1];
    PADDLE_ENFORCE_EQ(
        input_size, frame_size * 3,
167
        "The input_size must be 3 times of frame_size in GRUUnitOp.");
G
guosheng 已提交
168 169
    PADDLE_ENFORCE_EQ(
        weight_height, frame_size,
170
        "The shape of Weight matrix must be [frame_size, frame_size * 3].");
G
guosheng 已提交
171 172
    PADDLE_ENFORCE_EQ(
        weight_width, frame_size * 3,
173
        "The shape of Weight matrix must be [frame_size, frame_size * 3].");
Y
Yu Yang 已提交
174
    if (ctx->HasInput("Bias")) {
G
guosheng 已提交
175 176 177 178 179 180 181 182 183 184 185
      auto bias_dims = ctx->GetInputDim("Bias");
      int bias_height = bias_dims[0];
      int bias_width = bias_dims[1];
      PADDLE_ENFORCE_EQ(bias_height, 1,
                        "The shape of Bias must be [1, frame_size * 3].");
      PADDLE_ENFORCE_EQ(bias_width, frame_size * 3,
                        "The shape of Bias must be [1, frame_size * 3].");
      auto bias_grad_name = framework::GradVarName("Bias");
      if (ctx->HasOutput(bias_grad_name))
        ctx->SetOutputDim(bias_grad_name, bias_dims);
    }
186
    auto input_grad_name = framework::GradVarName("Input");
G
guosheng 已提交
187 188
    if (ctx->HasOutput(input_grad_name))
      ctx->SetOutputDim(input_grad_name, input_dims);
189
    auto hidden_prev_grad_name = framework::GradVarName("HiddenPrev");
G
guosheng 已提交
190 191
    if (ctx->HasOutput(hidden_prev_grad_name))
      ctx->SetOutputDim(hidden_prev_grad_name, hidden_prev_dims);
192
    auto weight_grad_name = framework::GradVarName("Weight");
G
guosheng 已提交
193 194 195 196 197 198 199 200 201 202 203 204
    if (ctx->HasOutput(weight_grad_name))
      ctx->SetOutputDim(weight_grad_name, weight_dims);
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
REGISTER_OP(gru_unit, ops::GRUUnitOp, ops::GRUUnitOpMaker, gru_unit_grad,
            ops::GRUUnitGradOp);
REGISTER_OP_CPU_KERNEL(gru_unit,
Y
Yu Yang 已提交
205 206
                       ops::GRUUnitKernel<paddle::platform::CPUPlace, float>,
                       ops::GRUUnitKernel<paddle::platform::CPUPlace, double>);
G
guosheng 已提交
207
REGISTER_OP_CPU_KERNEL(
Y
Yu Yang 已提交
208 209
    gru_unit_grad, ops::GRUUnitGradKernel<paddle::platform::CPUPlace, float>,
    ops::GRUUnitGradKernel<paddle::platform::CPUPlace, double>);