gru_unit_op.cc 10.8 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
G
guosheng 已提交
2

L
Luo Tao 已提交
3 4 5
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
G
guosheng 已提交
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
G
guosheng 已提交
8

L
Luo Tao 已提交
9 10 11 12 13
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. */
G
guosheng 已提交
14

Y
Yi Wang 已提交
15
#include "paddle/fluid/operators/gru_unit_op.h"
H
hong 已提交
16
#include <memory>
G
guosheng 已提交
17 18 19 20 21 22 23 24 25 26

namespace paddle {
namespace operators {

using framework::Tensor;

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

27 28 29 30 31 32 33 34 35 36
  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
37
                   "Output(%s) of GRUUnitOp should not be null.",
38 39 40 41 42 43
                   "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
44 45 46 47 48 49 50
    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,
51
        "The input_size must be 3 times of frame_size in GRUUnitOp.");
G
guosheng 已提交
52 53
    PADDLE_ENFORCE_EQ(
        weight_height, frame_size,
54
        "The shape of Weight matrix must be [frame_size, frame_size * 3].");
G
guosheng 已提交
55 56
    PADDLE_ENFORCE_EQ(
        weight_width, frame_size * 3,
57
        "The shape of Weight matrix must be [frame_size, frame_size * 3].");
Y
Yang Yang(Tony) 已提交
58
    if (ctx->HasInput("Bias")) {
G
guosheng 已提交
59 60 61 62 63 64 65 66
      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].");
    }
67 68 69
    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
70 71 72 73 74
  }
};

class GRUUnitOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
Y
Yu Yang 已提交
75
  void Make() override {
76
    AddInput("Input",
G
guosheng 已提交
77 78
             "(Tensor) Matrix with shape [batch_size, frame_size * 3] for the "
             "input.");
79
    AddInput("HiddenPrev",
G
guosheng 已提交
80 81
             "(Tensor) Matrix with shape [batch_size, frame_size] for the "
             "states of previous time step.");
K
kexinzhao 已提交
82 83 84 85 86 87 88 89 90 91 92
    AddInput(
        "Weight",
        "(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].");
    AddInput(
        "Bias",
        "(Tensor) Bias vector with shape [1, frame_size * 3] concatenating "
        "bias of the update gate, reset gate and output candidate.")
Y
Yang Yang(Tony) 已提交
93
        .AsDispensable();
94
    AddOutput("Gate",
G
guosheng 已提交
95
              "(Tensor) Matrix with shape [batch_size, frame_size * 3] for the "
K
kexinzhao 已提交
96
              "output of update gate, reset gate and output candidate.")
G
guosheng 已提交
97
        .AsIntermediate();
98
    AddOutput("ResetHiddenPrev",
G
guosheng 已提交
99
              "(Tensor) Matrix with shape [batch_size, frame_size] for the "
T
tianshuo78520a 已提交
100
              "reset hidden state of previous time step.")
G
guosheng 已提交
101
        .AsIntermediate();
102
    AddOutput("Hidden",
G
guosheng 已提交
103 104
              "(Tensor) The GRU hidden state of the current time step "
              "with shape [batch_size, frame_size].");
105 106 107 108 109 110 111 112 113 114
    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});
Q
Qiao Longfei 已提交
115 116
    AddAttr<bool>("origin_mode",
                  "bool"
117 118 119 120
                  "use origin mode in article <Learning Phrase Representations "
                  "using RNN Encoder–Decoder\n"
                  "for Statistical Machine "
                  "Translation>(https://arxiv.org/pdf/1406.1078.pdf)")
Q
Qiao Longfei 已提交
121
        .SetDefault(false);
G
guosheng 已提交
122
    AddComment(R"DOC(
G
guosheng 已提交
123
GRUUnit Operator implements partial calculations of the GRU unit as following:
K
kexinzhao 已提交
124 125

$$
G
guosheng 已提交
126 127 128 129
update \ gate: u_t = actGate(xu_t + W_u * h_{t-1} + b_u) \\
reset \ gate: r_t = actGate(xr_t + W_r * h_{t-1} + b_r)  \\
output \ candidate: {h}_t = actNode(xc_t + W_c * dot(r_t, h_{t-1}) + b_c) \\
output: h_t = dot((1 - u_t), h_{t-1}) + dot(u_t, {h}_t)
K
kexinzhao 已提交
130
$$
G
guosheng 已提交
131

G
guosheng 已提交
132 133
which is same as one time step of GRU Operator.

134
@note To implement the complete GRU unit, fully-connected operator must be
G
guosheng 已提交
135
used before to feed xu, xr and xc as the Input of GRUUnit operator.
K
kexinzhao 已提交
136

G
guosheng 已提交
137 138 139 140 141 142 143 144
)DOC");
  }
};

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

145 146 147 148
  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
149
                   "Input(%s) of GRUUnitGradOp should not be null.",
150 151 152 153 154 155
                   "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
156
                   "Input(%s) of GRUUnitGradOp should not be null.",
157 158
                   "ResetHiddenPrev");
    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Hidden")),
G
guosheng 已提交
159
                   "Input(%s@GRAD) of GRUUnitGradOp should not be null.",
160 161 162 163
                   "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
164 165 166 167 168 169 170
    // 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,
171
        "The input_size must be 3 times of frame_size in GRUUnitOp.");
G
guosheng 已提交
172 173
    PADDLE_ENFORCE_EQ(
        weight_height, frame_size,
174
        "The shape of Weight matrix must be [frame_size, frame_size * 3].");
G
guosheng 已提交
175 176
    PADDLE_ENFORCE_EQ(
        weight_width, frame_size * 3,
177
        "The shape of Weight matrix must be [frame_size, frame_size * 3].");
Y
Yu Yang 已提交
178
    if (ctx->HasInput("Bias")) {
G
guosheng 已提交
179 180 181 182 183 184 185 186 187 188 189
      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);
    }
190
    auto input_grad_name = framework::GradVarName("Input");
G
guosheng 已提交
191 192
    if (ctx->HasOutput(input_grad_name))
      ctx->SetOutputDim(input_grad_name, input_dims);
193
    auto hidden_prev_grad_name = framework::GradVarName("HiddenPrev");
G
guosheng 已提交
194 195
    if (ctx->HasOutput(hidden_prev_grad_name))
      ctx->SetOutputDim(hidden_prev_grad_name, hidden_prev_dims);
196
    auto weight_grad_name = framework::GradVarName("Weight");
G
guosheng 已提交
197 198 199
    if (ctx->HasOutput(weight_grad_name))
      ctx->SetOutputDim(weight_grad_name, weight_dims);
  }
200 201 202 203 204 205 206

  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
    return framework::OpKernelType(OperatorWithKernel::IndicateVarDataType(
                                       ctx, framework::GradVarName("Hidden")),
                                   ctx.device_context());
  }
G
guosheng 已提交
207 208
};

H
hong 已提交
209 210
template <typename T>
class GRUUnitGradOpMaker : public framework::SingleGradOpMaker<T> {
211
 public:
H
hong 已提交
212
  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
213 214

 protected:
215
  void Apply(GradOpPtr<T> op) const override {
216 217
    op->SetType("gru_unit_grad");

H
hong 已提交
218 219 220 221
    op->SetInput("Input", this->Input("Input"));
    op->SetInput("HiddenPrev", this->Input("HiddenPrev"));
    op->SetInput("Weight", this->Input("Weight"));
    op->SetInput("Bias", this->Input("Bias"));
222

H
hong 已提交
223 224 225
    op->SetInput("Gate", this->Output("Gate"));
    op->SetInput("ResetHiddenPrev", this->Output("ResetHiddenPrev"));
    op->SetInput(framework::GradVarName("Hidden"), this->OutputGrad("Hidden"));
226

H
hong 已提交
227
    op->SetAttrMap(this->Attrs());
228

H
hong 已提交
229
    op->SetOutput(framework::GradVarName("Input"), this->InputGrad("Input"));
230
    op->SetOutput(framework::GradVarName("HiddenPrev"),
H
hong 已提交
231 232 233
                  this->InputGrad("HiddenPrev"));
    op->SetOutput(framework::GradVarName("Weight"), this->InputGrad("Weight"));
    op->SetOutput(framework::GradVarName("Bias"), this->InputGrad("Bias"));
234 235 236
  }
};

237 238 239
DECLARE_NO_NEED_BUFFER_VARS_INFERENCE(GRUUnitGradOpNoNeedBufferVarInference,
                                      "Bias");

G
guosheng 已提交
240 241 242 243
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
244

Y
Yang Yang 已提交
245
REGISTER_OPERATOR(gru_unit, ops::GRUUnitOp, ops::GRUUnitOpMaker,
H
hong 已提交
246 247
                  ops::GRUUnitGradOpMaker<paddle::framework::OpDesc>,
                  ops::GRUUnitGradOpMaker<paddle::imperative::OpBase>);
248 249
REGISTER_OPERATOR(gru_unit_grad, ops::GRUUnitGradOp,
                  ops::GRUUnitGradOpNoNeedBufferVarInference);
250

G
guosheng 已提交
251
REGISTER_OP_CPU_KERNEL(
Q
QI JUN 已提交
252 253 254 255 256 257
    gru_unit, ops::GRUUnitKernel<paddle::platform::CPUDeviceContext, float>,
    ops::GRUUnitKernel<paddle::platform::CPUDeviceContext, double>);
REGISTER_OP_CPU_KERNEL(
    gru_unit_grad,
    ops::GRUUnitGradKernel<paddle::platform::CPUDeviceContext, float>,
    ops::GRUUnitGradKernel<paddle::platform::CPUDeviceContext, double>);