Skip to content

Commit a0af1ee

Browse files
guoshengCSguoshengCS
authored
Merge pull request #4443 from guoshengCS/add-GRUStepOp
Add gru_unit_op
2 parents 3b954e1 + ae1b29a commit a0af1ee

File tree

4 files changed

+577
-0
lines changed

4 files changed

+577
-0
lines changed

‎paddle/operators/gru_unit_op.cc‎

Lines changed: 210 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,210 @@
1+
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
2+
3+
Licensed under the Apache License, Version 2.0 (the "License");
4+
you may not use this file except in compliance with the License.
5+
You may obtain a copy of the License at
6+
7+
http://www.apache.org/licenses/LICENSE-2.0
8+
9+
Unless required by applicable law or agreed to in writing, software
10+
distributed under the License is distributed on an "AS IS" BASIS,
11+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12+
See the License for the specific language governing permissions and
13+
limitations under the License. */
14+
15+
#include "paddle/operators/gru_unit_op.h"
16+
17+
namespace paddle {
18+
namespace operators {
19+
20+
using framework::Tensor;
21+
22+
class GRUUnitOp : public framework::OperatorWithKernel {
23+
public:
24+
using framework::OperatorWithKernel::OperatorWithKernel;
25+
26+
protected:
27+
void InferShape(framework::InferShapeContext* ctx) const override {
28+
PADDLE_ENFORCE(ctx->HasInput("Input"),
29+
"Input(%s) of GRUUnitOp should not be null.", "Input");
30+
PADDLE_ENFORCE(ctx->HasInput("HiddenPrev"),
31+
"Input(%s) of GRUUnitOp should not be null.", "HiddenPrev");
32+
PADDLE_ENFORCE(ctx->HasInput("Weight"),
33+
"Input(%s) of GRUUnitOp should not be null.", "Weight");
34+
PADDLE_ENFORCE(ctx->HasOutput("Gate"),
35+
"Output(%s) of GRUUnitOp should not be null.", "Gate");
36+
PADDLE_ENFORCE(ctx->HasOutput("ResetHiddenPrev"),
37+
"Output(%s) of GRUUnitOp should not be null.",
38+
"ResetHiddenPrev");
39+
PADDLE_ENFORCE(ctx->HasOutput("Hidden"),
40+
"Output(%s) of GRUUnitOp should not be null.", "Hidden");
41+
auto input_dims = ctx->GetInputDim("Input");
42+
auto hidden_prev_dims = ctx->GetInputDim("HiddenPrev");
43+
auto weight_dims = ctx->GetInputDim("Weight");
44+
int batch_size = input_dims[0];
45+
int input_size = input_dims[1];
46+
int frame_size = hidden_prev_dims[1];
47+
int weight_height = weight_dims[0];
48+
int weight_width = weight_dims[1];
49+
PADDLE_ENFORCE_EQ(
50+
input_size, frame_size * 3,
51+
"The input_size must be 3 times of frame_size in GRUUnitOp.");
52+
PADDLE_ENFORCE_EQ(
53+
weight_height, frame_size,
54+
"The shape of Weight matrix must be [frame_size, frame_size * 3].");
55+
PADDLE_ENFORCE_EQ(
56+
weight_width, frame_size * 3,
57+
"The shape of Weight matrix must be [frame_size, frame_size * 3].");
58+
auto bias = Input("Bias");
59+
if (bias != framework::kEmptyVarName) {
60+
auto bias_dims = ctx->GetInputDim("Bias");
61+
int bias_height = bias_dims[0];
62+
int bias_width = bias_dims[1];
63+
PADDLE_ENFORCE_EQ(bias_height, 1,
64+
"The shape of Bias must be [1, frame_size * 3].");
65+
PADDLE_ENFORCE_EQ(bias_width, frame_size * 3,
66+
"The shape of Bias must be [1, frame_size * 3].");
67+
}
68+
ctx->SetOutputDim("Gate", {batch_size, frame_size * 3});
69+
ctx->SetOutputDim("ResetHiddenPrev", {batch_size, frame_size});
70+
ctx->SetOutputDim("Hidden", {batch_size, frame_size});
71+
}
72+
};
73+
74+
class GRUUnitOpMaker : public framework::OpProtoAndCheckerMaker {
75+
public:
76+
GRUUnitOpMaker(framework::OpProto* proto,
77+
framework::OpAttrChecker* op_checker)
78+
: OpProtoAndCheckerMaker(proto, op_checker) {
79+
AddInput("Input",
80+
"(Tensor) Matrix with shape [batch_size, frame_size * 3] for the "
81+
"input.");
82+
AddInput("HiddenPrev",
83+
"(Tensor) Matrix with shape [batch_size, frame_size] for the "
84+
"states of previous time step.");
85+
AddInput("Weight",
86+
"(Tensor) Weight matrix with shape [frame_size, frame_size * 3]. "
87+
"The elements continuous in memory can be divided into two parts. "
88+
"The first part are weights of the update gate and reset gate "
89+
"with shape [frame_size, frame_size * 2], and the second part are "
90+
"weights of output candidate with shape [frame_size, frame_size]");
91+
AddInput("Bias",
92+
"(Tensor) Bias vector with shape [1, frame_size * 3] concating "
93+
"bias of the update gate, reset gate and output candidate.");
94+
AddOutput("Gate",
95+
"(Tensor) Matrix with shape [batch_size, frame_size * 3] for the "
96+
"output of update gate, reset gate and output candidate")
97+
.AsIntermediate();
98+
AddOutput("ResetHiddenPrev",
99+
"(Tensor) Matrix with shape [batch_size, frame_size] for the "
100+
"reseted hidden state of previous time step.")
101+
.AsIntermediate();
102+
AddOutput("Hidden",
103+
"(Tensor) The GRU hidden state of the current time step "
104+
"with shape [batch_size, frame_size].");
105+
AddAttr<int>("activation",
106+
"(enum int, default tanh) "
107+
"The activation type used for output candidate {h}_t.")
108+
.SetDefault(tanh)
109+
.InEnum({identity, sigmoid, tanh, relu});
110+
AddAttr<int>("gate_activation",
111+
"(enum int, default sigmoid) "
112+
"The activation type used in update gate and reset gate.")
113+
.SetDefault(sigmoid)
114+
.InEnum({identity, sigmoid, tanh, relu});
115+
AddComment(R"DOC(
116+
GRUUnitOp implements part calculations of the GRU unit as following:
117+
118+
\f[
119+
update \ gate: u_t = actGate(xu_t + W_u * hidden_prev + bias_u) \\
120+
reset \ gate: r_t = actGate(xr_t + W_r * hidden_prev + bias_r) \\
121+
output \ candidate: {h}_t = actNode(xc_t + W_c * dot(r_t, hidden_prev) + bias_c) \\
122+
output: h_t = dot((1-u_t), {h}_t) + dot(u_t, hidden_prev)
123+
\f]
124+
125+
The rest of GRU unit can be completed by using FCOp's output as the input of GRUUnitOp.
126+
)DOC");
127+
}
128+
};
129+
130+
class GRUUnitGradOp : public framework::OperatorWithKernel {
131+
public:
132+
using framework::OperatorWithKernel::OperatorWithKernel;
133+
134+
protected:
135+
void InferShape(framework::InferShapeContext* ctx) const override {
136+
PADDLE_ENFORCE(ctx->HasInput("Input"),
137+
"Input(%s) of GRUUnitGradOp should not be null.", "Input");
138+
PADDLE_ENFORCE(ctx->HasInput("HiddenPrev"),
139+
"Input(%s) of GRUUnitGradOp should not be null.",
140+
"HiddenPrev");
141+
PADDLE_ENFORCE(ctx->HasInput("Weight"),
142+
"Input(%s) of GRUUnitGradOp should not be null.", "Weight");
143+
PADDLE_ENFORCE(ctx->HasInput("Gate"),
144+
"Input(%s) of GRUUnitGradOp should not be null.", "Gate");
145+
PADDLE_ENFORCE(ctx->HasInput("ResetHiddenPrev"),
146+
"Input(%s) of GRUUnitGradOp should not be null.",
147+
"ResetHiddenPrev");
148+
PADDLE_ENFORCE(ctx->HasInput("Hidden"),
149+
"Input(%s) of GRUUnitGradOp should not be null.", "Hidden");
150+
PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Gate")),
151+
"Input(%s@GRAD) of GRUUnitGradOp should not be null.",
152+
"Gate");
153+
PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("ResetHiddenPrev")),
154+
"Input(%s@GRAD) of GRUUnitGradOp should not be null.",
155+
"ResetHiddenPrev");
156+
PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Hidden")),
157+
"Input(%s@GRAD) of GRUUnitGradOp should not be null.",
158+
"Hidden");
159+
auto input_dims = ctx->GetInputDim("Input");
160+
auto hidden_prev_dims = ctx->GetInputDim("HiddenPrev");
161+
auto weight_dims = ctx->GetInputDim("Weight");
162+
// int batch_size = input_dims[0];
163+
int input_size = input_dims[1];
164+
int frame_size = hidden_prev_dims[1];
165+
int weight_height = weight_dims[0];
166+
int weight_width = weight_dims[1];
167+
PADDLE_ENFORCE_EQ(
168+
input_size, frame_size * 3,
169+
"The input_size must be 3 times of frame_size in GRUUnitOp.");
170+
PADDLE_ENFORCE_EQ(
171+
weight_height, frame_size,
172+
"The shape of Weight matrix must be [frame_size, frame_size * 3].");
173+
PADDLE_ENFORCE_EQ(
174+
weight_width, frame_size * 3,
175+
"The shape of Weight matrix must be [frame_size, frame_size * 3].");
176+
auto bias = Input("Bias");
177+
if (bias != framework::kEmptyVarName) {
178+
auto bias_dims = ctx->GetInputDim("Bias");
179+
int bias_height = bias_dims[0];
180+
int bias_width = bias_dims[1];
181+
PADDLE_ENFORCE_EQ(bias_height, 1,
182+
"The shape of Bias must be [1, frame_size * 3].");
183+
PADDLE_ENFORCE_EQ(bias_width, frame_size * 3,
184+
"The shape of Bias must be [1, frame_size * 3].");
185+
auto bias_grad_name = framework::GradVarName("Bias");
186+
if (ctx->HasOutput(bias_grad_name))
187+
ctx->SetOutputDim(bias_grad_name, bias_dims);
188+
}
189+
auto input_grad_name = framework::GradVarName("Input");
190+
if (ctx->HasOutput(input_grad_name))
191+
ctx->SetOutputDim(input_grad_name, input_dims);
192+
auto hidden_prev_grad_name = framework::GradVarName("HiddenPrev");
193+
if (ctx->HasOutput(hidden_prev_grad_name))
194+
ctx->SetOutputDim(hidden_prev_grad_name, hidden_prev_dims);
195+
auto weight_grad_name = framework::GradVarName("Weight");
196+
if (ctx->HasOutput(weight_grad_name))
197+
ctx->SetOutputDim(weight_grad_name, weight_dims);
198+
}
199+
};
200+
201+
} // namespace operators
202+
} // namespace paddle
203+
204+
namespace ops = paddle::operators;
205+
REGISTER_OP(gru_unit, ops::GRUUnitOp, ops::GRUUnitOpMaker, gru_unit_grad,
206+
ops::GRUUnitGradOp);
207+
REGISTER_OP_CPU_KERNEL(gru_unit,
208+
ops::GRUUnitKernel<paddle::platform::CPUPlace, float>);
209+
REGISTER_OP_CPU_KERNEL(
210+
gru_unit_grad, ops::GRUUnitGradKernel<paddle::platform::CPUPlace, float>);

‎paddle/operators/gru_unit_op.cu‎

Lines changed: 22 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,22 @@
1+
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
2+
3+
Licensed under the Apache License, Version 2.0 (the "License");
4+
you may not use this file except in compliance with the License.
5+
You may obtain a copy of the License at
6+
7+
http://www.apache.org/licenses/LICENSE-2.0
8+
9+
Unless required by applicable law or agreed to in writing, software
10+
distributed under the License is distributed on an "AS IS" BASIS,
11+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12+
See the License for the specific language governing permissions and
13+
limitations under the License. */
14+
15+
#define EIGEN_USE_GPU
16+
#include "paddle/operators/gru_unit_op.h"
17+
18+
namespace ops = paddle::operators;
19+
REGISTER_OP_GPU_KERNEL(gru_unit,
20+
ops::GRUUnitKernel<paddle::platform::GPUPlace, float>);
21+
REGISTER_OP_GPU_KERNEL(
22+
gru_unit_grad, ops::GRUUnitGradKernel<paddle::platform::GPUPlace, float>);

0 commit comments

Comments
 (0)