Skip to content
体验新版
项目
组织
正在加载...
登录
切换导航
打开侧边栏
PaddlePaddle
Paddle
提交
e9695f49
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2299
Star
20931
Fork
5422
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1423
列表
看板
标记
里程碑
合并请求
543
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
Paddle
项目概览
项目概览
详情
发布
仓库
仓库
文件
提交
分支
标签
贡献者
分支图
比较
Issue
1,423
Issue
1,423
列表
看板
标记
里程碑
合并请求
543
合并请求
543
Pages
分析
分析
仓库分析
DevOps
Wiki
0
Wiki
成员
成员
收起侧边栏
关闭侧边栏
动态
分支图
创建新Issue
提交
Issue看板
未验证
提交
e9695f49
编写于
11月 14, 2017
作者:
C
Cao Ying
提交者:
GitHub
11月 14, 2017
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #5014 from peterzhang2029/bi_tensor_prod_op
Add Bilinear Tensor Product operator.
上级
05c09084
c5d71077
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
406 addition
and
0 deletion
+406
-0
paddle/operators/bilinear_tensor_product_op.cc
paddle/operators/bilinear_tensor_product_op.cc
+159
-0
paddle/operators/bilinear_tensor_product_op.cu
paddle/operators/bilinear_tensor_product_op.cu
+26
-0
paddle/operators/bilinear_tensor_product_op.h
paddle/operators/bilinear_tensor_product_op.h
+184
-0
python/paddle/v2/framework/tests/test_bilinear_tensor_product_op.py
...dle/v2/framework/tests/test_bilinear_tensor_product_op.py
+37
-0
未找到文件。
paddle/operators/bilinear_tensor_product_op.cc
0 → 100644
浏览文件 @
e9695f49
/* 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/bilinear_tensor_product_op.h"
namespace
paddle
{
namespace
operators
{
using
framework
::
Tensor
;
class
BilinearTensorProductOp
:
public
framework
::
OperatorWithKernel
{
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
protected:
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"X"
),
"Input(X) should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Y"
),
"Input(Y) should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Weight"
),
"Input(Weight) should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
"Out"
),
"Output(Out) should not be null."
);
auto
x_dims
=
ctx
->
GetInputDim
(
"X"
);
auto
y_dims
=
ctx
->
GetInputDim
(
"Y"
);
auto
weight_dims
=
ctx
->
GetInputDim
(
"Weight"
);
PADDLE_ENFORCE_EQ
(
x_dims
.
size
(),
2UL
,
"The input(X) must be a 2D Tensor."
);
PADDLE_ENFORCE_EQ
(
y_dims
.
size
(),
2UL
,
"The input(Y) must be a 2D Tensor."
);
PADDLE_ENFORCE_EQ
(
weight_dims
.
size
(),
3UL
,
"The input(Weight) must be a 3D tensor."
);
PADDLE_ENFORCE_EQ
(
x_dims
[
0
],
y_dims
[
0
],
"The first dimension(batch_size) of input(X) must be "
"equal to the first dimension of the input(Y)."
);
PADDLE_ENFORCE_EQ
(
x_dims
[
1
],
weight_dims
[
1
],
"The second dimension of input(X) must be equal to "
"the second dimension of the input(Weight)."
);
PADDLE_ENFORCE_EQ
(
y_dims
[
1
],
weight_dims
[
2
],
"The second dimension of input(Y) must be equal to "
"the third dimension of the input(Weight)."
);
if
(
ctx
->
HasInput
(
"Bias"
))
{
auto
bias_dims
=
ctx
->
GetInputDim
(
"Bias"
);
PADDLE_ENFORCE
(
bias_dims
.
size
()
==
2UL
&&
bias_dims
[
0
]
==
1UL
,
"The Input(Bias) must be a 2-D tensor with "
"the 2nd dimension fixed to 1 (a row vector)."
);
PADDLE_ENFORCE_EQ
(
bias_dims
[
1
],
weight_dims
[
0
],
"The second dimension of input(Bias) must be equal "
"to the first dimension of the input(Weight)."
);
}
ctx
->
SetOutputDim
(
"Out"
,
{
x_dims
[
0
],
weight_dims
[
0
]});
ctx
->
ShareLoD
(
"X"
,
/*->*/
"Out"
);
}
};
class
BilinearTensorProductOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
BilinearTensorProductOpMaker
(
framework
::
OpProto
*
proto
,
framework
::
OpAttrChecker
*
op_checker
)
:
OpProtoAndCheckerMaker
(
proto
,
op_checker
)
{
AddInput
(
"X"
,
"The first input of bilinear_tensor_product operator."
);
AddInput
(
"Y"
,
"The second input of bilinear_tensor_product operator."
);
AddInput
(
"Weight"
,
"The learnable parameters of bilinear_tensor_product operator."
);
AddInput
(
"Bias"
,
"The learnable bias of bilinear_tensor_product operator."
)
.
AsDispensable
();
AddOutput
(
"Out"
,
"The output of bilinear_tensor_product operator."
);
AddComment
(
R"DOC(
Bilinear Tensor Product operator.
Given input X and Y, a 3D tensor weight, and bias. Each column of the
output is computed by one slice i = 1, . . . , k of the tensor:
M = (X W_i) \cdot Y
Out_i = \sum_i {M_i} + Bias_i
)DOC"
);
}
};
class
BilinearTensorProductOpGrad
:
public
framework
::
OperatorWithKernel
{
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
protected:
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"X"
),
"Input(X) should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Y"
),
"Input(Y) should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Weight"
),
"Input(Weight) should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
framework
::
GradVarName
(
"Out"
)),
"Input(Out@GRAD) should not be null."
);
auto
x_dims
=
ctx
->
GetInputDim
(
"X"
);
auto
y_dims
=
ctx
->
GetInputDim
(
"Y"
);
auto
weight_dims
=
ctx
->
GetInputDim
(
"Weight"
);
auto
out_dims
=
ctx
->
GetInputDim
(
framework
::
GradVarName
(
"Out"
));
PADDLE_ENFORCE_EQ
(
out_dims
.
size
(),
2UL
,
"The input(Out@GRAD) must be a 2D Tensor."
);
PADDLE_ENFORCE_EQ
(
x_dims
[
0
],
out_dims
[
0
],
"The first dimension(batch_size) of input(Out@GRAD) must be "
"equal to the first dimension of the Input(X)."
);
PADDLE_ENFORCE_EQ
(
weight_dims
[
0
],
out_dims
[
1
],
"The second dimension of input(Out@GRAD) must be equal to "
"the third dimension of the Input(Weight)."
);
if
(
ctx
->
HasInput
(
"Bias"
))
{
auto
bias_dims
=
ctx
->
GetInputDim
(
"Bias"
);
PADDLE_ENFORCE_EQ
(
bias_dims
[
1
],
out_dims
[
1
],
"The second dimension of input(Out@GRAD) must be equal to "
"the second dimension of the Input(Bias)."
);
auto
bias_grad_name
=
framework
::
GradVarName
(
"Bias"
);
if
(
ctx
->
HasOutput
(
bias_grad_name
))
ctx
->
SetOutputDim
(
bias_grad_name
,
bias_dims
);
}
auto
x_grad_name
=
framework
::
GradVarName
(
"X"
);
auto
y_grad_name
=
framework
::
GradVarName
(
"Y"
);
auto
weight_grad_name
=
framework
::
GradVarName
(
"Weight"
);
if
(
ctx
->
HasOutput
(
x_grad_name
))
{
ctx
->
SetOutputDim
(
x_grad_name
,
x_dims
);
}
if
(
ctx
->
HasOutput
(
y_grad_name
))
{
ctx
->
SetOutputDim
(
y_grad_name
,
y_dims
);
}
if
(
ctx
->
HasOutput
(
weight_grad_name
))
{
ctx
->
SetOutputDim
(
weight_grad_name
,
weight_dims
);
}
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OP
(
bilinear_tensor_product
,
ops
::
BilinearTensorProductOp
,
ops
::
BilinearTensorProductOpMaker
,
bilinear_tensor_product_grad
,
ops
::
BilinearTensorProductOpGrad
);
REGISTER_OP_CPU_KERNEL
(
bilinear_tensor_product
,
ops
::
BilinearTensorProductKernel
<
paddle
::
platform
::
CPUPlace
,
float
>
,
ops
::
BilinearTensorProductKernel
<
paddle
::
platform
::
CPUPlace
,
double
>
);
REGISTER_OP_CPU_KERNEL
(
bilinear_tensor_product_grad
,
ops
::
BilinearTensorProductGradKernel
<
paddle
::
platform
::
CPUPlace
,
float
>
,
ops
::
BilinearTensorProductGradKernel
<
paddle
::
platform
::
CPUPlace
,
double
>
);
paddle/operators/bilinear_tensor_product_op.cu
0 → 100644
浏览文件 @
e9695f49
/* 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. */
#define EIGEN_USE_GPU
#include "paddle/operators/bilinear_tensor_product_op.h"
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_GPU_KERNEL
(
bilinear_tensor_product
,
ops
::
BilinearTensorProductKernel
<
paddle
::
platform
::
GPUPlace
,
float
>
,
ops
::
BilinearTensorProductKernel
<
paddle
::
platform
::
GPUPlace
,
double
>
);
REGISTER_OP_GPU_KERNEL
(
bilinear_tensor_product_grad
,
ops
::
BilinearTensorProductGradKernel
<
paddle
::
platform
::
GPUPlace
,
float
>
,
ops
::
BilinearTensorProductGradKernel
<
paddle
::
platform
::
GPUPlace
,
double
>
);
paddle/operators/bilinear_tensor_product_op.h
0 → 100644
浏览文件 @
e9695f49
/* 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. */
#pragma once
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
#include "paddle/operators/math/math_function.h"
namespace
paddle
{
namespace
operators
{
using
framework
::
Tensor
;
template
<
typename
T
,
int
MajorType
=
Eigen
::
RowMajor
,
typename
IndexType
=
Eigen
::
DenseIndex
>
using
EigenMatrix
=
framework
::
EigenMatrix
<
T
,
MajorType
,
IndexType
>
;
template
<
typename
Place
,
typename
T
>
class
BilinearTensorProductKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
*
x
=
ctx
.
Input
<
Tensor
>
(
"X"
);
auto
*
y
=
ctx
.
Input
<
Tensor
>
(
"Y"
);
auto
*
weight
=
ctx
.
Input
<
Tensor
>
(
"Weight"
);
auto
*
bias
=
ctx
.
Input
<
Tensor
>
(
"Bias"
);
auto
*
out
=
ctx
.
Output
<
Tensor
>
(
"Out"
);
out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
y_mat
=
EigenMatrix
<
T
>::
From
(
*
y
);
auto
output_mat
=
EigenMatrix
<
T
>::
From
(
*
out
);
auto
batch_size
=
x
->
dims
()[
0
];
auto
weight_dims
=
weight
->
dims
();
int
out_dim
=
weight_dims
[
0
];
auto
x_dim
=
weight_dims
[
1
];
auto
y_dim
=
weight_dims
[
2
];
auto
place
=
ctx
.
GetEigenDevice
<
Place
>
();
// Create the intermediate variable to caculate the result of
// Input(X) multiplied by Input(Weight_i), the formula is:
// left_mul = X Weight_i.
Tensor
left_mul
;
left_mul
.
mutable_data
<
T
>
(
framework
::
make_ddim
({
batch_size
,
y_dim
}),
ctx
.
GetPlace
());
auto
left_mul_mat
=
EigenMatrix
<
T
>::
From
(
left_mul
);
for
(
int
i
=
0
;
i
<
out_dim
;
++
i
)
{
auto
output_col_vec
=
output_mat
.
chip
(
i
,
1
);
Tensor
weight_mat
=
weight
->
Slice
(
i
,
i
+
1
).
Resize
(
framework
::
make_ddim
({
x_dim
,
y_dim
}));
math
::
gemm
<
Place
,
T
>
(
ctx
.
device_context
(),
CblasNoTrans
,
CblasNoTrans
,
batch_size
,
y_dim
,
x_dim
,
1
,
x
->
data
<
T
>
(),
weight_mat
.
data
<
T
>
(),
0
,
left_mul
.
data
<
T
>
());
output_col_vec
.
device
(
place
)
=
(
left_mul_mat
*
y_mat
).
sum
(
Eigen
::
DSizes
<
int
,
1
>
(
1
));
}
if
(
bias
)
{
auto
bias_vec
=
EigenMatrix
<
T
>::
From
(
*
bias
);
Eigen
::
DSizes
<
int
,
2
>
bcast
(
batch_size
,
1
);
output_mat
.
device
(
place
)
=
bias_vec
.
broadcast
(
bcast
)
+
output_mat
;
}
}
};
template
<
typename
Place
,
typename
T
>
class
BilinearTensorProductGradKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
const
Tensor
*
x
=
ctx
.
Input
<
Tensor
>
(
"X"
);
const
Tensor
*
y
=
ctx
.
Input
<
Tensor
>
(
"Y"
);
const
Tensor
*
weight
=
ctx
.
Input
<
Tensor
>
(
"Weight"
);
Tensor
*
d_x
=
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"X"
));
Tensor
*
d_y
=
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"Y"
));
Tensor
*
d_weight
=
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"Weight"
));
Tensor
*
d_bias
=
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"Bias"
));
const
Tensor
*
d_out
=
ctx
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"Out"
));
auto
batch_size
=
x
->
dims
()[
0
];
auto
weight_dims
=
weight
->
dims
();
int
out_dim
=
weight_dims
[
0
];
auto
x_dim
=
weight_dims
[
1
];
auto
y_dim
=
weight_dims
[
2
];
auto
x_mat
=
EigenMatrix
<
T
>::
From
(
*
x
);
auto
y_mat
=
EigenMatrix
<
T
>::
From
(
*
y
);
auto
d_out_mat
=
EigenMatrix
<
T
>::
From
(
*
d_out
);
auto
place
=
ctx
.
GetEigenDevice
<
Place
>
();
// Create the intermediate variable to caculate the Output(Y@Grad).
Tensor
x_scale
;
x_scale
.
mutable_data
<
T
>
(
framework
::
make_ddim
({
batch_size
,
x_dim
}),
ctx
.
GetPlace
());
auto
x_scale_mat
=
EigenMatrix
<
T
>::
From
(
x_scale
);
// Create the intermediate variable to caculate the Output(X@Grad).
Tensor
y_scale
;
y_scale
.
mutable_data
<
T
>
(
framework
::
make_ddim
({
batch_size
,
y_dim
}),
ctx
.
GetPlace
());
auto
y_scale_mat
=
EigenMatrix
<
T
>::
From
(
y_scale
);
math
::
SetConstant
<
Place
,
T
>
set_zero
;
// Set Output(X@Grad) be zero.
if
(
d_x
)
{
d_x
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
set_zero
(
ctx
.
device_context
(),
d_x
,
static_cast
<
T
>
(
0
));
}
// Set Output(Y@Grad) be zero.
if
(
d_y
)
{
d_y
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
set_zero
(
ctx
.
device_context
(),
d_y
,
static_cast
<
T
>
(
0
));
}
// Caculate the Output(X@Grad) and Output(Y@Grad).
if
(
d_x
||
d_y
)
{
Eigen
::
DSizes
<
int
,
2
>
bcast_for_x
(
1
,
y_dim
);
Eigen
::
DSizes
<
int
,
2
>
bcast_for_y
(
1
,
x_dim
);
for
(
int
i
=
0
;
i
<
out_dim
;
++
i
)
{
Tensor
weight_i
=
weight
->
Slice
(
i
,
i
+
1
).
Resize
(
framework
::
make_ddim
({
x_dim
,
y_dim
}));
auto
output_vec
=
d_out_mat
.
chip
(
i
,
1
);
if
(
d_x
)
{
y_scale_mat
.
device
(
place
)
=
output_vec
.
reshape
(
Eigen
::
DSizes
<
int
,
2
>
(
batch_size
,
1
))
.
broadcast
(
bcast_for_x
)
*
y_mat
;
math
::
gemm
<
Place
,
T
>
(
ctx
.
device_context
(),
CblasNoTrans
,
CblasTrans
,
batch_size
,
x_dim
,
y_dim
,
1
,
y_scale
.
data
<
T
>
(),
weight_i
.
data
<
T
>
(),
1
,
d_x
->
data
<
T
>
());
}
if
(
d_y
)
{
x_scale_mat
.
device
(
place
)
=
output_vec
.
reshape
(
Eigen
::
DSizes
<
int
,
2
>
(
batch_size
,
1
))
.
broadcast
(
bcast_for_y
)
*
x_mat
;
math
::
gemm
<
Place
,
T
>
(
ctx
.
device_context
(),
CblasNoTrans
,
CblasNoTrans
,
batch_size
,
y_dim
,
x_dim
,
1
,
x_scale
.
data
<
T
>
(),
weight_i
.
data
<
T
>
(),
1
,
d_y
->
data
<
T
>
());
}
}
}
// Caculate the gradient of Input(Weight).
if
(
d_weight
)
{
d_weight
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
Eigen
::
DSizes
<
int
,
2
>
bcast_for_weight
(
1
,
x_dim
);
for
(
int
i
=
0
;
i
<
out_dim
;
++
i
)
{
Tensor
d_weight_i
=
d_weight
->
Slice
(
i
,
i
+
1
).
Resize
(
framework
::
make_ddim
({
x_dim
,
y_dim
}));
auto
output_vec
=
d_out_mat
.
chip
(
i
,
1
);
x_scale_mat
.
device
(
place
)
=
output_vec
.
reshape
(
Eigen
::
DSizes
<
int
,
2
>
(
batch_size
,
1
))
.
broadcast
(
bcast_for_weight
)
*
x_mat
;
math
::
gemm
<
Place
,
T
>
(
ctx
.
device_context
(),
CblasTrans
,
CblasNoTrans
,
x_dim
,
y_dim
,
batch_size
,
1
,
x_scale
.
data
<
T
>
(),
y
->
data
<
T
>
(),
0
,
d_weight_i
.
data
<
T
>
());
}
}
// Caculate the gradient of Input(Bias).
if
(
d_bias
)
{
d_bias
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
d_bias_mat
=
EigenMatrix
<
T
>::
From
(
*
d_bias
);
d_bias_mat
.
device
(
place
)
=
d_out_mat
.
sum
(
Eigen
::
DSizes
<
int
,
1
>
(
0
));
}
}
};
}
// namespace operators
}
// namespace paddle
python/paddle/v2/framework/tests/test_bilinear_tensor_product_op.py
0 → 100644
浏览文件 @
e9695f49
import
unittest
import
numpy
as
np
from
op_test
import
OpTest
class
TestBilinearTensorProductOp
(
OpTest
):
def
setUp
(
self
):
self
.
op_type
=
"bilinear_tensor_product"
batch_size
=
6
size0
=
3
size1
=
4
size2
=
5
a
=
np
.
random
.
random
((
batch_size
,
size0
)).
astype
(
"float32"
)
b
=
np
.
random
.
random
((
batch_size
,
size1
)).
astype
(
"float32"
)
w
=
np
.
random
.
random
((
size2
,
size0
,
size1
)).
astype
(
"float32"
)
bias
=
np
.
random
.
random
((
1
,
size2
)).
astype
(
"float32"
)
output
=
np
.
zeros
((
batch_size
,
size2
)).
astype
(
"float32"
)
for
i
in
range
(
size2
):
w_i
=
w
[
i
,
:,
:]
output
[:,
i
]
=
np
.
sum
(
np
.
matmul
(
a
,
w_i
)
*
b
,
axis
=
1
)
self
.
inputs
=
{
'X'
:
a
,
'Y'
:
b
,
'Weight'
:
w
,
'Bias'
:
bias
,
}
self
.
outputs
=
{
'Out'
:
output
+
bias
}
def
test_check_output
(
self
):
self
.
check_output
()
def
test_check_grad_normal
(
self
):
self
.
check_grad
([
'X'
,
'Y'
,
'Weight'
,
'Bias'
],
'Out'
)
if
__name__
==
"__main__"
:
unittest
.
main
()
编辑
预览
Markdown
is supported
0%
请重试
或
添加新附件
.
添加附件
取消
You are about to add
0
people
to the discussion. Proceed with caution.
先完成此消息的编辑!
取消
想要评论请
注册
或
登录