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
PaddleDetection
提交
db1b128f
P
PaddleDetection
项目概览
PaddlePaddle
/
PaddleDetection
大约 1 年 前同步成功
通知
694
Star
11112
Fork
2696
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
184
列表
看板
标记
里程碑
合并请求
40
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
PaddleDetection
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
184
Issue
184
列表
看板
标记
里程碑
合并请求
40
合并请求
40
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
体验新版 GitCode,发现更多精彩内容 >>
提交
db1b128f
编写于
3月 26, 2018
作者:
D
dzhwinter
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
"add details"
上级
53c8c36a
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
114 addition
and
47 deletion
+114
-47
paddle/fluid/operators/sequence_expand_op.h
paddle/fluid/operators/sequence_expand_op.h
+114
-47
未找到文件。
paddle/fluid/operators/sequence_expand_op.h
浏览文件 @
db1b128f
...
...
@@ -13,15 +13,19 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <numeric> // std::itoa
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/memory/memcpy.h"
#include "paddle/fluid/
platform/device_context
.h"
#include "paddle/fluid/
operators/math/math_function
.h"
namespace
paddle
{
namespace
operators
{
using
LoDTensor
=
framework
::
LoDTensor
;
template
<
typename
T
,
int
MajorType
=
Eigen
::
RowMajor
,
typename
IndexType
=
Eigen
::
DenseIndex
>
using
EigenMatrix
=
framework
::
EigenMatrix
<
T
,
MajorType
,
IndexType
>
;
template
<
typename
DeviceContext
,
typename
T
>
struct
SequenceExpandFunctor
{
...
...
@@ -38,23 +42,35 @@ template <typename T>
struct
SequenceExpandFunctor
<
platform
::
CPUDeviceContext
,
T
>
{
void
operator
()(
const
platform
::
CPUDeviceContext
&
context
,
const
LoDTensor
&
x
,
LoDTensor
*
out
)
{
auto
x_dims
=
x
.
dims
();
size_t
element_len
=
framework
::
product
(
x_dims
)
/
x_dims
[
0
];
const
T
*
x_data
=
x
.
data
<
T
>
();
T
*
out_data
=
out
->
mutable_data
<
T
>
(
context
.
GetPlace
());
auto
out_starts
=
out
->
lod
().
back
();
for
(
size_t
i
=
0
;
i
<
out_starts
.
size
()
-
1
;
i
++
)
{
int
scale
=
out_starts
[
i
+
1
]
-
out_starts
[
i
];
Eigen
::
TensorMap
<
Eigen
::
Tensor
<
const
T
,
2
,
Eigen
::
RowMajor
,
Eigen
::
DenseIndex
>>
x_t
(
x_data
,
1
,
element_len
);
Eigen
::
TensorMap
<
Eigen
::
Tensor
<
T
,
2
,
Eigen
::
RowMajor
,
Eigen
::
DenseIndex
>>
out_t
(
out_data
,
scale
,
element_len
);
Eigen
::
array
<
int
,
2
>
cast
({{
scale
,
1
}});
out_t
.
device
(
*
context
.
eigen_device
())
=
x_t
.
broadcast
(
cast
);
x_data
+=
element_len
;
out_data
+=
element_len
*
scale
;
auto
&
out_lod
=
out
->
lod
()[
0
];
framework
::
Vector
<
size_t
>
x_lod
;
if
(
x
.
lod
()
==
1
)
{
x_lod
=
x
.
lod
()[
0
];
}
else
{
x_lod
.
reserve
(
out_lod
.
size
());
std
::
itoa
(
x_lod
.
begin
(),
x_lod
.
end
(),
0
);
// fill 0 ~ out_lod.size()-1
}
int
out_offset
=
0
;
auto
&
eigen_place
=
*
context
.
eigen_device
();
for
(
size_t
i
=
1
;
i
<
out_lod
.
size
();
++
i
)
{
int
repeat_num
=
y_lod
[
ref_level
][
i
]
-
y_lod
[
ref_level
][
i
-
1
];
int
x_start
=
x_lod
[
i
-
1
];
int
x_end
=
x_lod
[
i
];
int
x_seq_len
=
x_end
-
x_start
;
if
(
repeat_num
>
0
)
{
auto
x_sub_tensor
=
x
->
Slice
(
x_start
,
x_end
);
x_sub_tensor
.
Resize
({
1
,
x_sub_tensor
.
numel
()});
int
out_start
=
out_offset
;
if
(
x_lod
.
size
()
==
1
)
{
out_start
=
out_lod
[
0
][
out_offset
];
}
auto
out_sub_tensor
=
out
->
Slice
(
out_start
,
out_start
+
x_seq_len
*
repeat_num
);
out_sub_tensor
.
Resize
({
repeat_num
,
x_sub_tensor
.
dims
()[
1
]});
EigenMatrix
<
T
>::
From
(
out_sub_tensor
).
device
(
eigen_place
)
=
EigenMatrix
<
T
>::
From
(
x_sub_tensor
)
.
broadcast
(
Eigen
::
array
<
int
,
2
>
({{
repeat_num
,
1
}}));
}
}
}
};
...
...
@@ -64,15 +80,42 @@ class SequenceExpandKernel : public framework::OpKernel<T> {
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
auto
*
x
=
context
.
Input
<
LoDTensor
>
(
"X"
);
auto
*
out
=
context
.
Output
<
LoDTensor
>
(
"Out"
);
auto
x_dims
=
x
->
dims
();
auto
*
y
=
context
.
Input
<
LoDTensor
>
(
"Y"
);
PADDLE_ENFORCE
(
!
y
->
lod
().
empty
(),
"y should have lod"
);
PADDLE_ENFORCE_EQ
(
static_cast
<
size_t
>
(
x_dims
[
0
]),
y
->
lod
().
back
().
size
()
-
1
,
"The size of last lod level in Input(Y)"
"must be equal to dims[0] of Input(X)."
);
out
->
set_lod
(
y
->
lod
());
auto
*
out
=
context
.
Output
<
LoDTensor
>
(
"Out"
);
int
ref_level
=
context
.
Attr
<
int
>
(
"ref_level"
);
auto
&
x_lod
=
x
->
lod
();
auto
&
y_lod
=
y
->
lod
();
if
(
ref_level
==
-
1
)
ref_level
=
y_lod
.
size
()
-
1
;
out
->
mutable_data
<
T
>
(
context
.
GetPlace
());
if
(
y_lod
[
ref_level
].
size
()
<=
1
)
{
framework
::
TensorCopy
(
*
x
,
context
.
GetPlace
(),
out
);
return
;
}
auto
&
out_lod
=
*
out
->
mutable_lod
();
// x lod level is at most 1.
if
(
x_lod
.
size
()
==
0
)
{
out_lod
=
y_lod
[
ref_level
];
}
else
if
(
x_lod
.
size
()
==
1
)
{
out_lod
.
resize
(
1
);
out_lod
[
0
]
=
{
0
};
int
out_offset
=
0
;
for
(
size_t
i
=
1
;
i
<
y_lod
[
ref_level
].
size
();
++
i
)
{
int
repeat_num
=
y_lod
[
ref_level
][
i
]
-
y_lod
[
ref_level
][
i
-
1
];
int
x_start
=
x_lod
[
0
][
i
-
1
];
int
x_end
=
x_lod
[
0
][
i
];
int
x_seq_len
=
x_end
-
x_start
;
for
(
int
j
=
0
;
j
<
repeat_num
;
++
j
)
{
out_lod
[
0
].
push_back
(
out_lod
[
0
].
back
()
+
x_seq_len
);
out_offset
++
;
}
}
}
SequenceExpandFunctor
<
DeviceContext
,
T
>
functor
;
functor
(
context
.
template
device_context
<
DeviceContext
>(),
*
x
,
out
);
}
...
...
@@ -94,21 +137,31 @@ template <typename T>
struct
SequenceExpandGradFunctor
<
platform
::
CPUDeviceContext
,
T
>
{
void
operator
()(
const
platform
::
CPUDeviceContext
&
context
,
const
LoDTensor
&
x
,
const
LoDTensor
&
out
,
const
LoDTensor
&
dout
,
LoDTensor
*
dx
)
{
auto
out_last_level
=
out
.
lod
().
back
();
const
T
*
d_out_data
=
dout
.
data
<
T
>
();
T
*
d_x_data
=
dx
->
mutable_data
<
T
>
(
context
.
GetPlace
());
size_t
element_len
=
dout
.
numel
()
/
dout
.
dims
()[
0
];
for
(
size_t
i
=
0
;
i
<
out_last_level
.
size
()
-
1
;
++
i
)
{
size_t
repeat
=
out_last_level
[
i
+
1
]
-
out_last_level
[
i
];
Eigen
::
TensorMap
<
Eigen
::
Tensor
<
const
T
,
2
,
Eigen
::
RowMajor
,
Eigen
::
DenseIndex
>>
d_out_t
(
d_out_data
,
static_cast
<
int
>
(
repeat
),
element_len
);
Eigen
::
TensorMap
<
Eigen
::
Tensor
<
T
,
1
,
Eigen
::
RowMajor
,
Eigen
::
DenseIndex
>>
d_x_t
(
d_x_data
,
static_cast
<
int
>
(
element_len
));
d_x_t
.
device
(
*
context
.
eigen_device
())
=
d_out_t
.
sum
(
Eigen
::
array
<
int
,
1
>
({{
0
}}));
d_out_data
+=
(
repeat
*
element_len
);
d_x_data
+=
element_len
;
auto
&
dev_ctx
=
context
.
template
device_context
<
DeviceContext
>();
math
::
SetConstant
<
DeviceContext
,
T
>
set_zero
;
set_zero
(
dev_ctx
,
g_x
,
static_cast
<
T
>
(
0
));
int
g_out_offset
=
0
;
for
(
size_t
i
=
1
;
i
<
y_lod
[
ref_level
].
size
();
++
i
)
{
int
repeat_num
=
y_lod
[
ref_level
][
i
]
-
y_lod
[
ref_level
][
i
-
1
];
if
(
repeat_num
>
0
)
{
int
x_start
=
i
-
1
;
int
x_end
=
i
;
if
(
x_lod
.
size
()
==
1
)
{
x_start
=
x_lod
[
0
][
i
-
1
];
x_end
=
x_lod
[
0
][
i
];
}
int
x_seq_len
=
x_end
-
x_start
;
auto
g_x_sub
=
g_x
->
Slice
(
x_start
,
x_end
);
g_x_sub
.
Resize
(
flatten_to_1d
(
g_x_sub
.
dims
()));
int
g_out_end
=
g_out_offset
+
repeat_num
*
x_seq_len
;
auto
g_out_sub
=
g_out
->
Slice
(
g_out_offset
,
g_out_end
);
g_out_sub
.
Resize
({
repeat_num
,
g_x_sub
.
dims
()[
0
]});
math
::
ColwiseSum
<
DeviceContext
,
T
>
col_sum
;
col_sum
(
dev_ctx
,
g_out_sub
,
&
g_x_sub
);
g_out_offset
+=
repeat_num
*
x_seq_len
;
}
}
}
};
...
...
@@ -117,15 +170,29 @@ template <typename DeviceContext, typename T>
class
SequenceExpandGradKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
auto
*
g_out
=
context
.
Input
<
LoDTensor
>
(
framework
::
GradVarName
(
"Out"
));
auto
*
x
=
context
.
Input
<
LoDTensor
>
(
"X"
);
auto
*
out
=
context
.
Input
<
LoDTensor
>
(
"Out"
);
auto
*
d_out
=
context
.
Input
<
LoDTensor
>
(
framework
::
GradVarName
(
"Out"
));
auto
*
y
=
context
.
Input
<
LoDTensor
>
(
"Y"
);
auto
*
g_x
=
context
.
Output
<
LoDTensor
>
(
framework
::
GradVarName
(
"X"
));
int
ref_level
=
context
.
Attr
<
int
>
(
"ref_level"
);
g_x
->
mutable_data
<
T
>
(
context
.
GetPlace
());
g_x
->
set_lod
(
x
->
lod
());
auto
&
x_lod
=
x
->
lod
();
auto
&
y_lod
=
y
->
lod
();
if
(
ref_level
==
-
1
)
ref_level
=
y_lod
.
size
()
-
1
;
// just copy the gradient
if
(
y_lod
[
ref_level
].
size
()
<=
1
)
{
framework
::
TensorCopy
(
*
g_out
,
context
.
GetPlace
(),
g_x
);
return
;
}
auto
*
d_x
=
context
.
Output
<
LoDTensor
>
(
framework
::
GradVarName
(
"X"
));
d_x
->
set_lod
(
x
->
lod
());
SequenceExpandGradFunctor
<
DeviceContext
,
T
>
functor
;
functor
(
context
.
template
device_context
<
DeviceContext
>(),
*
x
,
*
out
,
*
d
_out
,
d
_x
);
functor
(
context
.
template
device_context
<
DeviceContext
>(),
*
x
,
*
y
,
*
g
_out
,
g
_x
);
}
};
...
...
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