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
783f9ead
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 2 年 前同步成功
通知
2325
Star
20933
Fork
5424
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
You need to sign in or sign up before continuing.
提交
783f9ead
编写于
1月 02, 2018
作者:
S
sweetsky0901
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
del using in .h
上级
e12d1a1c
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
47 addition
and
34 deletion
+47
-34
paddle/operators/norm_op.h
paddle/operators/norm_op.h
+47
-34
未找到文件。
paddle/operators/norm_op.h
浏览文件 @
783f9ead
...
@@ -19,13 +19,6 @@ limitations under the License. */
...
@@ -19,13 +19,6 @@ limitations under the License. */
namespace
paddle
{
namespace
paddle
{
namespace
operators
{
namespace
operators
{
template
<
typename
T
,
int
MajorType
=
Eigen
::
RowMajor
,
typename
IndexType
=
Eigen
::
DenseIndex
>
using
EigenVector
=
framework
::
EigenVector
<
T
,
MajorType
,
IndexType
>
;
template
<
typename
T
,
int
MajorType
=
Eigen
::
RowMajor
,
typename
IndexType
=
Eigen
::
DenseIndex
>
using
EigenMatrix
=
framework
::
EigenMatrix
<
T
,
MajorType
,
IndexType
>
;
template
<
typename
DeviceContext
,
typename
T
,
typename
AttrType
=
T
>
template
<
typename
DeviceContext
,
typename
T
,
typename
AttrType
=
T
>
class
NormKernel
:
public
framework
::
OpKernel
<
T
>
{
class
NormKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
public:
...
@@ -42,29 +35,37 @@ class NormKernel : public framework::OpKernel<T> {
...
@@ -42,29 +35,37 @@ class NormKernel : public framework::OpKernel<T> {
int
fea_len
=
height
*
width
;
int
fea_len
=
height
*
width
;
auto
*
place
=
auto
*
place
=
context
.
template
device_context
<
DeviceContext
>().
eigen_device
();
context
.
template
device_context
<
DeviceContext
>().
eigen_device
();
auto
x
=
EigenMatrix
<
T
>::
From
(
auto
x
=
*
in_x
,
framework
::
make_ddim
({
batch_size
,
fea_len
*
channels
}));
framework
::
EigenMatrix
<
T
,
Eigen
::
RowMajor
,
Eigen
::
DenseIndex
>::
From
(
*
in_x
,
framework
::
make_ddim
({
batch_size
,
fea_len
*
channels
}));
// get square
// get square
framework
::
Tensor
x_square
;
framework
::
Tensor
x_square
;
x_square
.
mutable_data
<
T
>
(
in_x
->
dims
(),
context
.
GetPlace
());
x_square
.
mutable_data
<
T
>
(
in_x
->
dims
(),
context
.
GetPlace
());
auto
x_square_eigen
=
EigenMatrix
<
T
>::
From
(
auto
x_square_eigen
=
x_square
,
framework
::
make_ddim
({
batch_size
,
fea_len
*
channels
}));
framework
::
EigenMatrix
<
T
,
Eigen
::
RowMajor
,
Eigen
::
DenseIndex
>::
From
(
x_square
,
framework
::
make_ddim
({
batch_size
,
fea_len
*
channels
}));
x_square_eigen
.
device
(
*
place
)
=
x
.
square
();
x_square_eigen
.
device
(
*
place
)
=
x
.
square
();
auto
scale_eigen
=
EigenVector
<
T
>::
Flatten
(
*
scale
);
auto
scale_eigen
=
framework
::
EigenVector
<
T
,
Eigen
::
RowMajor
,
Eigen
::
DenseIndex
>::
Flatten
(
*
scale
);
for
(
int
n
=
0
;
n
<
batch_size
;
++
n
)
{
for
(
int
n
=
0
;
n
<
batch_size
;
++
n
)
{
framework
::
Tensor
in_x_batch
=
in_x
->
Slice
(
n
,
n
+
1
);
framework
::
Tensor
in_x_batch
=
in_x
->
Slice
(
n
,
n
+
1
);
auto
in_x_batch_eigen
=
EigenMatrix
<
T
>::
From
(
auto
in_x_batch_eigen
=
in_x_batch
,
framework
::
make_ddim
({
channels
,
fea_len
}));
framework
::
EigenMatrix
<
T
,
Eigen
::
RowMajor
,
Eigen
::
DenseIndex
>::
From
(
in_x_batch
,
framework
::
make_ddim
({
channels
,
fea_len
}));
framework
::
Tensor
x_square_batch
=
x_square
.
Slice
(
n
,
n
+
1
);
framework
::
Tensor
x_square_batch
=
x_square
.
Slice
(
n
,
n
+
1
);
auto
x_square_batch_eigen
=
EigenMatrix
<
T
>::
From
(
auto
x_square_batch_eigen
=
x_square_batch
,
framework
::
make_ddim
({
channels
,
fea_len
}));
framework
::
EigenMatrix
<
T
,
Eigen
::
RowMajor
,
Eigen
::
DenseIndex
>::
From
(
x_square_batch
,
framework
::
make_ddim
({
channels
,
fea_len
}));
framework
::
Tensor
out_batch
=
out
->
Slice
(
n
,
n
+
1
);
framework
::
Tensor
out_batch
=
out
->
Slice
(
n
,
n
+
1
);
auto
out_batch_eigen
=
EigenMatrix
<
T
>::
From
(
auto
out_batch_eigen
=
out_batch
,
framework
::
make_ddim
({
channels
,
fea_len
}));
framework
::
EigenMatrix
<
T
,
Eigen
::
RowMajor
,
Eigen
::
DenseIndex
>::
From
(
out_batch
,
framework
::
make_ddim
({
channels
,
fea_len
}));
framework
::
Tensor
tmp_tensor
;
framework
::
Tensor
tmp_tensor
;
tmp_tensor
.
mutable_data
<
T
>
(
framework
::
make_ddim
({
1
,
fea_len
}),
tmp_tensor
.
mutable_data
<
T
>
(
framework
::
make_ddim
({
1
,
fea_len
}),
context
.
GetPlace
());
context
.
GetPlace
());
auto
tmp
=
EigenVector
<
T
>::
Flatten
(
tmp_tensor
);
auto
tmp
=
framework
::
EigenVector
<
T
,
Eigen
::
RowMajor
,
Eigen
::
DenseIndex
>::
Flatten
(
tmp_tensor
);
// get colsum and sqrt , inverse
// get colsum and sqrt , inverse
auto
dim
=
Eigen
::
array
<
int
,
1
>
({{
0
}});
auto
dim
=
Eigen
::
array
<
int
,
1
>
({{
0
}});
tmp
.
device
(
*
place
)
=
x_square_batch_eigen
.
sum
(
dim
);
tmp
.
device
(
*
place
)
=
x_square_batch_eigen
.
sum
(
dim
);
...
@@ -102,40 +103,52 @@ class NormGradKernel : public framework::OpKernel<T> {
...
@@ -102,40 +103,52 @@ class NormGradKernel : public framework::OpKernel<T> {
auto
*
place
=
auto
*
place
=
context
.
template
device_context
<
DeviceContext
>().
eigen_device
();
context
.
template
device_context
<
DeviceContext
>().
eigen_device
();
auto
scale_eigen
=
EigenVector
<
T
>::
Flatten
(
*
scale
);
auto
scale_eigen
=
auto
x
=
EigenMatrix
<
T
>::
From
(
framework
::
EigenVector
<
T
,
Eigen
::
RowMajor
,
Eigen
::
DenseIndex
>::
Flatten
(
*
in_x
,
framework
::
make_ddim
({
batch_size
,
fea_len
*
channels
}));
*
scale
);
auto
x
=
framework
::
EigenMatrix
<
T
,
Eigen
::
RowMajor
,
Eigen
::
DenseIndex
>::
From
(
*
in_x
,
framework
::
make_ddim
({
batch_size
,
fea_len
*
channels
}));
// get square
// get square
framework
::
Tensor
x_square
;
framework
::
Tensor
x_square
;
x_square
.
mutable_data
<
T
>
(
in_x
->
dims
(),
context
.
GetPlace
());
x_square
.
mutable_data
<
T
>
(
in_x
->
dims
(),
context
.
GetPlace
());
auto
x_square_eigen
=
EigenMatrix
<
T
>::
From
(
auto
x_square_eigen
=
x_square
,
framework
::
make_ddim
({
batch_size
,
fea_len
*
channels
}));
framework
::
EigenMatrix
<
T
,
Eigen
::
RowMajor
,
Eigen
::
DenseIndex
>::
From
(
x_square
,
framework
::
make_ddim
({
batch_size
,
fea_len
*
channels
}));
x_square_eigen
.
device
(
*
place
)
=
x
.
square
();
x_square_eigen
.
device
(
*
place
)
=
x
.
square
();
for
(
int
n
=
0
;
n
<
batch_size
;
++
n
)
{
for
(
int
n
=
0
;
n
<
batch_size
;
++
n
)
{
framework
::
Tensor
in_x_batch
=
in_x
->
Slice
(
n
,
n
+
1
);
framework
::
Tensor
in_x_batch
=
in_x
->
Slice
(
n
,
n
+
1
);
auto
in_x_batch_eigen
=
EigenMatrix
<
T
>::
From
(
auto
in_x_batch_eigen
=
in_x_batch
,
framework
::
make_ddim
({
channels
,
fea_len
}));
framework
::
EigenMatrix
<
T
,
Eigen
::
RowMajor
,
Eigen
::
DenseIndex
>::
From
(
in_x_batch
,
framework
::
make_ddim
({
channels
,
fea_len
}));
framework
::
Tensor
in_g_batch
=
in_x_grad
->
Slice
(
n
,
n
+
1
);
framework
::
Tensor
in_g_batch
=
in_x_grad
->
Slice
(
n
,
n
+
1
);
auto
in_g_batch_eigen
=
EigenMatrix
<
T
>::
From
(
auto
in_g_batch_eigen
=
in_g_batch
,
framework
::
make_ddim
({
channels
,
fea_len
}));
framework
::
EigenMatrix
<
T
,
Eigen
::
RowMajor
,
Eigen
::
DenseIndex
>::
From
(
in_g_batch
,
framework
::
make_ddim
({
channels
,
fea_len
}));
framework
::
Tensor
x_square_batch
=
x_square
.
Slice
(
n
,
n
+
1
);
framework
::
Tensor
x_square_batch
=
x_square
.
Slice
(
n
,
n
+
1
);
auto
x_square_batch_eigen
=
EigenMatrix
<
T
>::
From
(
auto
x_square_batch_eigen
=
x_square_batch
,
framework
::
make_ddim
({
channels
,
fea_len
}));
framework
::
EigenMatrix
<
T
,
Eigen
::
RowMajor
,
Eigen
::
DenseIndex
>::
From
(
x_square_batch
,
framework
::
make_ddim
({
channels
,
fea_len
}));
framework
::
Tensor
outg_batch
=
out_grad
->
Slice
(
n
,
n
+
1
);
framework
::
Tensor
outg_batch
=
out_grad
->
Slice
(
n
,
n
+
1
);
auto
outg_batch_eigen
=
EigenMatrix
<
T
>::
From
(
auto
outg_batch_eigen
=
outg_batch
,
framework
::
make_ddim
({
channels
,
fea_len
}));
framework
::
EigenMatrix
<
T
,
Eigen
::
RowMajor
,
Eigen
::
DenseIndex
>::
From
(
outg_batch
,
framework
::
make_ddim
({
channels
,
fea_len
}));
framework
::
Tensor
tmp_tensor
;
framework
::
Tensor
tmp_tensor
;
tmp_tensor
.
mutable_data
<
T
>
(
framework
::
make_ddim
({
1
,
fea_len
}),
tmp_tensor
.
mutable_data
<
T
>
(
framework
::
make_ddim
({
1
,
fea_len
}),
context
.
GetPlace
());
context
.
GetPlace
());
auto
tmp_eigen
=
EigenVector
<
T
>::
Flatten
(
tmp_tensor
);
auto
tmp_eigen
=
framework
::
EigenVector
<
T
,
Eigen
::
RowMajor
,
Eigen
::
DenseIndex
>::
Flatten
(
tmp_tensor
);
auto
dim
=
Eigen
::
array
<
int
,
1
>
({{
0
}});
auto
dim
=
Eigen
::
array
<
int
,
1
>
({{
0
}});
tmp_eigen
.
device
(
*
place
)
=
(
in_x_batch_eigen
*
outg_batch_eigen
).
sum
(
dim
);
tmp_eigen
.
device
(
*
place
)
=
(
in_x_batch_eigen
*
outg_batch_eigen
).
sum
(
dim
);
framework
::
Tensor
norm_tmp_tensor
;
framework
::
Tensor
norm_tmp_tensor
;
norm_tmp_tensor
.
mutable_data
<
T
>
(
framework
::
make_ddim
({
1
,
fea_len
}),
norm_tmp_tensor
.
mutable_data
<
T
>
(
framework
::
make_ddim
({
1
,
fea_len
}),
context
.
GetPlace
());
context
.
GetPlace
());
auto
norm_tmp_eigen
=
EigenVector
<
T
>::
Flatten
(
norm_tmp_tensor
);
auto
norm_tmp_eigen
=
framework
::
EigenVector
<
T
,
Eigen
::
RowMajor
,
Eigen
::
DenseIndex
>::
Flatten
(
norm_tmp_tensor
);
norm_tmp_eigen
.
device
(
*
place
)
=
norm_tmp_eigen
.
device
(
*
place
)
=
(
x_square_batch_eigen
.
sum
(
dim
)
+
epsilon
).
sqrt
();
(
x_square_batch_eigen
.
sum
(
dim
)
+
epsilon
).
sqrt
();
Eigen
::
array
<
int
,
2
>
broadcast_dim_col
;
Eigen
::
array
<
int
,
2
>
broadcast_dim_col
;
...
...
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