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
兽拳
Paddle
提交
94cee3d6
P
Paddle
项目概览
兽拳
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
1
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1
列表
看板
标记
里程碑
合并请求
0
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
Issue
1
列表
看板
标记
里程碑
合并请求
0
合并请求
0
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
94cee3d6
编写于
8月 03, 2017
作者:
Z
Zhaolong Xing
提交者:
GitHub
8月 03, 2017
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #3163 from NHZlX/fix_conv_1x1
ignore im2col if not necessary in conv 1 * 1
上级
28db1491
fa10677a
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
90 addition
and
48 deletion
+90
-48
paddle/function/ConvOp.h
paddle/function/ConvOp.h
+7
-0
paddle/function/GemmConvOp.cpp
paddle/function/GemmConvOp.cpp
+83
-48
未找到文件。
paddle/function/ConvOp.h
浏览文件 @
94cee3d6
...
...
@@ -109,6 +109,13 @@ protected:
return
filter
[
filter
.
ndims
()
-
1
];
}
// determine whether im2col needs to be performed
inline
bool
isNeedIm2col
(
const
TensorShape
&
filter
)
const
{
return
!
(
getFilterHeight
(
filter
)
==
1
&&
getFilterWidth
(
filter
)
==
1
&&
strideH
()
==
1
&&
strideW
()
==
1
&&
paddingH
()
==
0
&&
paddingW
()
==
0
);
}
std
::
vector
<
size_t
>
strides_
;
std
::
vector
<
size_t
>
paddings_
;
...
...
paddle/function/GemmConvOp.cpp
浏览文件 @
94cee3d6
...
...
@@ -66,16 +66,23 @@ public:
real
*
inputData
=
inputs
[
0
].
data
<
real
>
();
real
*
filterData
=
inputs
[
1
].
data
<
real
>
();
real
*
outputData
=
outputs
[
0
].
data
<
real
>
();
bool
needIm2col
=
isNeedIm2col
(
filter
);
TensorShape
imShape
=
TensorShape
({
inputChannels
/
groups_
,
inputHeight
,
inputWidth
});
TensorShape
colShape
=
TensorShape
({
inputChannels
/
groups_
,
filterHeight
,
filterWidth
,
outputHeight
,
outputWidth
});
resizeBuffer
<
Device
>
(
colShape
.
getElements
());
real
*
colData
=
reinterpret_cast
<
real
*>
(
memory_
->
getBuf
());
TensorShape
colShape
;
real
*
colData
=
NULL
;
if
(
needIm2col
)
{
colShape
=
TensorShape
({
inputChannels
/
groups_
,
filterHeight
,
filterWidth
,
outputHeight
,
outputWidth
});
resizeBuffer
<
Device
>
(
colShape
.
getElements
());
colData
=
reinterpret_cast
<
real
*>
(
memory_
->
getBuf
());
}
Im2ColFunctor
<
kCFO
,
Device
,
real
>
im2col
;
GemmFunctor
<
Device
,
real
>
gemm
;
...
...
@@ -86,15 +93,18 @@ public:
for
(
size_t
i
=
0
;
i
<
batchSize
;
i
++
)
{
for
(
size_t
g
=
0
;
g
<
groups_
;
g
++
)
{
im2col
(
inputData
+
g
*
inputOffset
,
imShape
,
colData
,
colShape
,
strideH
(),
strideW
(),
paddingH
(),
paddingW
());
if
(
needIm2col
)
{
im2col
(
inputData
+
g
*
inputOffset
,
imShape
,
colData
,
colShape
,
strideH
(),
strideW
(),
paddingH
(),
paddingW
());
}
else
{
colData
=
inputData
+
g
*
inputOffset
;
}
int
M
=
outputChannels
/
groups_
;
int
N
=
outputHeight
*
outputWidth
;
int
K
=
inputChannels
/
groups_
*
filterHeight
*
filterWidth
;
...
...
@@ -159,19 +169,27 @@ public:
real
*
outputGrad
=
inputs
[
0
].
data
<
real
>
();
real
*
filterData
=
inputs
[
1
].
data
<
real
>
();
real
*
inputGrad
=
outputs
[
0
].
data
<
real
>
();
bool
needIm2col
=
isNeedIm2col
(
filter
);
TensorShape
imShape
=
TensorShape
({
inputChannels
/
groups_
,
inputHeight
,
inputWidth
});
TensorShape
colShape
=
TensorShape
({
inputChannels
/
groups_
,
filterHeight
,
filterWidth
,
outputHeight
,
outputWidth
});
resizeBuffer
<
Device
>
(
colShape
.
getElements
());
real
*
colData
=
reinterpret_cast
<
real
*>
(
memory_
->
getBuf
());
TensorShape
colShape
;
real
*
colData
=
NULL
;
if
(
needIm2col
)
{
colShape
=
TensorShape
({
inputChannels
/
groups_
,
filterHeight
,
filterWidth
,
outputHeight
,
outputWidth
});
resizeBuffer
<
Device
>
(
colShape
.
getElements
());
colData
=
reinterpret_cast
<
real
*>
(
memory_
->
getBuf
());
}
Col2ImFunctor
<
kCFO
,
Device
,
real
>
col2im
;
GemmFunctor
<
Device
,
real
>
gemm
;
size_t
inputOffset
=
imShape
.
getElements
();
size_t
outputOffset
=
(
outputChannels
/
groups_
)
*
outputHeight
*
outputWidth
;
...
...
@@ -182,6 +200,11 @@ public:
int
K
=
outputChannels
/
groups_
;
int
N
=
outputHeight
*
outputWidth
;
int
M
=
inputChannels
/
groups_
*
filterHeight
*
filterWidth
;
real
scale
=
0.0
f
;
if
(
!
needIm2col
)
{
colData
=
inputGrad
+
g
*
inputOffset
;
scale
=
1.0
f
;
}
gemm
(
CblasTrans
,
CblasNoTrans
,
M
,
...
...
@@ -192,17 +215,19 @@ public:
M
,
outputGrad
+
g
*
outputOffset
,
N
,
0.0
f
,
scale
,
colData
,
N
);
col2im
(
inputGrad
+
g
*
inputOffset
,
imShape
,
colData
,
colShape
,
strideH
(),
strideW
(),
paddingH
(),
paddingW
());
if
(
needIm2col
)
{
col2im
(
inputGrad
+
g
*
inputOffset
,
imShape
,
colData
,
colShape
,
strideH
(),
strideW
(),
paddingH
(),
paddingW
());
}
}
inputGrad
+=
inputChannels
*
inputHeight
*
inputWidth
;
outputGrad
+=
outputChannels
*
outputHeight
*
outputWidth
;
...
...
@@ -255,16 +280,23 @@ public:
real
*
outputGrad
=
inputs
[
0
].
data
<
real
>
();
real
*
inputData
=
inputs
[
1
].
data
<
real
>
();
real
*
filterGrad
=
outputs
[
0
].
data
<
real
>
();
bool
needIm2col
=
isNeedIm2col
(
filter
);
TensorShape
imShape
=
TensorShape
({
inputChannels
/
groups_
,
inputHeight
,
inputWidth
});
TensorShape
colShape
=
TensorShape
({
inputChannels
/
groups_
,
filterHeight
,
filterWidth
,
outputHeight
,
outputWidth
});
resizeBuffer
<
Device
>
(
colShape
.
getElements
());
real
*
colData
=
reinterpret_cast
<
real
*>
(
memory_
->
getBuf
());
TensorShape
colShape
;
real
*
colData
=
NULL
;
if
(
needIm2col
)
{
colShape
=
TensorShape
({
inputChannels
/
groups_
,
filterHeight
,
filterWidth
,
outputHeight
,
outputWidth
});
resizeBuffer
<
Device
>
(
colShape
.
getElements
());
colData
=
reinterpret_cast
<
real
*>
(
memory_
->
getBuf
());
}
Im2ColFunctor
<
kCFO
,
Device
,
real
>
im2col
;
GemmFunctor
<
Device
,
real
>
gemm
;
...
...
@@ -274,15 +306,18 @@ public:
size_t
filterOffset
=
filter
.
getElements
()
/
groups_
;
for
(
size_t
i
=
0
;
i
<
batchSize
;
i
++
)
{
for
(
size_t
g
=
0
;
g
<
groups_
;
g
++
)
{
im2col
(
inputData
+
g
*
inputOffset
,
imShape
,
colData
,
colShape
,
strideH
(),
strideW
(),
paddingH
(),
paddingW
());
if
(
needIm2col
)
{
im2col
(
inputData
+
g
*
inputOffset
,
imShape
,
colData
,
colShape
,
strideH
(),
strideW
(),
paddingH
(),
paddingW
());
}
else
{
colData
=
inputData
+
g
*
inputOffset
;
}
int
M
=
outputChannels
/
groups_
;
int
K
=
outputHeight
*
outputWidth
;
int
N
=
inputChannels
/
groups_
*
filterHeight
*
filterWidth
;
...
...
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