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-Lite
提交
98e69581
P
Paddle-Lite
项目概览
PaddlePaddle
/
Paddle-Lite
通知
332
Star
4
Fork
1
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
271
列表
看板
标记
里程碑
合并请求
78
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
Paddle-Lite
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
271
Issue
271
列表
看板
标记
里程碑
合并请求
78
合并请求
78
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
98e69581
编写于
7月 30, 2020
作者:
H
HappyAngel
提交者:
GitHub
7月 30, 2020
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[arm] change prior box implement (#4013)
* update prior profile, test=develop * fix review. test=develop * test=develop
上级
4f3cd537
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
334 addition
and
24 deletion
+334
-24
lite/core/mir/fusion/conv_conv_fuser.cc
lite/core/mir/fusion/conv_conv_fuser.cc
+3
-3
lite/kernels/arm/prior_box_compute.cc
lite/kernels/arm/prior_box_compute.cc
+324
-20
lite/kernels/arm/prior_box_compute.h
lite/kernels/arm/prior_box_compute.h
+7
-1
未找到文件。
lite/core/mir/fusion/conv_conv_fuser.cc
浏览文件 @
98e69581
...
...
@@ -117,11 +117,11 @@ void ConvConvFuser::InsertNewNode(SSAGraph* graph, const key2nodes_t& matched) {
<<
" must be 1"
;
}
for
(
int
i
=
0
;
i
<
paddings1
.
size
();
i
++
)
{
CHECK_EQ
(
paddings1
[
i
],
1
)
<<
"paddings
["
<<
i
<<
"]: "
<<
paddings1
[
i
]
<<
" must be
1
"
;
CHECK_EQ
(
paddings1
[
i
],
0
)
<<
"paddings1
["
<<
i
<<
"]: "
<<
paddings1
[
i
]
<<
" must be
0
"
;
}
for
(
int
i
=
0
;
i
<
dilations1
.
size
();
i
++
)
{
CHECK_EQ
(
dilations1
[
i
],
1
)
<<
"dilations["
<<
i
<<
"]: "
<<
dilations1
[
i
]
CHECK_EQ
(
dilations1
[
i
],
1
)
<<
"dilations
1
["
<<
i
<<
"]: "
<<
dilations1
[
i
]
<<
" must be 1"
;
}
// comupte new_wight and new bias
...
...
lite/kernels/arm/prior_box_compute.cc
浏览文件 @
98e69581
...
...
@@ -13,9 +13,11 @@
// limitations under the License.
#include "lite/kernels/arm/prior_box_compute.h"
#include <algorithm>
#include <string>
#include <vector>
#include "lite/backends/arm/math/funcs.h"
#include "lite/core/target_wrapper.h"
namespace
paddle
{
namespace
lite
{
...
...
@@ -46,9 +48,301 @@ inline void ExpandAspectRatios(const std::vector<float>& input_aspect_ratior,
}
}
void
PriorBoxCompute
::
Run
()
{
auto
&
param
=
Param
<
operators
::
PriorBoxParam
>
();
inline
void
fast_free
(
void
*
ptr
)
{
if
(
ptr
)
{
free
(
static_cast
<
void
**>
(
ptr
)[
-
1
]);
}
}
void
density_prior_box
(
const
lite
::
Tensor
*
input
,
const
lite
::
Tensor
*
image
,
lite
::
Tensor
*
boxes
,
lite
::
Tensor
*
variances
,
const
std
::
vector
<
float
>&
min_size_
,
const
std
::
vector
<
float
>&
fixed_size_
,
const
std
::
vector
<
float
>&
fixed_ratio_
,
const
std
::
vector
<
int
>&
density_size_
,
const
std
::
vector
<
float
>&
max_size_
,
const
std
::
vector
<
float
>&
aspect_ratio_
,
const
std
::
vector
<
float
>&
variance_
,
int
img_w_
,
int
img_h_
,
float
step_w_
,
float
step_h_
,
float
offset_
,
int
prior_num_
,
bool
is_flip_
,
bool
is_clip_
,
const
std
::
vector
<
std
::
string
>&
order_
,
bool
min_max_aspect_ratios_order
)
{
// compute output shape
int
win1
=
input
->
dims
()[
3
];
int
hin1
=
input
->
dims
()[
2
];
DDim
shape_out
({
hin1
,
win1
,
prior_num_
,
4
});
boxes
->
Resize
(
shape_out
);
variances
->
Resize
(
shape_out
);
float
*
_cpu_data
=
boxes
->
mutable_data
<
float
>
();
float
*
_variance_data
=
variances
->
mutable_data
<
float
>
();
const
int
width
=
win1
;
const
int
height
=
hin1
;
int
img_width
=
img_w_
;
int
img_height
=
img_h_
;
if
(
img_width
==
0
||
img_height
==
0
)
{
img_width
=
image
->
dims
()[
3
];
img_height
=
image
->
dims
()[
2
];
}
float
step_w
=
step_w_
;
float
step_h
=
step_h_
;
if
(
step_w
==
0
||
step_h
==
0
)
{
step_w
=
static_cast
<
float
>
(
img_width
)
/
width
;
step_h
=
static_cast
<
float
>
(
img_height
)
/
height
;
}
float
offset
=
offset_
;
int
step_average
=
static_cast
<
int
>
((
step_w
+
step_h
)
*
0.5
);
// add
int
channel_size
=
height
*
width
*
prior_num_
*
4
;
int
idx
=
0
;
for
(
int
h
=
0
;
h
<
height
;
++
h
)
{
for
(
int
w
=
0
;
w
<
width
;
++
w
)
{
float
center_x
=
(
w
+
offset
)
*
step_w
;
float
center_y
=
(
h
+
offset
)
*
step_h
;
float
box_width
;
float
box_height
;
if
(
fixed_size_
.
size
()
>
0
)
{
// add
for
(
int
s
=
0
;
s
<
fixed_size_
.
size
();
++
s
)
{
int
fixed_size
=
fixed_size_
[
s
];
int
com_idx
=
0
;
box_width
=
fixed_size
;
box_height
=
fixed_size
;
if
(
fixed_ratio_
.
size
()
>
0
)
{
for
(
int
r
=
0
;
r
<
fixed_ratio_
.
size
();
++
r
)
{
float
ar
=
fixed_ratio_
[
r
];
int
density
=
density_size_
[
s
];
int
shift
=
step_average
/
density
;
float
box_width_ratio
=
fixed_size_
[
s
]
*
sqrt
(
ar
);
float
box_height_ratio
=
fixed_size_
[
s
]
/
sqrt
(
ar
);
for
(
int
p
=
0
;
p
<
density
;
++
p
)
{
for
(
int
c
=
0
;
c
<
density
;
++
c
)
{
float
center_x_temp
=
center_x
-
step_average
/
2.0
f
+
shift
/
2.
f
+
c
*
shift
;
float
center_y_temp
=
center_y
-
step_average
/
2.0
f
+
shift
/
2.
f
+
p
*
shift
;
// xmin
_cpu_data
[
idx
++
]
=
(
center_x_temp
-
box_width_ratio
/
2.
f
)
/
img_width
>=
0
?
(
center_x_temp
-
box_width_ratio
/
2.
f
)
/
img_width
:
0
;
// ymin
_cpu_data
[
idx
++
]
=
(
center_y_temp
-
box_height_ratio
/
2.
f
)
/
img_height
>=
0
?
(
center_y_temp
-
box_height_ratio
/
2.
f
)
/
img_height
:
0
;
// xmax
_cpu_data
[
idx
++
]
=
(
center_x_temp
+
box_width_ratio
/
2.
f
)
/
img_width
<=
1
?
(
center_x_temp
+
box_width_ratio
/
2.
f
)
/
img_width
:
1
;
// ymax
_cpu_data
[
idx
++
]
=
(
center_y_temp
+
box_height_ratio
/
2.
f
)
/
img_height
<=
1
?
(
center_y_temp
+
box_height_ratio
/
2.
f
)
/
img_height
:
1
;
}
}
}
}
else
{
// this code for density anchor box
if
(
density_size_
.
size
()
>
0
)
{
CHECK_EQ
(
fixed_size_
.
size
(),
density_size_
.
size
())
<<
"fixed_size_ should be same with density_size_"
;
int
density
=
density_size_
[
s
];
int
shift
=
fixed_size_
[
s
]
/
density
;
for
(
int
r
=
0
;
r
<
density
;
++
r
)
{
for
(
int
c
=
0
;
c
<
density
;
++
c
)
{
float
center_x_temp
=
center_x
-
fixed_size
/
2.
f
+
shift
/
2.
f
+
c
*
shift
;
float
center_y_temp
=
center_y
-
fixed_size
/
2.
f
+
shift
/
2.
f
+
r
*
shift
;
// xmin
_cpu_data
[
idx
++
]
=
(
center_x_temp
-
box_width
/
2.
f
)
/
img_width
>=
0
?
(
center_x_temp
-
box_width
/
2.
f
)
/
img_width
:
0
;
// ymin
_cpu_data
[
idx
++
]
=
(
center_y_temp
-
box_height
/
2.
f
)
/
img_height
>=
0
?
(
center_y_temp
-
box_height
/
2.
f
)
/
img_height
:
0
;
// xmax
_cpu_data
[
idx
++
]
=
(
center_x_temp
+
box_width
/
2.
f
)
/
img_width
<=
1
?
(
center_x_temp
+
box_width
/
2.
f
)
/
img_width
:
1
;
// ymax
_cpu_data
[
idx
++
]
=
(
center_y_temp
+
box_height
/
2.
f
)
/
img_height
<=
1
?
(
center_y_temp
+
box_height
/
2.
f
)
/
img_height
:
1
;
}
}
}
// rest of priors: will never come here!!!
for
(
int
r
=
0
;
r
<
aspect_ratio_
.
size
();
++
r
)
{
float
ar
=
aspect_ratio_
[
r
];
if
(
fabs
(
ar
-
1.
)
<
1e-6
)
{
continue
;
}
int
density
=
density_size_
[
s
];
int
shift
=
fixed_size_
[
s
]
/
density
;
float
box_width_ratio
=
fixed_size_
[
s
]
*
sqrt
(
ar
);
float
box_height_ratio
=
fixed_size_
[
s
]
/
sqrt
(
ar
);
for
(
int
p
=
0
;
p
<
density
;
++
p
)
{
for
(
int
c
=
0
;
c
<
density
;
++
c
)
{
float
center_x_temp
=
center_x
-
fixed_size
/
2.
f
+
shift
/
2.
f
+
c
*
shift
;
float
center_y_temp
=
center_y
-
fixed_size
/
2.
f
+
shift
/
2.
f
+
p
*
shift
;
// xmin
_cpu_data
[
idx
++
]
=
(
center_x_temp
-
box_width_ratio
/
2.
f
)
/
img_width
>=
0
?
(
center_x_temp
-
box_width_ratio
/
2.
f
)
/
img_width
:
0
;
// ymin
_cpu_data
[
idx
++
]
=
(
center_y_temp
-
box_height_ratio
/
2.
f
)
/
img_height
>=
0
?
(
center_y_temp
-
box_height_ratio
/
2.
f
)
/
img_height
:
0
;
// xmax
_cpu_data
[
idx
++
]
=
(
center_x_temp
+
box_width_ratio
/
2.
f
)
/
img_width
<=
1
?
(
center_x_temp
+
box_width_ratio
/
2.
f
)
/
img_width
:
1
;
// ymax
_cpu_data
[
idx
++
]
=
(
center_y_temp
+
box_height_ratio
/
2.
f
)
/
img_height
<=
1
?
(
center_y_temp
+
box_height_ratio
/
2.
f
)
/
img_height
:
1
;
}
}
}
}
}
}
else
{
float
*
min_buf
=
reinterpret_cast
<
float
*>
(
TargetWrapper
<
TARGET
(
kHost
)
>::
Malloc
(
sizeof
(
float
)
*
4
));
float
*
max_buf
=
reinterpret_cast
<
float
*>
(
TargetWrapper
<
TARGET
(
kHost
)
>::
Malloc
(
sizeof
(
float
)
*
4
));
float
*
com_buf
=
reinterpret_cast
<
float
*>
(
TargetWrapper
<
TARGET
(
kHost
)
>::
Malloc
(
sizeof
(
float
)
*
aspect_ratio_
.
size
()
*
4
));
for
(
int
s
=
0
;
s
<
min_size_
.
size
();
++
s
)
{
int
min_idx
=
0
;
int
max_idx
=
0
;
int
com_idx
=
0
;
int
min_size
=
min_size_
[
s
];
// first prior: aspect_ratio = 1, size = min_size
box_width
=
box_height
=
min_size
;
//! xmin
min_buf
[
min_idx
++
]
=
(
center_x
-
box_width
/
2.
f
)
/
img_width
;
//! ymin
min_buf
[
min_idx
++
]
=
(
center_y
-
box_height
/
2.
f
)
/
img_height
;
//! xmax
min_buf
[
min_idx
++
]
=
(
center_x
+
box_width
/
2.
f
)
/
img_width
;
//! ymax
min_buf
[
min_idx
++
]
=
(
center_y
+
box_height
/
2.
f
)
/
img_height
;
if
(
max_size_
.
size
()
>
0
)
{
int
max_size
=
max_size_
[
s
];
//! second prior: aspect_ratio = 1, size = sqrt(min_size * max_size)
box_width
=
box_height
=
sqrtf
(
min_size
*
max_size
);
//! xmin
max_buf
[
max_idx
++
]
=
(
center_x
-
box_width
/
2.
f
)
/
img_width
;
//! ymin
max_buf
[
max_idx
++
]
=
(
center_y
-
box_height
/
2.
f
)
/
img_height
;
//! xmax
max_buf
[
max_idx
++
]
=
(
center_x
+
box_width
/
2.
f
)
/
img_width
;
//! ymax
max_buf
[
max_idx
++
]
=
(
center_y
+
box_height
/
2.
f
)
/
img_height
;
}
//! rest of priors
for
(
int
r
=
0
;
r
<
aspect_ratio_
.
size
();
++
r
)
{
float
ar
=
aspect_ratio_
[
r
];
if
(
fabs
(
ar
-
1.
)
<
1e-6
)
{
continue
;
}
box_width
=
min_size
*
sqrt
(
ar
);
box_height
=
min_size
/
sqrt
(
ar
);
//! xmin
com_buf
[
com_idx
++
]
=
(
center_x
-
box_width
/
2.
f
)
/
img_width
;
//! ymin
com_buf
[
com_idx
++
]
=
(
center_y
-
box_height
/
2.
f
)
/
img_height
;
//! xmax
com_buf
[
com_idx
++
]
=
(
center_x
+
box_width
/
2.
f
)
/
img_width
;
//! ymax
com_buf
[
com_idx
++
]
=
(
center_y
+
box_height
/
2.
f
)
/
img_height
;
}
if
(
min_max_aspect_ratios_order
)
{
memcpy
(
_cpu_data
+
idx
,
min_buf
,
sizeof
(
float
)
*
min_idx
);
idx
+=
min_idx
;
memcpy
(
_cpu_data
+
idx
,
max_buf
,
sizeof
(
float
)
*
max_idx
);
idx
+=
max_idx
;
memcpy
(
_cpu_data
+
idx
,
com_buf
,
sizeof
(
float
)
*
com_idx
);
idx
+=
com_idx
;
}
else
{
memcpy
(
_cpu_data
+
idx
,
min_buf
,
sizeof
(
float
)
*
min_idx
);
idx
+=
min_idx
;
memcpy
(
_cpu_data
+
idx
,
com_buf
,
sizeof
(
float
)
*
com_idx
);
idx
+=
com_idx
;
memcpy
(
_cpu_data
+
idx
,
max_buf
,
sizeof
(
float
)
*
max_idx
);
idx
+=
max_idx
;
}
}
TargetWrapper
<
TARGET
(
kHost
)
>::
Free
(
min_buf
);
TargetWrapper
<
TARGET
(
kHost
)
>::
Free
(
max_buf
);
TargetWrapper
<
TARGET
(
kHost
)
>::
Free
(
com_buf
);
}
}
}
//! clip the prior's coordinate such that it is within [0, 1]
if
(
is_clip_
)
{
for
(
int
d
=
0
;
d
<
channel_size
;
++
d
)
{
_cpu_data
[
d
]
=
std
::
min
(
std
::
max
(
_cpu_data
[
d
],
0.
f
),
1.
f
);
}
}
//! set the variance.
int
count
=
0
;
for
(
int
h
=
0
;
h
<
height
;
++
h
)
{
for
(
int
w
=
0
;
w
<
width
;
++
w
)
{
for
(
int
i
=
0
;
i
<
prior_num_
;
++
i
)
{
for
(
int
j
=
0
;
j
<
4
;
++
j
)
{
_variance_data
[
count
]
=
variance_
[
j
];
++
count
;
}
}
}
}
}
void
PriorBoxCompute
::
ReInitWhenNeeded
()
{
auto
&
param
=
this
->
template
Param
<
param_t
>();
auto
input_dims
=
param
.
input
->
dims
();
auto
image_dims
=
param
.
image
->
dims
();
if
(
last_input_shape_
==
input_dims
&&
last_image_shape_
==
image_dims
)
{
return
;
}
bool
is_flip
=
param
.
flip
;
bool
is_clip
=
param
.
clip
;
std
::
vector
<
float
>
min_size
=
param
.
min_sizes
;
...
...
@@ -66,25 +360,35 @@ void PriorBoxCompute::Run() {
prior_num
+=
max_size
.
size
();
std
::
vector
<
std
::
string
>
order
=
param
.
order
;
bool
min_max_aspect_ratios_order
=
param
.
min_max_aspect_ratios_order
;
density_prior_box
(
param
.
input
,
param
.
image
,
&
boxes_tmp_
,
&
variances_tmp_
,
min_size
,
std
::
vector
<
float
>
(),
std
::
vector
<
float
>
(),
std
::
vector
<
int
>
(),
max_size
,
aspect_ratios_vec
,
variance
,
img_w
,
img_h
,
step_w
,
step_h
,
offset
,
prior_num
,
is_flip
,
is_clip
,
order
,
min_max_aspect_ratios_order
);
last_input_shape_
=
input_dims
;
last_image_shape_
=
image_dims
;
}
lite
::
arm
::
math
::
prior_box
(
param
.
input
,
param
.
image
,
&
param
.
boxes
,
&
param
.
variances
,
min_size
,
max_size
,
aspect_ratios_vec
,
variance
,
img_w
,
img_h
,
step_w
,
step_h
,
offset
,
prior_num
,
is_flip
,
is_clip
,
order
,
min_max_aspect_ratios_order
);
void
PriorBoxCompute
::
Run
()
{
auto
&
param
=
this
->
template
Param
<
param_t
>();
param
.
boxes
->
CopyDataFrom
(
boxes_tmp_
);
param
.
variances
->
CopyDataFrom
(
variances_tmp_
);
}
}
// namespace arm
...
...
lite/kernels/arm/prior_box_compute.h
浏览文件 @
98e69581
...
...
@@ -26,8 +26,14 @@ class PriorBoxCompute : public KernelLite<TARGET(kARM), PRECISION(kFloat)> {
using
param_t
=
operators
::
PriorBoxParam
;
void
Run
()
override
;
void
ReInitWhenNeeded
()
override
;
virtual
~
PriorBoxCompute
()
=
default
;
private:
Tensor
boxes_tmp_
;
Tensor
variances_tmp_
;
DDim
last_input_shape_
;
DDim
last_image_shape_
;
};
}
// namespace arm
...
...
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