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
wzpei9
Paddle
提交
6e17babe
P
Paddle
项目概览
wzpei9
/
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
6e17babe
编写于
1月 30, 2018
作者:
X
xzl
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
More efficient, add check on python side
上级
b5ea0483
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
26 addition
and
30 deletion
+26
-30
paddle/operators/CMakeLists.txt
paddle/operators/CMakeLists.txt
+0
-1
paddle/operators/math/depthwise_conv.cu
paddle/operators/math/depthwise_conv.cu
+24
-28
python/paddle/v2/fluid/layers/nn.py
python/paddle/v2/fluid/layers/nn.py
+2
-1
未找到文件。
paddle/operators/CMakeLists.txt
浏览文件 @
6e17babe
...
...
@@ -159,7 +159,6 @@ if (WITH_GPU)
op_library
(
conv_op SRCS conv_op.cc conv_op.cu.cc conv_cudnn_op.cu.cc DEPS
vol2col depthwise_conv
)
# op_library(conv_op SRCS conv_op.cc conv_op.cu.cc conv_cudnn_op.cu.cc DEPS vol2col)
op_library
(
edit_distance_op SRCS edit_distance_op.cc edit_distance_op.cu DEPS math_function
)
op_library
(
pool_op SRCS pool_op.cc pool_op.cu.cc pool_cudnn_op.cu.cc DEPS pooling
)
op_library
(
conv_transpose_op SRCS conv_transpose_op.cc conv_transpose_op.cu.cc
...
...
paddle/operators/math/depthwise_conv.cu
浏览文件 @
6e17babe
...
...
@@ -46,16 +46,18 @@ __global__ void KernelDepthwiseConv(
-
padding_height
+
h_out
*
stride_height
+
filter_height
-
1
;
const
int
w_in_end
=
-
padding_width
+
w_out
*
stride_width
+
filter_width
-
1
;
const
int
in_offset
=
((
batch
*
input_channels
+
c_in
)
*
input_height
)
*
input_width
;
if
((
h_in_start
>=
0
)
&&
(
h_in_end
<
input_height
)
&&
(
w_in_start
>=
0
)
&&
(
w_in_end
<
input_width
))
{
for
(
int
kh
=
0
;
kh
<
filter_height
;
++
kh
)
{
for
(
int
kw
=
0
;
kw
<
filter_width
;
++
kw
)
{
const
int
h_in
=
-
padding_height
+
h_out
*
stride_height
+
kh
;
const
int
w_in
=
-
padding_width
+
w_out
*
stride_width
+
kw
;
const
int
offset
=
((
batch
*
input_channels
+
c_in
)
*
input_height
+
h_in
)
*
input_width
+
w_in
;
const
int
h_in
=
h_in_start
+
kh
;
const
int
w_in
=
w_in_start
+
kw
;
const
int
offset
=
in_offset
+
h_in
*
input_width
+
w_in
;
value
+=
(
*
weight
)
*
input_data
[
offset
];
++
weight
;
}
...
...
@@ -63,14 +65,11 @@ __global__ void KernelDepthwiseConv(
}
else
{
for
(
int
kh
=
0
;
kh
<
filter_height
;
++
kh
)
{
for
(
int
kw
=
0
;
kw
<
filter_width
;
++
kw
)
{
const
int
h_in
=
-
padding_height
+
h_out
*
stride_heigh
t
+
kh
;
const
int
w_in
=
-
padding_width
+
w_out
*
stride_width
+
kw
;
const
int
h_in
=
h_in_star
t
+
kh
;
const
int
w_in
=
w_in_start
+
kw
;
if
((
h_in
>=
0
)
&&
(
h_in
<
input_height
)
&&
(
w_in
>=
0
)
&&
(
w_in
<
input_width
))
{
const
int
offset
=
((
batch
*
input_channels
+
c_in
)
*
input_height
+
h_in
)
*
input_width
+
w_in
;
const
int
offset
=
in_offset
+
h_in
*
input_width
+
w_in
;
value
+=
(
*
weight
)
*
input_data
[
offset
];
}
++
weight
;
...
...
@@ -159,36 +158,33 @@ __global__ void KernelDepthwiseConvFilterGrad(
const
int
h_in_end
=
-
padding_height
+
h_out
*
stride_height
+
filter_height
;
const
int
w_in_end
=
-
padding_width
+
w_out
*
stride_width
+
filter_width
;
const
int
in_offset
=
(
batch
*
input_channels
+
c_in
)
*
input_height
*
input_width
;
T
*
addr_offset
=
filter_grad_data
+
c_out
*
filter_height
*
filter_width
;
if
((
h_in_start
>=
0
)
&&
(
h_in_end
<
input_height
)
&&
(
w_in_start
>=
0
)
&&
(
w_in_end
<
input_width
))
{
for
(
int
kw
=
0
;
kw
<
filter_width
;
kw
++
)
{
for
(
int
kh
=
0
;
kh
<
filter_height
;
kh
++
)
{
const
int
h_in
=
-
padding_height
+
h_out
*
stride_height
+
kh
;
const
int
w_in
=
-
padding_width
+
w_out
*
stride_width
+
kw
;
const
int
offset
=
((
batch
*
input_channels
+
c_in
)
*
input_height
+
h_in
)
*
input_width
+
w_in
;
const
int
h_in
=
h_in_start
+
kh
;
const
int
w_in
=
w_in_start
+
kw
;
const
int
offset
=
in_offset
+
h_in
*
input_width
+
w_in
;
const
T
diff_temp
=
output_grad_data
[
index
]
*
input_data
[
offset
];
T
*
addr
=
filter_grad_data
+
c_out
*
filter_height
*
filter_width
+
kh
*
filter_width
+
kw
;
T
*
addr
=
addr_offset
+
kh
*
filter_width
+
kw
;
paddle
::
platform
::
CudaAtomicAdd
(
addr
,
diff_temp
);
}
}
}
else
{
for
(
int
kw
=
0
;
kw
<
filter_width
;
kw
++
)
{
for
(
int
kh
=
0
;
kh
<
filter_height
;
kh
++
)
{
const
int
h_in
=
-
padding_height
+
h_out
*
stride_heigh
t
+
kh
;
const
int
w_in
=
-
padding_width
+
w_out
*
stride_width
+
kw
;
const
int
h_in
=
h_in_star
t
+
kh
;
const
int
w_in
=
w_in_start
+
kw
;
if
((
h_in
>=
0
)
&&
(
h_in
<
input_height
)
&&
(
w_in
>=
0
)
&&
(
w_in
<
input_width
))
{
const
int
offset
=
((
batch
*
input_channels
+
c_in
)
*
input_height
+
h_in
)
*
input_width
+
w_in
;
const
int
offset
=
in_offset
+
h_in
*
input_width
+
w_in
;
const
T
diff_temp
=
output_grad_data
[
index
]
*
input_data
[
offset
];
T
*
addr
=
filter_grad_data
+
c_out
*
filter_height
*
filter_width
+
kh
*
filter_width
+
kw
;
T
*
addr
=
addr_offset
+
kh
*
filter_width
+
kw
;
paddle
::
platform
::
CudaAtomicAdd
(
addr
,
diff_temp
);
}
}
...
...
python/paddle/v2/fluid/layers/nn.py
浏览文件 @
6e17babe
...
...
@@ -1013,7 +1013,8 @@ def conv2d(input,
num_channels
=
input
.
shape
[
1
]
l_type
=
'conv2d'
if
num_channels
==
groups
and
not
use_cudnn
:
if
(
num_channels
==
groups
and
num_filters
%
num_channels
==
0
and
not
use_cudnn
):
l_type
=
'depthwise_conv'
helper
=
LayerHelper
(
l_type
,
**
locals
())
...
...
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