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
2340ceda
P
Paddle
项目概览
PaddlePaddle
/
Paddle
12 个月 前同步成功
通知
2292
Star
20931
Fork
5422
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
体验新版 GitCode,发现更多精彩内容 >>
提交
2340ceda
编写于
9月 13, 2017
作者:
H
hedaoyuan
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Add groups in convolution GemmConvGradKernel.
上级
fb46345f
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
39 addition
and
29 deletion
+39
-29
paddle/operators/gemm_conv_op.h
paddle/operators/gemm_conv_op.h
+39
-29
未找到文件。
paddle/operators/gemm_conv_op.h
浏览文件 @
2340ceda
...
...
@@ -82,19 +82,16 @@ class GemmConvKernel : public framework::OpKernel {
int
in_step
=
input_channels
/
groups
;
int
out_step
=
output_channels
/
groups
;
for
(
int
i
=
0
;
i
<
batch_size
;
i
++
)
{
Tensor
in_slice_batch
=
input
->
Slice
<
T
>
(
i
,
i
+
1
).
Resize
(
input_shape
);
Tensor
out_slice_batch
=
output
->
Slice
<
T
>
(
i
,
i
+
1
).
Resize
(
output_matrix_shape
);
Tensor
in_batch
=
input
->
Slice
<
T
>
(
i
,
i
+
1
).
Resize
(
input_shape
);
Tensor
out_batch
=
output
->
Slice
<
T
>
(
i
,
i
+
1
).
Resize
(
output_matrix_shape
);
for
(
int
g
=
0
;
g
<
groups
;
g
++
)
{
// im2col
Tensor
in_slice
=
in_slice_batch
.
Slice
<
T
>
(
g
*
in_step
,
(
g
+
1
)
*
in_step
);
Tensor
in_slice
=
in_batch
.
Slice
<
T
>
(
g
*
in_step
,
(
g
+
1
)
*
in_step
);
im2col
(
in_slice
,
col
,
strides
[
0
],
strides
[
1
],
paddings
[
0
],
paddings
[
1
],
device_context
);
// gemm
Tensor
out_slice
=
out_slice_batch
.
Slice
<
T
>
(
g
*
out_step
,
(
g
+
1
)
*
out_step
);
Tensor
out_slice
=
out_batch
.
Slice
<
T
>
(
g
*
out_step
,
(
g
+
1
)
*
out_step
);
Tensor
filter_slice
=
filter
.
Slice
<
T
>
(
g
*
out_step
,
(
g
+
1
)
*
out_step
);
math
::
matmul
<
Place
,
T
>
(
filter_slice
,
false
,
col_matrix
,
false
,
T
(
1.0
),
&
out_slice
,
T
(
0.0
),
device_context
);
...
...
@@ -125,12 +122,13 @@ class GemmConvGradKernel : public framework::OpKernel {
std
::
vector
<
int
>
strides
=
context
.
Attr
<
std
::
vector
<
int
>>
(
"strides"
);
std
::
vector
<
int
>
paddings
=
context
.
Attr
<
std
::
vector
<
int
>>
(
"paddings"
);
//
int groups = context.Attr<int>("groups");
int
groups
=
context
.
Attr
<
int
>
(
"groups"
);
int
batch_size
=
input
->
dims
()[
0
];
int
input_channels
=
input
->
dims
()[
1
];
int
filter_height
=
filter
.
dims
()[
filter
.
dims
().
size
()
-
2
];
int
filter_width
=
filter
.
dims
()[
filter
.
dims
().
size
()
-
1
];
int
output_channels
=
output_grad
->
dims
()[
1
];
int
output_height
=
output_grad
->
dims
()[
2
];
int
output_width
=
output_grad
->
dims
()[
3
];
...
...
@@ -141,11 +139,11 @@ class GemmConvGradKernel : public framework::OpKernel {
paddle
::
operators
::
math
::
ColFormat
::
kCFO
,
Place
,
T
>
im2col
;
// use col_shape in the im2col and col2im calculation
framework
::
DDim
col_shape
=
{
input_channels
,
filter_height
,
filter_width
,
output_height
,
output_width
};
framework
::
DDim
col_shape
=
{
input_channels
/
groups
,
filter_height
,
filter_width
,
output_height
,
output_width
};
// use col_matrix_shape in the gemm calculation
framework
::
DDim
col_matrix_shape
=
{
input_channels
*
filter_height
*
filter_width
,
input_channels
/
groups
*
filter_height
*
filter_width
,
output_height
*
output_width
};
Tensor
col
;
col
.
mutable_data
<
T
>
(
col_shape
,
context
.
GetPlace
());
...
...
@@ -176,26 +174,38 @@ class GemmConvGradKernel : public framework::OpKernel {
// convolution backward input operator: gemm + col2im
// convolution backward weight operator: im2col + gemm
int
in_step
=
input_channels
/
groups
;
int
out_step
=
output_channels
/
groups
;
for
(
int
i
=
0
;
i
<
batch_size
;
i
++
)
{
// gemm
Tensor
out_slice
=
Tensor
out_grad_batch
=
output_grad
->
Slice
<
T
>
(
i
,
i
+
1
).
Resize
(
output_matrix_shape
);
math
::
matmul
<
Place
,
T
>
(
filter
,
true
,
out_slice
,
false
,
T
(
1.0
),
&
col_matrix
,
T
(
0.0
),
device_context
);
// col2im
Tensor
in_grad_slice
=
input_grad
->
Slice
<
T
>
(
i
,
i
+
1
).
Resize
(
input_shape
);
col2im
(
in_grad_slice
,
col
,
strides
[
0
],
strides
[
1
],
paddings
[
0
],
paddings
[
1
],
device_context
);
// im2col
Tensor
in_slice
=
input
->
Slice
<
T
>
(
i
,
i
+
1
).
Resize
(
input_shape
);
im2col
(
in_slice
,
col
,
strides
[
0
],
strides
[
1
],
paddings
[
0
],
paddings
[
1
],
device_context
);
// gemm
math
::
matmul
<
Place
,
T
>
(
out_slice
,
false
,
col_matrix
,
true
,
T
(
1.0
),
&
filter_grad
,
T
(
1.0
),
device_context
);
Tensor
in_grad_batch
=
input_grad
->
Slice
<
T
>
(
i
,
i
+
1
).
Resize
(
input_shape
);
Tensor
in_batch
=
input
->
Slice
<
T
>
(
i
,
i
+
1
).
Resize
(
input_shape
);
for
(
int
g
=
0
;
g
<
groups
;
g
++
)
{
// gemm
Tensor
out_grad_slice
=
out_grad_batch
.
Slice
<
T
>
(
g
*
out_step
,
(
g
+
1
)
*
out_step
);
Tensor
filter_slice
=
filter
.
Slice
<
T
>
(
g
*
out_step
,
(
g
+
1
)
*
out_step
);
math
::
matmul
<
Place
,
T
>
(
filter_slice
,
true
,
out_grad_slice
,
false
,
T
(
1.0
),
&
col_matrix
,
T
(
0.0
),
device_context
);
// col2im
Tensor
in_grad_slice
=
in_grad_batch
.
Slice
<
T
>
(
g
*
in_step
,
(
g
+
1
)
*
in_step
);
col2im
(
in_grad_slice
,
col
,
strides
[
0
],
strides
[
1
],
paddings
[
0
],
paddings
[
1
],
device_context
);
// im2col
Tensor
in_slice
=
in_batch
.
Slice
<
T
>
(
g
*
in_step
,
(
g
+
1
)
*
in_step
);
im2col
(
in_slice
,
col
,
strides
[
0
],
strides
[
1
],
paddings
[
0
],
paddings
[
1
],
device_context
);
// gemm
Tensor
filter_grad_slice
=
filter_grad
.
Slice
<
T
>
(
g
*
out_step
,
(
g
+
1
)
*
out_step
);
math
::
matmul
<
Place
,
T
>
(
out_grad_slice
,
false
,
col_matrix
,
true
,
T
(
1.0
),
&
filter_grad_slice
,
T
(
1.0
),
device_context
);
}
}
}
};
...
...
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