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
BaiXuePrincess
Paddle
提交
36e26a53
P
Paddle
项目概览
BaiXuePrincess
/
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
0
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
0
Issue
0
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
提交
36e26a53
编写于
11月 28, 2018
作者:
Q
Qiao Longfei
提交者:
qingqing01
11月 28, 2018
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Optimize bilinear tensor product op (#14485)
* optimize bilinear_tensor_product * add set zero to set grad to 0.
上级
4ec9de01
变更
1
隐藏空白更改
内联
并排
Showing
1 changed file
with
30 addition
and
31 deletion
+30
-31
paddle/fluid/operators/bilinear_tensor_product_op.h
paddle/fluid/operators/bilinear_tensor_product_op.h
+30
-31
未找到文件。
paddle/fluid/operators/bilinear_tensor_product_op.h
浏览文件 @
36e26a53
...
...
@@ -70,7 +70,7 @@ class BilinearTensorProductKernel : public framework::OpKernel<T> {
if
(
bias
)
{
auto
bias_vec
=
EigenMatrix
<
T
>::
From
(
*
bias
);
Eigen
::
DSizes
<
int
,
2
>
bcast
(
batch_size
,
1
);
output_mat
.
device
(
place
)
=
bias_vec
.
broadcast
(
bcast
)
+
output_mat
;
output_mat
.
device
(
place
)
=
bias_vec
.
broadcast
(
bcast
)
.
eval
()
+
output_mat
;
}
}
};
...
...
@@ -99,13 +99,13 @@ class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
auto
d_out_mat
=
EigenMatrix
<
T
>::
From
(
*
d_out
);
auto
&
place
=
*
ctx
.
template
device_context
<
DeviceContext
>().
eigen_device
();
auto
&
dev_ctx
=
ctx
.
template
device_context
<
DeviceContext
>();
// Create the intermediate variable to caculate the Output(Y@Grad).
// Create the intermediate variable to ca
l
culate the Output(Y@Grad).
Tensor
x_scale
;
x_scale
.
mutable_data
<
T
>
(
framework
::
make_ddim
({
batch_size
,
x_dim
}),
ctx
.
GetPlace
());
auto
x_scale_mat
=
EigenMatrix
<
T
>::
From
(
x_scale
);
// Create the intermediate variable to caculate the Output(X@Grad).
// Create the intermediate variable to ca
l
culate the Output(X@Grad).
Tensor
y_scale
;
y_scale
.
mutable_data
<
T
>
(
framework
::
make_ddim
({
batch_size
,
y_dim
}),
ctx
.
GetPlace
());
...
...
@@ -113,65 +113,64 @@ class BilinearTensorProductGradKernel : public framework::OpKernel<T> {
math
::
SetConstant
<
DeviceContext
,
T
>
set_zero
;
// Set Output(X@Grad) be zero.
if
(
d_x
)
{
d_x
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
set_zero
(
dev_ctx
,
d_x
,
static_cast
<
T
>
(
0
));
}
// Set Output(Y@Grad) be zero.
if
(
d_y
)
{
d_y
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
set_zero
(
dev_ctx
,
d_y
,
static_cast
<
T
>
(
0
));
}
if
(
d_weight
)
{
d_weight
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
}
auto
blas
=
math
::
GetBlas
<
DeviceContext
,
T
>
(
ctx
);
// Caculate the Output(X@Grad) and Output(Y@Grad).
if
(
d_x
||
d_y
)
{
if
(
d_x
||
d_y
||
d_weight
)
{
Eigen
::
DSizes
<
int
,
2
>
bcast_for_x
(
1
,
y_dim
);
Eigen
::
DSizes
<
int
,
2
>
bcast_for_y
(
1
,
x_dim
);
Eigen
::
DSizes
<
int
,
2
>
bcast_for_weight
(
1
,
x_dim
);
for
(
int
i
=
0
;
i
<
out_dim
;
++
i
)
{
Tensor
weight_i
=
weight
->
Slice
(
i
,
i
+
1
).
Resize
(
framework
::
make_ddim
({
x_dim
,
y_dim
}));
auto
output_vec
=
d_out_mat
.
chip
(
i
,
1
);
if
(
d_x
)
{
y_scale_mat
.
device
(
place
)
=
output_vec
.
reshape
(
Eigen
::
DSizes
<
int
,
2
>
(
batch_size
,
1
))
.
broadcast
(
bcast_for_x
)
*
.
broadcast
(
bcast_for_x
)
.
eval
()
*
y_mat
;
blas
.
GEMM
(
CblasNoTrans
,
CblasTrans
,
batch_size
,
x_dim
,
y_dim
,
1
,
y_scale
.
data
<
T
>
(),
weight_i
.
data
<
T
>
(),
1
,
d_x
->
data
<
T
>
());
}
if
(
d_y
)
{
x_scale_mat
.
device
(
place
)
=
if
(
d_y
||
d_weight
)
{
auto
output_vec_y
=
output_vec
.
reshape
(
Eigen
::
DSizes
<
int
,
2
>
(
batch_size
,
1
))
.
broadcast
(
bcast_for_y
)
*
x_mat
;
blas
.
GEMM
(
CblasNoTrans
,
CblasNoTrans
,
batch_size
,
y_dim
,
x_dim
,
1
,
x_scale
.
data
<
T
>
(),
weight_i
.
data
<
T
>
(),
1
,
d_y
->
data
<
T
>
());
.
broadcast
(
bcast_for_y
)
.
eval
();
x_scale_mat
.
device
(
place
)
=
output_vec_y
*
x_mat
;
if
(
d_y
)
{
blas
.
GEMM
(
CblasNoTrans
,
CblasNoTrans
,
batch_size
,
y_dim
,
x_dim
,
1
,
x_scale
.
data
<
T
>
(),
weight_i
.
data
<
T
>
(),
1
,
d_y
->
data
<
T
>
());
}
if
(
d_weight
)
{
Tensor
d_weight_i
=
d_weight
->
Slice
(
i
,
i
+
1
).
Resize
(
framework
::
make_ddim
({
x_dim
,
y_dim
}));
blas
.
GEMM
(
CblasTrans
,
CblasNoTrans
,
x_dim
,
y_dim
,
batch_size
,
1
,
x_scale
.
data
<
T
>
(),
y
->
data
<
T
>
(),
0
,
d_weight_i
.
data
<
T
>
());
}
}
}
}
// Caculate the gradient of Input(Weight).
if
(
d_weight
)
{
d_weight
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
Eigen
::
DSizes
<
int
,
2
>
bcast_for_weight
(
1
,
x_dim
);
for
(
int
i
=
0
;
i
<
out_dim
;
++
i
)
{
Tensor
d_weight_i
=
d_weight
->
Slice
(
i
,
i
+
1
).
Resize
(
framework
::
make_ddim
({
x_dim
,
y_dim
}));
auto
output_vec
=
d_out_mat
.
chip
(
i
,
1
);
x_scale_mat
.
device
(
place
)
=
output_vec
.
reshape
(
Eigen
::
DSizes
<
int
,
2
>
(
batch_size
,
1
))
.
broadcast
(
bcast_for_weight
)
*
x_mat
;
blas
.
GEMM
(
CblasTrans
,
CblasNoTrans
,
x_dim
,
y_dim
,
batch_size
,
1
,
x_scale
.
data
<
T
>
(),
y
->
data
<
T
>
(),
0
,
d_weight_i
.
data
<
T
>
());
}
}
// Caculate the gradient of Input(Bias).
// calculate the gradient of Input(Bias).
if
(
d_bias
)
{
d_bias
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
d_bias_mat
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
d_bias
);
...
...
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