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
机器未来
Paddle
提交
21c0f874
P
Paddle
项目概览
机器未来
/
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 >>
未验证
提交
21c0f874
编写于
12月 05, 2018
作者:
J
Jiabin Yang
提交者:
GitHub
12月 05, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #14728 from JiabinYang/optimize_hs_op
Optimize hs op
上级
439af8d5
7cbcb2b8
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
46 addition
and
17 deletion
+46
-17
paddle/fluid/operators/hierarchical_sigmoid_op.cc
paddle/fluid/operators/hierarchical_sigmoid_op.cc
+1
-0
paddle/fluid/operators/hierarchical_sigmoid_op.h
paddle/fluid/operators/hierarchical_sigmoid_op.h
+0
-1
paddle/fluid/operators/math/matrix_bit_code.cc
paddle/fluid/operators/math/matrix_bit_code.cc
+41
-16
paddle/fluid/operators/math/matrix_bit_code.h
paddle/fluid/operators/math/matrix_bit_code.h
+4
-0
未找到文件。
paddle/fluid/operators/hierarchical_sigmoid_op.cc
浏览文件 @
21c0f874
...
...
@@ -158,6 +158,7 @@ class HierarchicalSigmoidGradOp : public framework::OperatorWithKernel {
ctx
->
SetOutputDim
(
framework
::
GradVarName
(
"W"
),
ctx
->
GetInputDim
(
"W"
));
}
ctx
->
SetOutputDim
(
framework
::
GradVarName
(
"X"
),
ctx
->
GetInputDim
(
"X"
));
ctx
->
ShareLoD
(
"X"
,
/*->*/
framework
::
GradVarName
(
"X"
));
}
protected:
...
...
paddle/fluid/operators/hierarchical_sigmoid_op.h
浏览文件 @
21c0f874
...
...
@@ -185,7 +185,6 @@ class HierarchicalSigmoidGradOpKernel : public framework::OpKernel<T> {
ctx
.
Output
<
framework
::
SelectedRows
>
(
framework
::
GradVarName
(
"W"
));
w_grad
->
set_rows
(
real_rows
);
// Build a map of id -> row_index to speed up finding the index of one id
w_grad
->
SyncIndex
();
w_grad
->
set_height
(
w
.
dims
()[
0
]);
auto
*
w_grad_value
=
w_grad
->
mutable_value
();
framework
::
DDim
temp_dim
(
w
.
dims
());
...
...
paddle/fluid/operators/math/matrix_bit_code.cc
浏览文件 @
21c0f874
...
...
@@ -89,6 +89,8 @@ template <typename T>
void
MatrixBitCodeFunctor
<
T
>::
Mul
(
framework
::
Tensor
*
tmat
,
const
framework
::
Tensor
&
weight
,
const
framework
::
Tensor
&
input
)
{
auto
blas
=
GetBlas
<
platform
::
CPUDeviceContext
,
T
>
(
platform
::
CPUDeviceContext
());
size_t
num_samples
=
tmat
->
dims
()[
0
];
size_t
tmat_width
=
tmat
->
dims
()[
1
];
size_t
input_width
=
input
.
dims
()[
1
];
...
...
@@ -99,13 +101,12 @@ void MatrixBitCodeFunctor<T>::Mul(framework::Tensor* tmat,
for
(
size_t
i
=
0
;
i
<
num_samples
;
++
i
)
{
auto
code
=
code_table_
->
get_code
(
i
);
int
code_length
=
code
->
get_length
();
const
T
*
input_row
=
input_value
+
input_width
*
i
;
for
(
int
j
=
0
;
j
<
code_length
;
++
j
)
{
size_t
index
=
code
->
calc_index
(
j
);
const
T
*
weight_row
=
weight_value
+
weight_width
*
index
;
T
sum
=
static_cast
<
T
>
(
0.0
);
for
(
size_t
k
=
0
;
k
<
input_width
;
++
k
)
{
sum
+=
weight_value
[
weight_width
*
index
+
k
]
*
input_value
[
input_width
*
i
+
k
];
}
sum
=
blas
.
DOT
(
input_width
,
weight_row
,
input_row
);
tmat_value
[
i
*
tmat_width
+
j
]
+=
sum
;
}
}
...
...
@@ -115,6 +116,8 @@ template <typename T>
void
MatrixBitCodeFunctor
<
T
>::
MulGradWeight
(
const
framework
::
Tensor
&
tmat
,
framework
::
Tensor
*
weight
,
const
framework
::
Tensor
&
input
)
{
auto
blas
=
GetBlas
<
platform
::
CPUDeviceContext
,
T
>
(
platform
::
CPUDeviceContext
());
size_t
num_samples
=
tmat
.
dims
()[
0
];
size_t
input_width
=
input
.
dims
()[
1
];
size_t
tmat_width
=
tmat
.
dims
()[
1
];
...
...
@@ -122,16 +125,25 @@ void MatrixBitCodeFunctor<T>::MulGradWeight(const framework::Tensor& tmat,
auto
tmat_value
=
tmat
.
data
<
T
>
();
auto
weight_value
=
weight
->
data
<
T
>
();
auto
input_value
=
input
.
data
<
T
>
();
std
::
unordered_map
<
int
,
std
::
vector
<
std
::
pair
<
T
,
const
T
*>>>
ops
;
for
(
size_t
i
=
0
;
i
<
num_samples
;
++
i
)
{
auto
code
=
code_table_
->
get_code
(
i
);
int
code_length
=
code
->
get_length
();
const
T
*
input_value_row
=
input_value
+
input_width
*
i
;
const
T
*
tmat_row
=
tmat_value
+
i
*
tmat_width
;
for
(
int
j
=
0
;
j
<
code_length
;
++
j
)
{
size_t
index
=
code
->
calc_index
(
j
);
for
(
size_t
k
=
0
;
k
<
input_width
;
++
k
)
{
weight_value
[
weight_width
*
index
+
k
]
+=
tmat_value
[
i
*
tmat_width
+
j
]
*
input_value
[
input_width
*
i
+
k
];
}
ops
[
code
->
calc_index
(
j
)].
emplace_back
(
tmat_row
[
j
],
input_value_row
);
}
}
for
(
auto
&
op
:
ops
)
{
auto
&
op_in_row
=
op
.
second
;
for
(
auto
&
pair
:
op_in_row
)
{
auto
&
scale
=
pair
.
first
;
auto
*
input_row
=
pair
.
second
;
T
*
weight_row
=
weight_value
+
op
.
first
*
weight_width
;
blas
.
AXPY
(
input_width
,
scale
,
input_row
,
weight_row
);
}
}
}
...
...
@@ -140,6 +152,8 @@ template <typename T>
void
MatrixBitCodeFunctor
<
T
>::
MulGradWeight
(
const
framework
::
Tensor
&
tmat
,
framework
::
SelectedRows
*
weight
,
const
framework
::
Tensor
&
input
)
{
auto
blas
=
GetBlas
<
platform
::
CPUDeviceContext
,
T
>
(
platform
::
CPUDeviceContext
());
size_t
num_samples
=
tmat
.
dims
()[
0
];
size_t
input_width
=
input
.
dims
()[
1
];
size_t
tmat_width
=
tmat
.
dims
()[
1
];
...
...
@@ -147,17 +161,28 @@ void MatrixBitCodeFunctor<T>::MulGradWeight(const framework::Tensor& tmat,
auto
tmat_value
=
tmat
.
data
<
T
>
();
auto
weight_value
=
weight
->
mutable_value
()
->
data
<
T
>
();
auto
input_value
=
input
.
data
<
T
>
();
std
::
unordered_map
<
int
,
std
::
vector
<
std
::
pair
<
T
,
const
T
*>>>
ops
;
ops
.
reserve
(
weight
->
rows
().
size
());
for
(
size_t
i
=
0
;
i
<
num_samples
;
++
i
)
{
auto
code
=
code_table_
->
get_code
(
i
);
int
code_length
=
code
->
get_length
();
const
T
*
input_value_row
=
input_value
+
input_width
*
i
;
const
T
*
tmat_row
=
tmat_value
+
i
*
tmat_width
;
for
(
int
j
=
0
;
j
<
code_length
;
++
j
)
{
size_t
index
=
code
->
calc_index
(
j
);
for
(
size_t
k
=
0
;
k
<
input_width
;
++
k
)
{
int64_t
row_index
=
weight
->
GetIndexFromId
(
static_cast
<
int64_t
>
(
index
));
weight_value
[
row_index
*
weight_width
+
k
]
+=
tmat_value
[
i
*
tmat_width
+
j
]
*
input_value
[
input_width
*
i
+
k
];
}
ops
[
code
->
calc_index
(
j
)].
emplace_back
(
tmat_row
[
j
],
input_value_row
);
}
}
for
(
auto
&
row
:
weight
->
rows
())
{
auto
&
op_in_row
=
ops
[
row
];
for
(
auto
&
pair
:
op_in_row
)
{
auto
&
scale
=
pair
.
first
;
auto
*
input_row
=
pair
.
second
;
blas
.
AXPY
(
input_width
,
scale
,
input_row
,
weight_value
);
}
weight_value
+=
weight_width
;
}
}
...
...
paddle/fluid/operators/math/matrix_bit_code.h
浏览文件 @
21c0f874
...
...
@@ -13,10 +13,14 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <unordered_map>
#include <utility>
#include <vector>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/lod_tensor.h"
#include "paddle/fluid/framework/selected_rows.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/operators/math/blas.h"
#include "paddle/fluid/platform/device_context.h"
#if defined(_WIN32)
...
...
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