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
39ef5736
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2298
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
39ef5736
编写于
12月 25, 2017
作者:
Y
Yang Yu
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Use tranform to rewrite adam
上级
cbe973f9
变更
1
显示空白变更内容
内联
并排
Showing
1 changed file
with
93 addition
and
39 deletion
+93
-39
paddle/operators/adam_op.h
paddle/operators/adam_op.h
+93
-39
未找到文件。
paddle/operators/adam_op.h
浏览文件 @
39ef5736
...
...
@@ -13,59 +13,113 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include
"paddle/framework/eigen.h"
#include
<math.h> // for sqrt in CPU and CUDA
#include "paddle/framework/op_registry.h"
#include "paddle/operators/detail/safe_ref.h"
#include "paddle/platform/transform.h"
namespace
paddle
{
namespace
operators
{
template
<
typename
T
>
struct
AdamFunctor
{
T
beta1_
;
T
beta2_
;
T
epsilon_
;
const
T
*
beta1_pow_
;
const
T
*
beta2_pow_
;
const
T
*
moment1_
;
T
*
moment1_out_
;
const
T
*
moment2_
;
T
*
moment2_out_
;
const
T
*
lr_
;
const
T
*
grad_
;
const
T
*
param_
;
AdamFunctor
(
T
beta1
,
T
beta2
,
T
epsilon
,
const
T
*
beta1_pow
,
const
T
*
beta2_pow
,
const
T
*
mom1
,
T
*
mom1_out
,
const
T
*
mom2
,
T
*
mom2_out
,
const
T
*
lr
,
const
T
*
grad
,
const
T
*
param
)
:
beta1_
(
beta1
),
beta2_
(
beta2
),
epsilon_
(
epsilon
),
beta1_pow_
(
beta1_pow
),
beta2_pow_
(
beta2_pow
),
moment1_
(
mom1
),
moment1_out_
(
mom1_out
),
moment2_
(
mom2
),
moment2_out_
(
mom2_out
),
lr_
(
lr
),
grad_
(
grad
),
param_
(
param
)
{}
// From param[i] --> param_out[i];
inline
HOSTDEVICE
T
operator
()(
const
T
&
p
)
const
{
size_t
i
=
&
p
-
param_
;
// Merge all memory access together.
T
g
=
grad_
[
i
];
T
mom1
=
moment1_
[
i
];
T
mom2
=
moment2_
[
i
];
T
lr
=
*
lr_
;
T
beta1_pow
=
*
beta1_pow_
;
T
beta2_pow
=
*
beta2_pow_
;
// Calculation
lr
=
lr
*
sqrt
(
1
-
beta2_pow
)
/
(
1
-
beta1_pow
);
mom1
=
beta1_
*
mom1
+
(
1
-
beta1_
)
*
g
;
mom2
=
beta2_
*
mom2
+
(
1
-
beta2_
)
*
g
*
g
;
T
new_p
=
p
-
lr
*
(
mom1
/
(
sqrt
(
mom2
)
+
epsilon_
));
// Write back to global memory
moment1_out_
[
i
]
=
mom1
;
moment2_out_
[
i
]
=
mom2
;
return
new_p
;
}
};
template
<
typename
DeviceContext
,
typename
T
>
class
AdamOpKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
param_out_tensor
=
ctx
.
Output
<
framework
::
Tensor
>
(
"ParamOut"
);
auto
moment1_out_tensor
=
ctx
.
Output
<
framework
::
Tensor
>
(
"Moment1Out"
);
auto
moment2_out_tensor
=
ctx
.
Output
<
framework
::
Tensor
>
(
"Moment2Out"
);
param_out_tensor
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
moment1_out_tensor
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
moment2_out_tensor
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
using
paddle
::
framework
::
LoDTensor
;
using
paddle
::
operators
::
detail
::
Ref
;
T
beta1
=
static_cast
<
T
>
(
ctx
.
Attr
<
float
>
(
"beta1"
));
T
beta2
=
static_cast
<
T
>
(
ctx
.
Attr
<
float
>
(
"beta2"
));
T
epsilon
=
static_cast
<
T
>
(
ctx
.
Attr
<
float
>
(
"epsilon"
));
auto
&
param
=
Ref
(
ctx
.
Input
<
LoDTensor
>
(
"Param"
),
"Must set Param"
);
auto
&
grad
=
Ref
(
ctx
.
Input
<
LoDTensor
>
(
"Grad"
),
"Must set Grad"
);
auto
&
mom1
=
Ref
(
ctx
.
Input
<
LoDTensor
>
(
"Moment1"
),
"Must set Moment1"
);
auto
&
mom2
=
Ref
(
ctx
.
Input
<
LoDTensor
>
(
"Moment2"
),
"Must set Moment2"
);
auto
&
lr
=
Ref
(
ctx
.
Input
<
LoDTensor
>
(
"LearningRate"
),
"Must set LearningRate"
);
auto
&
beta1_pow
=
Ref
(
ctx
.
Input
<
LoDTensor
>
(
"Beta1Pow"
),
"Must set Beta1Pow"
);
auto
&
beta2_pow
=
Ref
(
ctx
.
Input
<
LoDTensor
>
(
"Beta2Pow"
),
"Must set Beta2Pow"
);
auto
&
param_out
=
Ref
(
ctx
.
Output
<
LoDTensor
>
(
"ParamOut"
),
"Must set ParamOut"
);
auto
&
mom1_out
=
Ref
(
ctx
.
Output
<
LoDTensor
>
(
"Moment1Out"
),
"Must set Moment1Out"
);
auto
&
mom2_out
=
Ref
(
ctx
.
Output
<
LoDTensor
>
(
"Moment2Out"
),
"Must set Moment1Out"
);
AdamFunctor
<
T
>
functor
(
beta1
,
beta2
,
epsilon
,
beta1_pow
.
template
data
<
T
>(),
beta2_pow
.
template
data
<
T
>(),
mom1
.
template
data
<
T
>(),
mom1_out
.
template
mutable_data
<
T
>(
ctx
.
GetPlace
()),
mom2
.
template
data
<
T
>(),
mom2_out
.
template
mutable_data
<
T
>(
ctx
.
GetPlace
()),
lr
.
template
data
<
T
>(),
grad
.
template
data
<
T
>(),
param
.
template
data
<
T
>());
auto
param
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
ctx
.
Input
<
framework
::
Tensor
>
(
"Param"
));
auto
grad
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
ctx
.
Input
<
framework
::
Tensor
>
(
"Grad"
));
auto
moment1
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
ctx
.
Input
<
framework
::
Tensor
>
(
"Moment1"
));
auto
moment2
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
ctx
.
Input
<
framework
::
Tensor
>
(
"Moment2"
));
auto
lr
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
ctx
.
Input
<
framework
::
Tensor
>
(
"LearningRate"
));
auto
beta1_pow
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
ctx
.
Input
<
framework
::
Tensor
>
(
"Beta1Pow"
));
auto
beta2_pow
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
ctx
.
Input
<
framework
::
Tensor
>
(
"Beta2Pow"
));
auto
param_out
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
param_out_tensor
);
auto
moment1_out
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
moment1_out_tensor
);
auto
moment2_out
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
moment2_out_tensor
);
auto
*
place
=
ctx
.
template
device_context
<
DeviceContext
>().
eigen_device
();
moment1_out
.
device
(
*
place
)
=
beta1
*
moment1
+
(
1
-
beta1
)
*
grad
;
moment2_out
.
device
(
*
place
)
=
beta2
*
moment2
+
(
1
-
beta2
)
*
grad
.
square
();
// All of these are tensors of 1 element
auto
lr_t
=
lr
*
(
1
-
beta2_pow
).
sqrt
()
/
(
1
-
beta1_pow
);
// Eigen does not support automatic broadcast
// Get dimensions of moment vector to broadcast lr_t
Eigen
::
DSizes
<
int
,
1
>
m_dsize
(
moment1_out_tensor
->
numel
());
param_out
.
device
(
*
place
)
=
param
-
lr_t
.
broadcast
(
m_dsize
)
*
(
moment1_out
/
(
moment2_out
.
sqrt
()
+
epsilon
));
const
T
*
in_ptr
=
param
.
template
data
<
T
>();
T
*
out_ptr
=
param_out
.
template
mutable_data
<
T
>(
ctx
.
GetPlace
());
platform
::
Transform
<
DeviceContext
>
trans
;
trans
(
static_cast
<
const
DeviceContext
&>
(
ctx
.
device_context
()),
in_ptr
,
in_ptr
+
param_out
.
numel
(),
out_ptr
,
functor
);
}
};
...
...
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