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
wmsofts
Paddle
提交
720b55cb
P
Paddle
项目概览
wmsofts
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
0
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
体验新版 GitCode,发现更多精彩内容 >>
提交
720b55cb
编写于
12月 17, 2018
作者:
T
tensor-tang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
enable crf decoding and layer norm refer code
上级
64a90b2f
变更
8
隐藏空白更改
内联
并排
Showing
8 changed file
with
119 addition
and
14 deletion
+119
-14
paddle/fluid/operators/crf_decoding_op.h
paddle/fluid/operators/crf_decoding_op.h
+4
-5
paddle/fluid/operators/jit/helper.cc
paddle/fluid/operators/jit/helper.cc
+4
-0
paddle/fluid/operators/jit/kernel_base.h
paddle/fluid/operators/jit/kernel_base.h
+18
-1
paddle/fluid/operators/jit/refer/CMakeLists.txt
paddle/fluid/operators/jit/refer/CMakeLists.txt
+2
-0
paddle/fluid/operators/jit/refer/refer.cc
paddle/fluid/operators/jit/refer/refer.cc
+3
-0
paddle/fluid/operators/jit/refer/refer.h
paddle/fluid/operators/jit/refer/refer.h
+80
-0
paddle/fluid/operators/jit/test.cc
paddle/fluid/operators/jit/test.cc
+1
-1
paddle/fluid/operators/layer_norm_op.h
paddle/fluid/operators/layer_norm_op.h
+7
-7
未找到文件。
paddle/fluid/operators/crf_decoding_op.h
浏览文件 @
720b55cb
...
...
@@ -16,7 +16,7 @@ limitations under the License. */
#include <limits>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/
math/jit_kernel
.h"
#include "paddle/fluid/operators/
jit/kernels
.h"
#include "paddle/fluid/operators/math/math_function.h"
namespace
paddle
{
...
...
@@ -82,10 +82,9 @@ class CRFDecodingOpKernel : public framework::OpKernel<T> {
Tensor
track
;
int
*
track_value
=
track
.
mutable_data
<
int
>
(
emission_dims
,
platform
::
CPUPlace
());
const
auto
&
ker
=
math
::
jitkernel
::
KernelPool
::
Instance
()
.
template
Get
<
math
::
jitkernel
::
CRFDecodeKernel
<
T
>
>
(
static_cast
<
int
>
(
tag_num
));
ker
->
Compute
(
static_cast
<
int
>
(
seq_len
),
x
,
w
,
alpha_value
,
track_value
);
auto
ker
=
jit
::
Get
<
jit
::
crfdecoding
,
jit
::
CRFDecoding
,
platform
::
CPUPlace
>
(
tag_num
);
ker
(
static_cast
<
int
>
(
seq_len
),
x
,
w
,
alpha_value
,
track_value
,
tag_num
);
T
max_score
=
-
std
::
numeric_limits
<
T
>::
max
();
int
max_i
=
0
;
for
(
size_t
i
=
0
;
i
<
tag_num
;
++
i
)
{
...
...
paddle/fluid/operators/jit/helper.cc
浏览文件 @
720b55cb
...
...
@@ -42,6 +42,8 @@ const char* to_string(KernelType kt) {
ONE_CASE
(
gruh1
);
ONE_CASE
(
gruhtpart1
);
ONE_CASE
(
gruhtpart2
);
ONE_CASE
(
crfdecoding
);
ONE_CASE
(
layernorm
);
default:
PADDLE_THROW
(
"Not support type: %d"
,
kt
);
return
"NOT JITKernel"
;
...
...
@@ -64,6 +66,8 @@ KernelType to_kerneltype(const std::string& act) {
}
else
if
(
lower
==
"tanh"
||
lower
==
"vtanh"
)
{
return
vtanh
;
}
PADDLE_THROW
(
"Not support type: %s, or forget to add this case"
,
act
);
return
non_kernel
;
}
...
...
paddle/fluid/operators/jit/kernel_base.h
浏览文件 @
720b55cb
...
...
@@ -37,7 +37,9 @@ typedef enum {
lstmc1h1
,
gruh1
,
gruhtpart1
,
gruhtpart2
gruhtpart2
,
crfdecoding
,
layernorm
}
KernelType
;
template
<
typename
T
>
...
...
@@ -109,6 +111,21 @@ struct GRUTuples {
typedef
void
(
*
func_type
)(
gru_t
*
,
const
gru_attr_t
*
);
};
template
<
typename
T
>
struct
CRFDecodingTuples
{
typedef
T
data_type
;
typedef
int
attr_type
;
typedef
void
(
*
func_type
)(
const
int
,
const
T
*
,
const
T
*
,
T
*
,
int
*
,
int
);
};
template
<
typename
T
>
struct
LayerNormTuples
{
typedef
T
data_type
;
typedef
int
attr_type
;
typedef
void
(
*
func_type
)(
T
*
,
T
*
,
T
*
,
T
*
,
const
T
*
,
const
T
*
,
int
,
const
float
,
int
);
};
// Just for adding to kernel pool without template
class
Kernel
{
public:
...
...
paddle/fluid/operators/jit/refer/CMakeLists.txt
浏览文件 @
720b55cb
...
...
@@ -23,3 +23,5 @@ USE_JITKERNEL_REFER(lstmc1h1)
USE_JITKERNEL_REFER
(
gruh1
)
USE_JITKERNEL_REFER
(
gruhtpart1
)
USE_JITKERNEL_REFER
(
gruhtpart2
)
USE_JITKERNEL_REFER
(
crfdecoding
)
USE_JITKERNEL_REFER
(
layernorm
)
paddle/fluid/operators/jit/refer/refer.cc
浏览文件 @
720b55cb
...
...
@@ -42,4 +42,7 @@ REGISTER_REFER_KERNEL(gruh1, GRUH1);
REGISTER_REFER_KERNEL
(
gruhtpart1
,
GRUHtPart1
);
REGISTER_REFER_KERNEL
(
gruhtpart2
,
GRUHtPart2
);
REGISTER_REFER_KERNEL
(
crfdecoding
,
CRFDecoding
);
REGISTER_REFER_KERNEL
(
layernorm
,
LayerNorm
);
#undef REGISTER_REFER_KERNEL
paddle/fluid/operators/jit/refer/refer.h
浏览文件 @
720b55cb
...
...
@@ -13,6 +13,9 @@
* limitations under the License. */
#pragma once
#include <cmath>
#include <limits>
#include "paddle/fluid/operators/jit/helper.h"
#include "paddle/fluid/operators/jit/kernel_base.h"
#include "paddle/fluid/platform/enforce.h"
...
...
@@ -242,6 +245,80 @@ void GRUHtPart2(gru_t* step, const gru_attr_t* attr) {
}
}
template
<
typename
T
>
void
CRFDecoding
(
const
int
seq_len
,
const
T
*
x
,
const
T
*
w
,
T
*
alpha
,
int
*
track
,
int
right
)
{
constexpr
int
state_trans_base_idx
=
2
;
for
(
int
i
=
0
;
i
<
right
;
++
i
)
{
alpha
[
i
]
=
w
[
i
]
+
x
[
i
];
}
for
(
int
k
=
1
;
k
<
seq_len
;
++
k
)
{
for
(
int
i
=
0
;
i
<
right
;
++
i
)
{
T
max_score
=
-
std
::
numeric_limits
<
T
>::
max
();
int
max_j
=
0
;
for
(
int
j
=
0
;
j
<
right
;
++
j
)
{
T
score
=
alpha
[(
k
-
1
)
*
right
+
j
]
+
w
[(
j
+
state_trans_base_idx
)
*
right
+
i
];
if
(
score
>
max_score
)
{
max_score
=
score
;
max_j
=
j
;
}
}
alpha
[
k
*
right
+
i
]
=
max_score
+
x
[
k
*
right
+
i
];
track
[
k
*
right
+
i
]
=
max_j
;
}
}
}
template
<
typename
T
>
void
LayerNorm
(
T
*
x
,
T
*
out
,
T
*
mean
,
T
*
var
,
const
T
*
scale
,
const
T
*
bias
,
int
height
,
const
float
epsilon
,
int
right
)
{
// get mean
for
(
int
i
=
0
;
i
<
height
;
i
++
)
{
T
sum
=
0.0
;
int
offset
=
i
*
right
;
for
(
int
j
=
0
;
j
<
right
;
j
++
)
{
sum
+=
x
[
offset
+
j
];
}
mean
[
i
]
=
sum
/
right
;
}
// get variance
for
(
int
i
=
0
;
i
<
height
;
i
++
)
{
T
sum
=
0.0
;
int
offset
=
i
*
right
;
for
(
int
j
=
0
;
j
<
right
;
j
++
)
{
sum
+=
(
x
[
offset
+
j
]
-
mean
[
i
])
*
(
x
[
offset
+
j
]
-
mean
[
i
]);
}
var
[
i
]
=
sum
/
right
;
}
for
(
int
i
=
0
;
i
<
height
;
i
++
)
{
int
offset
=
i
*
right
;
T
sqrt_var
=
std
::
sqrt
(
var
[
i
]
+
(
T
)
epsilon
);
for
(
int
j
=
0
;
j
<
right
;
j
++
)
{
out
[
offset
+
j
]
=
(
x
[
offset
+
j
]
-
mean
[
i
])
/
sqrt_var
;
}
}
if
(
scale
)
{
for
(
int
i
=
0
;
i
<
height
;
i
++
)
{
int
offset
=
i
*
right
;
for
(
int
j
=
0
;
j
<
right
;
j
++
)
{
out
[
offset
+
j
]
*=
scale
[
j
];
}
}
}
if
(
bias
)
{
for
(
int
i
=
0
;
i
<
height
;
i
++
)
{
int
offset
=
i
*
right
;
for
(
int
j
=
0
;
j
<
right
;
j
++
)
{
out
[
offset
+
j
]
+=
bias
[
j
];
}
}
}
}
#define DECLARE_REFER_KERNEL(name, tuples) \
template <typename T> \
class name##Kernel : public ReferKernel<tuples<T>> { \
...
...
@@ -275,6 +352,9 @@ DECLARE_REFER_KERNEL(GRUH1, GRUTuples);
DECLARE_REFER_KERNEL
(
GRUHtPart1
,
GRUTuples
);
DECLARE_REFER_KERNEL
(
GRUHtPart2
,
GRUTuples
);
DECLARE_REFER_KERNEL
(
CRFDecoding
,
CRFDecodingTuples
);
DECLARE_REFER_KERNEL
(
LayerNorm
,
LayerNormTuples
);
#undef DECLARE_REFER_KERNEL
}
// namespace refer
...
...
paddle/fluid/operators/jit/test.cc
浏览文件 @
720b55cb
...
...
@@ -515,7 +515,7 @@ TEST(JITKernel, gruhtpart2) {
TestGRUKernel
<
jit
::
gruhtpart2
,
double
,
paddle
::
platform
::
CPUPlace
>
();
}
// TODO(
TJ): refine the tests template
// TODO(
yihua/TJ): add crf decoding and layer norm unit tests
TEST
(
JITKernel
,
pool
)
{
// TODO(TJ): add some test
...
...
paddle/fluid/operators/layer_norm_op.h
浏览文件 @
720b55cb
...
...
@@ -19,7 +19,7 @@ limitations under the License. */
#include "paddle/fluid/operators/math/blas.h"
#if !defined(PADDLE_WITH_CUDA) && !defined(_WIN32) && !defined(__APPLE__) && \
!defined(__OSX__)
#include "paddle/fluid/operators/
math/jit_kernel
.h"
#include "paddle/fluid/operators/
jit/kernels
.h"
#endif
#include "paddle/fluid/operators/math/math_function.h"
...
...
@@ -229,12 +229,12 @@ class LayerNormKernel : public framework::OpKernel<T> {
PADDLE_ENFORCE_EQ
(
scale
->
numel
(),
right
);
PADDLE_ENFORCE_EQ
(
bias
->
numel
(),
right
);
const
auto
&
ker
=
math
::
jitkernel
::
KernelPool
::
Instance
()
.
template
Get
<
math
::
jitkernel
::
LayerNormKernel
<
T
>
>
(
static_cast
<
int
>
(
right
)
);
ker
->
Compute
(
x
.
data
<
T
>
(),
out
.
data
<
T
>
(),
mean
->
data
<
T
>
(),
var
->
data
<
T
>
(),
scale
->
data
<
T
>
(),
bias
->
data
<
T
>
(),
static_cast
<
int
>
(
left
),
static_cast
<
const
float
>
(
epsilon
)
);
auto
ker
=
jit
::
Get
<
jit
::
layernorm
,
jit
::
LayerNormTuples
,
platform
::
CPUPlace
>
(
right
);
ker
(
x
.
data
<
T
>
(),
out
.
data
<
T
>
(),
mean
->
data
<
T
>
(),
var
->
data
<
T
>
(),
scale
->
data
<
T
>
(),
bias
->
data
<
T
>
(),
static_cast
<
int
>
(
left
),
static_cast
<
const
float
>
(
epsilon
),
right
);
#endif
}
};
...
...
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