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
a815d6ab
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
a815d6ab
编写于
9月 29, 2017
作者:
Z
zhouxiao-coder
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
elu: Optimize gradient calculation;Add more comments
上级
a2657fea
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
89 addition
and
0 deletion
+89
-0
paddle/operators/activation_op.cc
paddle/operators/activation_op.cc
+25
-0
paddle/operators/activation_op.cu
paddle/operators/activation_op.cu
+4
-0
paddle/operators/activation_op.h
paddle/operators/activation_op.h
+40
-0
python/paddle/v2/framework/tests/test_activation_op.py
python/paddle/v2/framework/tests/test_activation_op.py
+20
-0
未找到文件。
paddle/operators/activation_op.cc
浏览文件 @
a815d6ab
...
...
@@ -174,6 +174,25 @@ class SoftReluOpMaker : public framework::OpProtoAndCheckerMaker {
}
};
template
<
typename
AttrType
>
class
ELUOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
ELUOpMaker
(
framework
::
OpProto
*
proto
,
framework
::
OpAttrChecker
*
op_checker
)
:
OpProtoAndCheckerMaker
(
proto
,
op_checker
)
{
AddInput
(
"X"
,
"Input of ELU operator, it shouldn't be empty. Input is flattened "
"and treated as a 1D array."
);
AddOutput
(
"Y"
,
"Output of ELU operator, has same shape as the input."
);
AddComment
(
"ELU activation operator. It applies this element-wise computation on "
"the input: f(x) = max(0, x) + min(0, alpha * (exp(x) - 1))."
"Check .. _Link: https://arxiv.org/abs/1511.07289 for more details"
);
AddAttr
<
AttrType
>
(
"alpha"
,
"alpha value in the elu formulation, default to 1."
)
.
SetDefault
(
static_cast
<
AttrType
>
(
1.
));
}
};
template
<
typename
AttrType
>
class
PowOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
...
...
@@ -311,6 +330,12 @@ REGISTER_OP_CPU_KERNEL(soft_relu,
REGISTER_OP_CPU_KERNEL
(
soft_relu_grad
,
ops
::
SoftReluGradKernel
<
paddle
::
platform
::
CPUPlace
,
float
>
);
REGISTER_OP
(
elu
,
ops
::
ActivationOp
,
ops
::
ELUOpMaker
<
float
>
,
elu_grad
,
ops
::
ActivationOpGrad
);
REGISTER_OP_CPU_KERNEL
(
elu
,
ops
::
ELUKernel
<
paddle
::
platform
::
CPUPlace
,
float
>
);
REGISTER_OP_CPU_KERNEL
(
elu_grad
,
ops
::
ELUGradKernel
<
paddle
::
platform
::
CPUPlace
,
float
>
);
REGISTER_OP
(
pow
,
ops
::
ActivationOp
,
ops
::
PowOpMaker
<
float
>
,
pow_grad
,
ops
::
ActivationOpGrad
);
REGISTER_OP_CPU_KERNEL
(
pow
,
ops
::
PowKernel
<
paddle
::
platform
::
CPUPlace
,
float
>
);
...
...
paddle/operators/activation_op.cu
浏览文件 @
a815d6ab
...
...
@@ -97,6 +97,10 @@ REGISTER_OP_GPU_KERNEL(soft_relu,
REGISTER_OP_GPU_KERNEL
(
soft_relu_grad
,
ops
::
SoftReluGradKernel
<
paddle
::
platform
::
GPUPlace
,
float
>
);
REGISTER_OP_GPU_KERNEL
(
elu
,
ops
::
ELUKernel
<
paddle
::
platform
::
GPUPlace
,
float
>
);
REGISTER_OP_GPU_KERNEL
(
elu_grad
,
ops
::
ELUGradKernel
<
paddle
::
platform
::
GPUPlace
,
float
>
);
REGISTER_OP_GPU_KERNEL
(
pow
,
ops
::
PowKernel
<
paddle
::
platform
::
GPUPlace
,
float
>
);
REGISTER_OP_GPU_KERNEL
(
pow_grad
,
ops
::
PowGradKernel
<
paddle
::
platform
::
GPUPlace
,
float
>
);
...
...
paddle/operators/activation_op.h
浏览文件 @
a815d6ab
...
...
@@ -296,6 +296,46 @@ class SoftReluGradKernel : public framework::OpKernel<T> {
}
};
template
<
typename
Place
,
typename
T
,
typename
AttrType
=
T
>
class
ELUKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
auto
*
X
=
context
.
Input
<
framework
::
Tensor
>
(
"X"
);
auto
*
Y
=
context
.
Output
<
framework
::
Tensor
>
(
"Y"
);
auto
alpha
=
static_cast
<
T
>
(
context
.
Attr
<
AttrType
>
(
"alpha"
));
Y
->
mutable_data
<
T
>
(
context
.
GetPlace
());
auto
x
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
X
);
auto
y
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
Y
);
auto
place
=
context
.
GetEigenDevice
<
Place
>
();
y
.
device
(
place
)
=
x
.
cwiseMax
(
static_cast
<
T
>
(
0
))
+
(
alpha
*
(
x
.
exp
()
-
static_cast
<
T
>
(
1
))).
cwiseMin
(
static_cast
<
T
>
(
0
));
}
};
template
<
typename
Place
,
typename
T
,
typename
AttrType
=
T
>
class
ELUGradKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
auto
*
X
=
context
.
Input
<
framework
::
Tensor
>
(
"X"
);
auto
*
Y
=
context
.
Input
<
framework
::
Tensor
>
(
"Y"
);
auto
*
dY
=
context
.
Input
<
framework
::
Tensor
>
(
framework
::
GradVarName
(
"Y"
));
auto
*
dX
=
context
.
Output
<
framework
::
Tensor
>
(
framework
::
GradVarName
(
"X"
));
auto
alpha
=
static_cast
<
T
>
(
context
.
Attr
<
AttrType
>
(
"alpha"
));
dX
->
mutable_data
<
T
>
(
context
.
GetPlace
());
auto
x
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
X
);
auto
y
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
Y
);
auto
dy
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
dY
);
auto
dx
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
dX
);
auto
place
=
context
.
GetEigenDevice
<
Place
>
();
dx
.
device
(
place
)
=
dy
*
(
x
>
static_cast
<
T
>
(
0
)).
template
cast
<
T
>()
+
dy
*
(
y
+
alpha
)
*
(
x
<
static_cast
<
T
>
(
0
)).
template
cast
<
T
>();
}
};
template
<
typename
Place
,
typename
T
,
typename
AttrType
=
T
>
class
PowKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
...
...
python/paddle/v2/framework/tests/test_activation_op.py
浏览文件 @
a815d6ab
...
...
@@ -144,6 +144,26 @@ class TestSoftRelu(OpTest):
self
.
check_grad
([
'X'
],
'Y'
,
max_relative_error
=
0.02
)
class
TestELU
(
OpTest
):
def
setUp
(
self
):
self
.
op_type
=
"elu"
x
=
np
.
random
.
uniform
(
-
3
,
3
,
[
4
,
4
]).
astype
(
"float32"
)
alpha
=
1.
# Note: unlike other Relu extensions, point 0 on standard ELU function (i.e. alpha = 1)
# is differentiable, so we can skip modifications like x[np.abs(x) < 0.005] = 0.02 here
self
.
inputs
=
{
'X'
:
x
}
self
.
attrs
=
{
'alpha'
:
alpha
}
self
.
outputs
=
{
'Y'
:
np
.
maximum
(
0
,
x
)
+
np
.
minimum
(
0
,
alpha
*
(
np
.
exp
(
x
)
-
1
))
}
def
test_check_output
(
self
):
self
.
check_output
()
def
test_check_grad
(
self
):
self
.
check_grad
([
'X'
],
'Y'
,
max_relative_error
=
0.02
)
class
TestReciprocal
(
OpTest
):
def
setUp
(
self
):
self
.
op_type
=
"reciprocal"
...
...
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