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
94697b05
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 >>
未验证
提交
94697b05
编写于
1月 30, 2018
作者:
C
Cao Ying
提交者:
GitHub
1月 30, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #7960 from kuke/label_smooth_op
Add label smooth operator.
上级
db8da025
a10caf7c
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
275 addition
and
0 deletion
+275
-0
paddle/operators/label_smooth_op.cc
paddle/operators/label_smooth_op.cc
+128
-0
paddle/operators/label_smooth_op.cu
paddle/operators/label_smooth_op.cu
+26
-0
paddle/operators/label_smooth_op.h
paddle/operators/label_smooth_op.h
+66
-0
python/paddle/v2/fluid/tests/test_label_smooth_op.py
python/paddle/v2/fluid/tests/test_label_smooth_op.py
+55
-0
未找到文件。
paddle/operators/label_smooth_op.cc
0 → 100644
浏览文件 @
94697b05
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/operators/label_smooth_op.h"
namespace
paddle
{
namespace
operators
{
class
LabelSmoothOp
:
public
framework
::
OperatorWithKernel
{
public:
LabelSmoothOp
(
const
std
::
string
&
type
,
const
framework
::
VariableNameMap
&
inputs
,
const
framework
::
VariableNameMap
&
outputs
,
const
framework
::
AttributeMap
&
attrs
)
:
OperatorWithKernel
(
type
,
inputs
,
outputs
,
attrs
)
{}
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"X"
),
"Input(X) of LabelSmoothOp should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
"Out"
),
"Output(Out) of LabelSmoothOp should not be null."
);
auto
in_dims
=
ctx
->
GetInputDim
(
"X"
);
if
(
ctx
->
HasInput
(
"PriorDist"
))
{
auto
noise_dims
=
ctx
->
GetInputDim
(
"PriorDist"
);
auto
noise_numel
=
paddle
::
framework
::
product
(
noise_dims
);
PADDLE_ENFORCE
(
in_dims
[
1
]
==
noise_numel
,
"The number of elements in Input(PriorDist) must be equal to the "
"dimension of each label."
);
}
ctx
->
ShareLoD
(
"X"
,
/*->*/
"Out"
);
ctx
->
SetOutputDim
(
"Out"
,
in_dims
);
}
};
class
LabelSmoothOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
LabelSmoothOpMaker
(
OpProto
*
proto
,
OpAttrChecker
*
op_checker
)
:
OpProtoAndCheckerMaker
(
proto
,
op_checker
)
{
AddInput
(
"X"
,
"(LoDTensor) The input labels of LabelSmooth operator. This "
"input can be batched labels in one-hot encoding or output from "
"softmax, with shape [N x K], where N is the batch size and K is "
"the number of classes"
);
AddInput
(
"PriorDist"
,
"(Tensor, optional)"
"The prior distribution to be added to the smoothed label. It is "
"fixed during training and the number of elements should be equal "
"to the dimension K of each label. Default is uniform "
"distribution and each element will be set to 1/K if not provided "
"in input."
)
.
AsDispensable
();
AddOutput
(
"Out"
,
"(loDTensor) The smoothed label of LabelSmooth operator. It has"
"the same shape and LoD with the Input(LoDTensor)."
);
AddAttr
<
float
>
(
"epsilon"
,
"(float, default 0.0f)"
"The smoothing parameter of LabelSmooth operator."
)
.
SetDefault
(
0.0
f
);
AddComment
(
R"DOC(
LabelSmooth Operator.
Label smoothing is a mechanism to regularize the classifier layer. In machine
learning, optimizing the log-likelihood of the correct label directly may
cause two problems. First, it may result in overfitting: if the model learns
to assign full probability to the ground-truth label for each training example,
it is not guaranteed to generalize. Second, it encourages the differences
between the largest logit and all others to become large, reducing the ability
of the model to adapt. Label smoothing is proposed to encourage the model to
be less confident, which replaces the ground-truth label $y$ with the weighted
sum of itself and some fixed distribution $\mu$, i.e.
$$
\tilde{y} = (1 - \epsilon) * y + \epsilon * \mu,
$$
where $(1 - \epsilon)$ and $\epsilon$ are the weights respectively, and
$\tilde{y}$ is the smoothed label. Usually uniform distribution is used for
$\mu$. This change in the ground-truth label is called label-smoothing
regularization or LSR.
See more details about label smoothing in https://arxiv.org/abs/1512.00567.
)DOC"
);
}
};
class
LabelSmoothGradOp
:
public
framework
::
OperatorWithKernel
{
public:
LabelSmoothGradOp
(
const
std
::
string
&
type
,
const
framework
::
VariableNameMap
&
inputs
,
const
framework
::
VariableNameMap
&
outputs
,
const
framework
::
AttributeMap
&
attrs
)
:
OperatorWithKernel
(
type
,
inputs
,
outputs
,
attrs
)
{}
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"X"
),
"Input(X) shouldn't be null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
framework
::
GradVarName
(
"Out"
)),
"Input(Out@GRAD) shouldn't be null."
);
ctx
->
SetOutputDim
(
framework
::
GradVarName
(
"X"
),
ctx
->
GetInputDim
(
"X"
));
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OP
(
label_smooth
,
ops
::
LabelSmoothOp
,
ops
::
LabelSmoothOpMaker
,
label_smooth_grad
,
ops
::
LabelSmoothGradOp
);
REGISTER_OP_CPU_KERNEL
(
label_smooth
,
ops
::
LabelSmoothKernel
<
paddle
::
platform
::
CPUDeviceContext
,
float
>
,
ops
::
LabelSmoothKernel
<
paddle
::
platform
::
CPUDeviceContext
,
double
>
);
REGISTER_OP_CPU_KERNEL
(
label_smooth_grad
,
ops
::
LabelSmoothGradKernel
<
paddle
::
platform
::
CPUDeviceContext
,
float
>
,
ops
::
LabelSmoothGradKernel
<
paddle
::
platform
::
CPUDeviceContext
,
double
>
);
paddle/operators/label_smooth_op.cu
0 → 100644
浏览文件 @
94697b05
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/operators/label_smooth_op.h"
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_CUDA_KERNEL
(
label_smooth
,
ops
::
LabelSmoothKernel
<
paddle
::
platform
::
CUDADeviceContext
,
float
>
,
ops
::
LabelSmoothKernel
<
paddle
::
platform
::
CUDADeviceContext
,
double
>
);
REGISTER_OP_CUDA_KERNEL
(
label_smooth_grad
,
ops
::
LabelSmoothGradKernel
<
paddle
::
platform
::
CUDADeviceContext
,
float
>
,
ops
::
LabelSmoothGradKernel
<
paddle
::
platform
::
CUDADeviceContext
,
double
>
);
paddle/operators/label_smooth_op.h
0 → 100644
浏览文件 @
94697b05
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
namespace
paddle
{
namespace
operators
{
template
<
typename
DeviceContext
,
typename
T
>
class
LabelSmoothKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
{
auto
*
out_t
=
ctx
.
Output
<
framework
::
LoDTensor
>
(
"Out"
);
auto
*
in_t
=
ctx
.
Input
<
framework
::
LoDTensor
>
(
"X"
);
auto
*
dist_t
=
ctx
.
Input
<
framework
::
Tensor
>
(
"PriorDist"
);
auto
label_dim
=
in_t
->
dims
()[
1
];
out_t
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
epsilon
=
ctx
.
Attr
<
float
>
(
"epsilon"
);
auto
out
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
out_t
);
auto
in
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
in_t
);
auto
&
dev
=
*
ctx
.
template
device_context
<
DeviceContext
>().
eigen_device
();
if
(
dist_t
)
{
auto
dist
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
dist_t
);
out
.
device
(
dev
)
=
static_cast
<
T
>
(
1
-
epsilon
)
*
in
+
epsilon
*
dist
.
broadcast
(
Eigen
::
DSizes
<
int
,
1
>
(
in_t
->
numel
()));
}
else
{
out
.
device
(
dev
)
=
static_cast
<
T
>
(
1
-
epsilon
)
*
in
+
static_cast
<
T
>
(
epsilon
/
label_dim
);
}
}
};
template
<
typename
DeviceContext
,
typename
T
>
class
LabelSmoothGradKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
{
auto
*
d_out_t
=
ctx
.
Input
<
framework
::
Tensor
>
(
framework
::
GradVarName
(
"Out"
));
auto
*
d_in_t
=
ctx
.
Output
<
framework
::
Tensor
>
(
framework
::
GradVarName
(
"X"
));
d_in_t
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
d_out
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
d_out_t
);
auto
d_in
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
d_in_t
);
auto
epsilon
=
ctx
.
Attr
<
float
>
(
"epsilon"
);
auto
&
dev
=
*
ctx
.
template
device_context
<
DeviceContext
>().
eigen_device
();
d_in
.
device
(
dev
)
=
static_cast
<
T
>
(
1
-
epsilon
)
*
d_out
;
}
};
}
// namespace operators
}
// namespace paddle
python/paddle/v2/fluid/tests/test_label_smooth_op.py
0 → 100644
浏览文件 @
94697b05
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import
unittest
import
numpy
as
np
from
op_test
import
OpTest
class
TestLabelSmoothOp
(
OpTest
):
def
config
(
self
):
self
.
op_type
=
"label_smooth"
self
.
epsilon
=
0.1
batch_size
,
self
.
label_dim
=
5
,
10
self
.
label
=
np
.
zeros
((
batch_size
,
self
.
label_dim
)).
astype
(
"float64"
)
nonzero_index
=
np
.
random
.
randint
(
self
.
label_dim
,
size
=
(
batch_size
))
self
.
label
[
np
.
arange
(
batch_size
),
nonzero_index
]
=
1
def
setUp
(
self
):
self
.
config
()
smoothed_label
=
(
1
-
self
.
epsilon
)
*
self
.
label
+
self
.
epsilon
/
self
.
label_dim
self
.
inputs
=
{
'X'
:
self
.
label
}
self
.
attrs
=
{
'epsilon'
:
self
.
epsilon
}
self
.
outputs
=
{
'Out'
:
smoothed_label
}
def
test_check_output
(
self
):
self
.
check_output
()
def
test_check_grad
(
self
):
self
.
check_grad
([
"X"
],
"Out"
)
class
TestLabelSmoothOpWithPriorDist
(
TestLabelSmoothOp
):
def
setUp
(
self
):
self
.
config
()
dist
=
np
.
random
.
random
((
1
,
self
.
label_dim
))
smoothed_label
=
(
1
-
self
.
epsilon
)
*
self
.
label
+
self
.
epsilon
*
dist
self
.
inputs
=
{
'X'
:
self
.
label
,
'PriorDist'
:
dist
}
self
.
attrs
=
{
'epsilon'
:
self
.
epsilon
}
self
.
outputs
=
{
'Out'
:
smoothed_label
}
if
__name__
==
'__main__'
:
unittest
.
main
()
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