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
5ed07ef1
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2299
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
提交
5ed07ef1
编写于
1月 30, 2018
作者:
Y
Yibing Liu
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Add more comments and enable the distribution's outside setting
上级
fcff9758
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
79 addition
and
13 deletion
+79
-13
paddle/operators/label_smooth_op.cc
paddle/operators/label_smooth_op.cc
+46
-2
paddle/operators/label_smooth_op.h
paddle/operators/label_smooth_op.h
+10
-2
python/paddle/v2/fluid/tests/test_label_smooth_op.py
python/paddle/v2/fluid/tests/test_label_smooth_op.py
+23
-9
未找到文件。
paddle/operators/label_smooth_op.cc
浏览文件 @
5ed07ef1
...
...
@@ -31,6 +31,14 @@ class LabelSmoothOp : public framework::OperatorWithKernel {
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
);
}
...
...
@@ -40,8 +48,22 @@ class LabelSmoothOpMaker : public framework::OpProtoAndCheckerMaker {
public:
LabelSmoothOpMaker
(
OpProto
*
proto
,
OpAttrChecker
*
op_checker
)
:
OpProtoAndCheckerMaker
(
proto
,
op_checker
)
{
AddInput
(
"X"
,
"The input label of LabelSmooth operator."
);
AddOutput
(
"Out"
,
"The smoothed label of LabelSmooth operator."
);
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."
)
...
...
@@ -49,6 +71,28 @@ class LabelSmoothOpMaker : public framework::OpProtoAndCheckerMaker {
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 itselft 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"
);
}
};
...
...
paddle/operators/label_smooth_op.h
浏览文件 @
5ed07ef1
...
...
@@ -26,6 +26,7 @@ class LabelSmoothKernel : public framework::OpKernel<T> {
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
());
...
...
@@ -33,8 +34,15 @@ class LabelSmoothKernel : public framework::OpKernel<T> {
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
();
out
.
device
(
dev
)
=
static_cast
<
T
>
(
1
-
epsilon
)
*
in
+
static_cast
<
T
>
(
epsilon
/
label_dim
);
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
);
}
}
};
...
...
python/paddle/v2/fluid/tests/test_label_smooth_op.py
浏览文件 @
5ed07ef1
...
...
@@ -18,16 +18,20 @@ from op_test import OpTest
class
TestLabelSmoothOp
(
OpTest
):
def
setUp
(
self
):
def
config
(
self
):
self
.
op_type
=
"label_smooth"
epsilon
=
0.1
batch_size
,
label_dim
=
5
,
10
label
=
np
.
zeros
((
batch_size
,
label_dim
)).
astype
(
"float64"
)
nonzero_index
=
np
.
random
.
randint
(
label_dim
,
size
=
(
batch_size
))
label
[
np
.
arange
(
batch_size
),
nonzero_index
]
=
1
smoothed_label
=
(
1
-
epsilon
)
*
label
+
epsilon
/
label_dim
self
.
inputs
=
{
'X'
:
label
}
self
.
attrs
=
{
'epsilon'
:
epsilon
}
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
):
...
...
@@ -37,5 +41,15 @@ class TestLabelSmoothOp(OpTest):
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