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
55ba7f61
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
55ba7f61
编写于
3月 14, 2019
作者:
S
sneaxiy
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
fix numeric error
test=develop
上级
487624e1
变更
4
显示空白变更内容
内联
并排
Showing
4 changed file
with
78 addition
and
22 deletion
+78
-22
paddle/fluid/operators/cross_entropy_op.cc
paddle/fluid/operators/cross_entropy_op.cc
+13
-1
paddle/fluid/operators/cross_entropy_op.h
paddle/fluid/operators/cross_entropy_op.h
+56
-17
python/paddle/fluid/layers/nn.py
python/paddle/fluid/layers/nn.py
+2
-0
python/paddle/fluid/tests/unittests/test_cross_entropy2_op.py
...on/paddle/fluid/tests/unittests/test_cross_entropy2_op.py
+7
-4
未找到文件。
paddle/fluid/operators/cross_entropy_op.cc
浏览文件 @
55ba7f61
...
...
@@ -248,10 +248,15 @@ class CrossEntropyOp2 : public CrossEntropyOpBase {
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
"XShape"
),
"Output(XShape) should be not null."
);
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
"MatchX"
),
"Output(MatchX) should be not null."
);
auto
x_dims
=
ctx
->
GetInputDim
(
"X"
);
auto
x_dims_vec
=
framework
::
vectorize
(
x_dims
);
x_dims_vec
.
push_back
(
0
);
ctx
->
SetOutputDim
(
"XShape"
,
framework
::
make_ddim
(
x_dims_vec
));
x_dims
[
x_dims
.
size
()
-
1
]
=
1
;
ctx
->
SetOutputDim
(
"MatchX"
,
x_dims
);
ctx
->
ShareLoD
(
"X"
,
/*->*/
"XShape"
);
}
...
...
@@ -265,6 +270,11 @@ class CrossEntropyGradientOp2 : public CrossEntropyGradientOpBase {
public:
using
CrossEntropyGradientOpBase
::
CrossEntropyGradientOpBase
;
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"MatchX"
),
"Input(MatchX) must exist"
);
CrossEntropyGradientOpBase
::
InferShape
(
ctx
);
}
protected:
virtual
framework
::
DDim
GetXDim
(
framework
::
InferShapeContext
*
ctx
)
const
{
auto
x_shape
=
ctx
->
GetInputDim
(
"XShape"
);
...
...
@@ -295,6 +305,8 @@ class CrossEntropyOpMaker2 : public framework::OpProtoAndCheckerMaker {
"with 'X' except that the last dimension size is 1. It "
"represents the cross entropy loss."
);
AddOutput
(
"XShape"
,
"Temporaily variable to save shape and LoD of X."
);
AddOutput
(
"MatchX"
,
"X value that matches label, used for gradient computation."
);
AddAttr
<
int
>
(
"ignore_index"
,
"(int, default -100), Specifies a target value that is"
"ignored and does not contribute to the input gradient."
...
...
@@ -327,7 +339,7 @@ class CrossEntropyGradOpDescMaker2 : public framework::SingleGradOpDescMaker {
std
::
unique_ptr
<
framework
::
OpDesc
>
op
(
new
framework
::
OpDesc
());
op
->
SetType
(
"cross_entropy_grad2"
);
op
->
SetInput
(
"Label"
,
Input
(
"Label"
));
op
->
SetInput
(
"
Y"
,
Output
(
"Y
"
));
op
->
SetInput
(
"
MatchX"
,
Output
(
"MatchX
"
));
op
->
SetInput
(
"XShape"
,
Output
(
"XShape"
));
op
->
SetInput
(
framework
::
GradVarName
(
"Y"
),
OutputGrad
(
"Y"
));
op
->
SetOutput
(
framework
::
GradVarName
(
"X"
),
InputGrad
(
"X"
));
...
...
paddle/fluid/operators/cross_entropy_op.h
浏览文件 @
55ba7f61
...
...
@@ -138,15 +138,48 @@ class CrossEntropyGradientOpKernel : public framework::OpKernel<T> {
}
};
template
<
typename
T
>
struct
HardLabelCrossEntropyForwardFunctor
{
HardLabelCrossEntropyForwardFunctor
(
const
T
*
x
,
T
*
y
,
T
*
match_x
,
const
int64_t
*
label
,
int64_t
ignore_index
,
int64_t
feature_size
)
:
x_
(
x
),
y_
(
y
),
match_x_
(
match_x
),
label_
(
label
),
ignore_index_
(
ignore_index
),
feature_size_
(
feature_size
)
{}
HOSTDEVICE
void
operator
()(
int64_t
idx
)
const
{
auto
label
=
label_
[
idx
];
if
(
label
!=
ignore_index_
)
{
auto
match_x
=
x_
[
idx
*
feature_size_
+
label
];
y_
[
idx
]
=
-
math
::
TolerableValue
<
T
>
()(
real_log
(
match_x
));
match_x_
[
idx
]
=
match_x
;
}
else
{
y_
[
idx
]
=
0
;
match_x_
[
idx
]
=
0
;
// any value is ok
}
}
const
T
*
x_
;
T
*
y_
;
T
*
match_x_
;
const
int64_t
*
label_
;
int64_t
ignore_index_
;
int64_t
feature_size_
;
};
template
<
typename
T
>
struct
HardLabelCrossEntropyBackwardFunctor
{
HardLabelCrossEntropyBackwardFunctor
(
T
*
dx
,
const
T
*
y
,
const
T
*
dy
,
HardLabelCrossEntropyBackwardFunctor
(
T
*
dx
,
const
T
*
dy
,
const
T
*
match_x
,
const
int64_t
*
label
,
int64_t
ignore_index
,
int64_t
feature_size
)
:
dx_
(
dx
),
y_
(
y
),
dy_
(
dy
),
match_x_
(
match_x
),
label_
(
label
),
ignore_index_
(
ignore_index
),
feature_size_
(
feature_size
)
{}
...
...
@@ -156,15 +189,15 @@ struct HardLabelCrossEntropyBackwardFunctor {
auto
col_idx
=
idx
%
feature_size_
;
auto
label
=
label_
[
row_idx
];
if
(
label
==
col_idx
&&
label
!=
ignore_index_
)
{
dx_
[
idx
]
=
-
dy_
[
row_idx
]
*
real_exp
(
y_
[
row_idx
])
;
dx_
[
idx
]
=
-
dy_
[
row_idx
]
/
match_x_
[
row_idx
]
;
}
else
{
dx_
[
idx
]
=
0
;
}
}
T
*
dx_
;
const
T
*
y_
;
const
T
*
dy_
;
const
T
*
match_x_
;
const
int64_t
*
label_
;
int64_t
ignore_index_
;
int64_t
feature_size_
;
...
...
@@ -174,20 +207,26 @@ template <typename DeviceContext, typename T>
class
CrossEntropyOpKernel2
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
*
x_original
=
ctx
.
Input
<
Tensor
>
(
"X"
);
int
rank
=
x_original
->
dims
().
size
();
auto
x
=
framework
::
ReshapeToMatrix
(
*
x_original
,
rank
-
1
);
auto
label
=
framework
::
ReshapeToMatrix
(
*
ctx
.
Input
<
Tensor
>
(
"Label"
),
rank
-
1
);
auto
*
x
=
ctx
.
Input
<
Tensor
>
(
"X"
);
auto
*
label
=
ctx
.
Input
<
Tensor
>
(
"Label"
);
auto
*
y
=
ctx
.
Output
<
Tensor
>
(
"Y"
);
y
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
*
match_x
=
ctx
.
Output
<
Tensor
>
(
"MatchX"
);
auto
&
x_dims
=
x
->
dims
();
auto
feature_size
=
x_dims
[
x_dims
.
size
()
-
1
];
auto
batch_size
=
framework
::
product
(
x
->
dims
())
/
feature_size
;
auto
*
p_x
=
x
->
data
<
T
>
();
auto
*
p_label
=
label
->
data
<
int64_t
>
();
auto
*
p_y
=
y
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
*
p_match_x
=
match_x
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
ignore_index
=
ctx
.
Attr
<
int
>
(
"ignore_index"
);
math
::
CrossEntropyFunctor
<
DeviceContext
,
T
>
()(
ctx
.
template
device_context
<
DeviceContext
>(),
y
,
&
x
,
&
label
,
false
,
ignore_index
);
platform
::
ForRange
<
DeviceContext
>
for_range
(
ctx
.
template
device_context
<
DeviceContext
>(),
batch_size
);
for_range
(
HardLabelCrossEntropyForwardFunctor
<
T
>
(
p_x
,
p_y
,
p_match_x
,
p_label
,
ignore_index
,
feature_size
));
}
};
...
...
@@ -196,13 +235,13 @@ class CrossEntropyGradientOpKernel2 : public framework::OpKernel<T> {
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
*
dx
=
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"X"
));
auto
*
y
=
ctx
.
Input
<
Tensor
>
(
"Y"
);
auto
*
dy
=
ctx
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"Y"
));
auto
*
match_x
=
ctx
.
Input
<
Tensor
>
(
"MatchX"
);
auto
*
label
=
ctx
.
Input
<
Tensor
>
(
"Label"
);
auto
*
p_dx
=
dx
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
*
p_y
=
y
->
data
<
T
>
();
auto
*
p_dy
=
dy
->
data
<
T
>
();
auto
*
p_match_x
=
match_x
->
data
<
T
>
();
auto
*
p_label
=
label
->
data
<
int64_t
>
();
int64_t
ignore_index
=
ctx
.
Attr
<
int
>
(
"ignore_index"
);
...
...
@@ -214,7 +253,7 @@ class CrossEntropyGradientOpKernel2 : public framework::OpKernel<T> {
ctx
.
template
device_context
<
DeviceContext
>(),
batch_size
*
feature_size
);
for_range
(
HardLabelCrossEntropyBackwardFunctor
<
T
>
(
p_dx
,
p_
y
,
p_dy
,
p_label
,
ignore_index
,
feature_size
));
p_dx
,
p_
dy
,
p_match_x
,
p_label
,
ignore_index
,
feature_size
));
}
};
...
...
python/paddle/fluid/layers/nn.py
浏览文件 @
55ba7f61
...
...
@@ -1450,11 +1450,13 @@ def cross_entropy2(input, label, ignore_index=kIgnoreIndex):
helper
=
LayerHelper
(
'cross_entropy2'
,
**
locals
())
out
=
helper
.
create_variable_for_type_inference
(
dtype
=
input
.
dtype
)
xshape
=
helper
.
create_variable_for_type_inference
(
dtype
=
input
.
dtype
)
match_x
=
helper
.
create_variable_for_type_inference
(
dtype
=
input
.
dtype
)
helper
.
append_op
(
type
=
'cross_entropy2'
,
inputs
=
{
'X'
:
[
input
],
'Label'
:
[
label
]},
outputs
=
{
'Y'
:
[
out
],
'MatchX'
:
[
match_x
],
'XShape'
:
[
xshape
]},
attrs
=
{
'ignore_index'
:
ignore_index
})
return
out
...
...
python/paddle/fluid/tests/unittests/test_cross_entropy2_op.py
浏览文件 @
55ba7f61
...
...
@@ -24,11 +24,13 @@ class CrossEntropy2OpTestBase(OpTest):
def
calc_output
(
self
,
logits
,
label
,
ignore_index
):
ret
=
np
.
zeros
(
shape
=
label
.
shape
,
dtype
=
logits
.
dtype
)
match_x
=
np
.
zeros
(
shape
=
label
.
shape
,
dtype
=
logits
.
dtype
)
for
idx
in
six
.
moves
.
range
(
label
.
shape
[
0
]):
if
label
[
idx
]
==
ignore_index
:
continue
ret
[
idx
]
=
-
np
.
log
(
logits
[
idx
][
label
[
idx
]])
return
ret
match_x
[
idx
]
=
logits
[
idx
][
label
[
idx
]]
ret
[
idx
]
=
-
np
.
log
(
match_x
[
idx
])
return
ret
,
match_x
def
setUp
(
self
):
self
.
shape
,
self
.
dtype
,
self
.
ignore_index
=
self
.
initParameters
()
...
...
@@ -39,12 +41,13 @@ class CrossEntropy2OpTestBase(OpTest):
label
=
np
.
random
.
random_integers
(
low
=
0
,
high
=
feature_size
-
1
,
size
=
self
.
shape
[
0
:
-
1
]
+
[
1
]).
astype
(
'int64'
)
outputs
=
self
.
calc_output
(
outputs
,
match_x
=
self
.
calc_output
(
np
.
reshape
(
logits
,
[
batch_size
,
feature_size
]),
np
.
reshape
(
label
,
[
batch_size
,
1
]),
self
.
ignore_index
)
self
.
inputs
=
{
'X'
:
logits
,
'Label'
:
label
}
self
.
outputs
=
{
'Y'
:
np
.
reshape
(
outputs
,
label
.
shape
),
'MatchX'
:
np
.
reshape
(
match_x
,
label
.
shape
),
'XShape'
:
np
.
zeros
(
shape
=
logits
.
shape
,
dtype
=
logits
.
dtype
)
}
...
...
@@ -57,7 +60,7 @@ class CrossEntropy2OpTestBase(OpTest):
self
.
check_grad
(
inputs_to_check
=
[
'X'
],
output_names
=
[
'Y'
],
no_grad_set
=
[
'XShape'
,
'Label'
])
no_grad_set
=
[
'XShape'
,
'
MatchX'
,
'
Label'
])
class
CrossEntropy2OpTest2
(
CrossEntropy2OpTestBase
):
...
...
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