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
马赛克医院
Paddle
提交
1e9fd407
P
Paddle
项目概览
马赛克医院
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
1
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
1e9fd407
编写于
3月 13, 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
combine op files
test=develop
上级
b26e9bd2
变更
7
隐藏空白更改
内联
并排
Showing
7 changed file
with
345 addition
and
447 deletion
+345
-447
paddle/fluid/operators/cross_entropy2_op.cc
paddle/fluid/operators/cross_entropy2_op.cc
+0
-137
paddle/fluid/operators/cross_entropy2_op.cu
paddle/fluid/operators/cross_entropy2_op.cu
+0
-29
paddle/fluid/operators/cross_entropy2_op.h
paddle/fluid/operators/cross_entropy2_op.h
+0
-110
paddle/fluid/operators/cross_entropy_op.cc
paddle/fluid/operators/cross_entropy_op.cc
+254
-2
paddle/fluid/operators/cross_entropy_op.cu
paddle/fluid/operators/cross_entropy_op.cu
+10
-0
paddle/fluid/operators/cross_entropy_op.h
paddle/fluid/operators/cross_entropy_op.h
+81
-0
paddle/fluid/operators/cross_entropy_op_base.h
paddle/fluid/operators/cross_entropy_op_base.h
+0
-169
未找到文件。
paddle/fluid/operators/cross_entropy2_op.cc
已删除
100644 → 0
浏览文件 @
b26e9bd2
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
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/fluid/operators/cross_entropy2_op.h"
#include <memory>
#include <string>
#include <unordered_map>
#include "paddle/fluid/operators/cross_entropy_op_base.h"
namespace
paddle
{
namespace
operators
{
class
CrossEntropyOp2
:
public
CrossEntropyOpBase
{
public:
using
CrossEntropyOpBase
::
CrossEntropyOpBase
;
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
CrossEntropyOpBase
::
InferShape
(
ctx
);
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
"XShape"
),
"Output(XShape) 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
));
ctx
->
ShareLoD
(
"X"
,
/*->*/
"XShape"
);
}
protected:
bool
IsSoftLabel
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
return
false
;
}
};
class
CrossEntropyGradientOp2
:
public
CrossEntropyGradientOpBase
{
public:
using
CrossEntropyGradientOpBase
::
CrossEntropyGradientOpBase
;
protected:
virtual
framework
::
DDim
GetXDim
(
framework
::
InferShapeContext
*
ctx
)
const
{
auto
x_shape
=
ctx
->
GetInputDim
(
"XShape"
);
return
framework
::
DDim
(
x_shape
.
Get
(),
x_shape
.
size
()
-
1
);
}
virtual
const
char
*
VarNameWithXLoD
()
const
{
return
"XShape"
;
}
virtual
bool
IsSoftLabel
(
framework
::
InferShapeContext
*
ctx
)
const
{
return
false
;
}
};
class
CrossEntropyOpMaker2
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
void
Make
()
override
{
AddInput
(
"X"
,
"(Tensor, default Tensor<float>), a tensor whose last dimension "
"size is equal to the number of classes. This input is a "
"probability computed by the previous operator, which is almost "
"always the result of a softmax operator."
);
AddInput
(
"Label"
,
"(Tensor), the tensor which represents the ground truth. It has the "
"same shape with 'X' except the last dimension. One hot Tensor."
);
AddOutput
(
"Y"
,
"(Tensor, default Tensor<float>), a tensor whose shape is same "
"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."
);
AddAttr
<
int
>
(
"ignore_index"
,
"(int, default -100), Specifies a target value that is"
"ignored and does not contribute to the input gradient."
"Only valid if soft_label is set to False"
)
.
SetDefault
(
-
100
);
AddComment
(
R"DOC(
Hard-label CrossEntropy Operator.
The input 'X' and 'Label' will first be logically flattened to 2-D matrixs.
The matrix's second dimension(row length) is as same as the original last
dimension, and the first dimension(column length) is the product of all other
original dimensions. Then the softmax computation will take palce on each raw
of flattened matrixs.
Only support hard label.
Both the input X and Label can carry the LoD (Level of Details) information,
or not. But the output only shares the LoD information with input X.
)DOC"
);
}
};
class
CrossEntropyGradOpMaker2
:
public
framework
::
SingleGradOpDescMaker
{
public:
using
framework
::
SingleGradOpDescMaker
::
SingleGradOpDescMaker
;
protected:
std
::
unique_ptr
<
framework
::
OpDesc
>
Apply
()
const
override
{
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
(
"XShape"
,
Output
(
"XShape"
));
op
->
SetInput
(
framework
::
GradVarName
(
"Y"
),
OutputGrad
(
"Y"
));
op
->
SetOutput
(
framework
::
GradVarName
(
"X"
),
InputGrad
(
"X"
));
op
->
SetAttrMap
(
Attrs
());
return
op
;
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
using
CPUCtx
=
paddle
::
platform
::
CPUDeviceContext
;
REGISTER_OPERATOR
(
cross_entropy2
,
ops
::
CrossEntropyOp2
,
ops
::
CrossEntropyOpMaker2
,
ops
::
CrossEntropyOpInferVarType
,
ops
::
CrossEntropyGradOpMaker2
);
REGISTER_OPERATOR
(
cross_entropy_grad2
,
ops
::
CrossEntropyGradientOp2
);
REGISTER_OP_CPU_KERNEL
(
cross_entropy2
,
ops
::
CrossEntropyOpKernel2
<
CPUCtx
,
float
>
,
ops
::
CrossEntropyOpKernel2
<
CPUCtx
,
double
>
);
REGISTER_OP_CPU_KERNEL
(
cross_entropy_grad2
,
ops
::
CrossEntropyGradientOpKernel2
<
CPUCtx
,
float
>
,
ops
::
CrossEntropyGradientOpKernel2
<
CPUCtx
,
double
>
);
paddle/fluid/operators/cross_entropy2_op.cu
已删除
100644 → 0
浏览文件 @
b26e9bd2
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
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/fluid/operators/cross_entropy2_op.h"
#include "paddle/fluid/platform/float16.h"
namespace
plat
=
paddle
::
platform
;
namespace
ops
=
paddle
::
operators
;
using
CUDACtx
=
paddle
::
platform
::
CUDADeviceContext
;
REGISTER_OP_CUDA_KERNEL
(
cross_entropy2
,
ops
::
CrossEntropyOpKernel2
<
CUDACtx
,
float
>
,
ops
::
CrossEntropyOpKernel2
<
CUDACtx
,
double
>
,
ops
::
CrossEntropyOpKernel2
<
CUDACtx
,
plat
::
float16
>
);
REGISTER_OP_CUDA_KERNEL
(
cross_entropy_grad2
,
ops
::
CrossEntropyGradientOpKernel2
<
CUDACtx
,
float
>
,
ops
::
CrossEntropyGradientOpKernel2
<
CUDACtx
,
double
>
,
ops
::
CrossEntropyGradientOpKernel2
<
CUDACtx
,
plat
::
float16
>
);
paddle/fluid/operators/cross_entropy2_op.h
已删除
100644 → 0
浏览文件 @
b26e9bd2
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
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 <cmath>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math.h"
#include "paddle/fluid/operators/math/cross_entropy.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/platform/for_range.h"
namespace
paddle
{
namespace
operators
{
using
Tensor
=
framework
::
Tensor
;
template
<
typename
T
>
struct
CrossEntropyBackwardFunctor
{
CrossEntropyBackwardFunctor
(
T
*
dx
,
const
T
*
y
,
const
T
*
dy
,
const
int64_t
*
label
,
int64_t
ignore_index
,
int64_t
feature_size
)
:
dx_
(
dx
),
y_
(
y
),
dy_
(
dy
),
label_
(
label
),
ignore_index_
(
ignore_index
),
feature_size_
(
feature_size
)
{}
HOSTDEVICE
void
operator
()(
int64_t
idx
)
const
{
auto
row_idx
=
idx
/
feature_size_
;
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
]);
}
else
{
dx_
[
idx
]
=
0
;
}
}
T
*
dx_
;
const
T
*
y_
;
const
T
*
dy_
;
const
int64_t
*
label_
;
int64_t
ignore_index_
;
int64_t
feature_size_
;
};
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
*
y
=
ctx
.
Output
<
Tensor
>
(
"Y"
);
y
->
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
);
}
};
template
<
typename
DeviceContext
,
typename
T
>
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
*
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_label
=
label
->
data
<
int64_t
>
();
int64_t
ignore_index
=
ctx
.
Attr
<
int
>
(
"ignore_index"
);
int
rank
=
dx
->
dims
().
size
();
int64_t
feature_size
=
dx
->
dims
()[
rank
-
1
];
int64_t
batch_size
=
framework
::
product
(
dx
->
dims
())
/
feature_size
;
platform
::
ForRange
<
DeviceContext
>
for_range
(
ctx
.
template
device_context
<
DeviceContext
>(),
batch_size
*
feature_size
);
for_range
(
CrossEntropyBackwardFunctor
<
T
>
(
p_dx
,
p_y
,
p_dy
,
p_label
,
ignore_index
,
feature_size
));
}
};
}
// namespace operators
}
// namespace paddle
paddle/fluid/operators/cross_entropy_op.cc
浏览文件 @
1e9fd407
...
...
@@ -14,11 +14,154 @@ limitations under the License. */
#include "paddle/fluid/operators/cross_entropy_op.h"
#include <string>
#include "paddle/fluid/operators/cross_entropy_op_base.h"
namespace
paddle
{
namespace
operators
{
class
CrossEntropyOpBase
:
public
framework
::
OperatorWithKernel
{
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"X"
),
"Input(X) should be not null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Label"
),
"Input(Label) should be not null."
);
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
"Y"
),
"Output(Y) should be not null."
);
auto
x_dims
=
ctx
->
GetInputDim
(
"X"
);
auto
label_dims
=
ctx
->
GetInputDim
(
"Label"
);
int
rank
=
x_dims
.
size
();
PADDLE_ENFORCE_EQ
(
rank
,
label_dims
.
size
(),
"Input(X) and Input(Label) shall have the same rank."
);
bool
check
=
true
;
if
((
!
ctx
->
IsRuntime
())
&&
(
framework
::
product
(
x_dims
)
<=
0
||
framework
::
product
(
label_dims
)
<=
0
))
{
check
=
false
;
}
if
(
check
)
{
PADDLE_ENFORCE_EQ
(
framework
::
slice_ddim
(
x_dims
,
0
,
rank
-
1
),
framework
::
slice_ddim
(
label_dims
,
0
,
rank
-
1
),
"Input(X) and Input(Label) shall have the same shape "
"except the last dimension."
);
}
if
(
IsSoftLabel
(
ctx
))
{
if
(
check
)
{
PADDLE_ENFORCE_EQ
(
x_dims
[
rank
-
1
],
label_dims
[
rank
-
1
],
"If Attr(soft_label) == true, the last dimension of "
"Input(X) and Input(Label) should be equal."
);
}
}
else
{
PADDLE_ENFORCE_EQ
(
label_dims
[
rank
-
1
],
1UL
,
"If Attr(softLabel) == false, the last dimension of "
"Input(Label) should be 1."
);
}
auto
y_dims
=
x_dims
;
y_dims
[
rank
-
1
]
=
1
;
ctx
->
SetOutputDim
(
"Y"
,
y_dims
);
ctx
->
ShareLoD
(
"X"
,
/*->*/
"Y"
);
}
protected:
// Explicitly set that the data type of computation kernel of cross_entropy
// is determined by its input "X".
framework
::
OpKernelType
GetExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
return
framework
::
OpKernelType
(
ctx
.
Input
<
Tensor
>
(
"X"
)
->
type
(),
ctx
.
device_context
());
}
virtual
bool
IsSoftLabel
(
framework
::
InferShapeContext
*
ctx
)
const
{
return
ctx
->
Attrs
().
Get
<
bool
>
(
"soft_label"
);
}
};
class
CrossEntropyGradientOpBase
:
public
framework
::
OperatorWithKernel
{
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
{
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Label"
),
"Input(Label) should be not null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
framework
::
GradVarName
(
"Y"
)),
"Input(Y@GRAD) shoudl be not null."
);
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
framework
::
GradVarName
(
"X"
)),
"Output(X@GRAD) should be not null."
);
auto
x_dims
=
GetXDim
(
ctx
);
auto
label_dims
=
ctx
->
GetInputDim
(
"Label"
);
auto
dy_dims
=
ctx
->
GetInputDim
(
framework
::
GradVarName
(
"Y"
));
int
rank
=
x_dims
.
size
();
PADDLE_ENFORCE_EQ
(
dy_dims
.
size
(),
rank
,
"Input(Y@Grad) and Input(X) should have the same rank."
);
PADDLE_ENFORCE_EQ
(
label_dims
.
size
(),
rank
,
"Input(Label) and Input(X) should have the same rank."
);
bool
check
=
true
;
if
((
!
ctx
->
IsRuntime
())
&&
(
framework
::
product
(
x_dims
)
<=
0
||
framework
::
product
(
label_dims
)
<=
0
))
{
check
=
false
;
}
if
(
check
)
{
PADDLE_ENFORCE_EQ
(
framework
::
slice_ddim
(
x_dims
,
0
,
rank
-
1
),
framework
::
slice_ddim
(
label_dims
,
0
,
rank
-
1
),
"The Input(X) and Input(Label) should have the same "
"shape except the last dimension."
);
PADDLE_ENFORCE_EQ
(
framework
::
slice_ddim
(
x_dims
,
0
,
rank
-
1
),
framework
::
slice_ddim
(
dy_dims
,
0
,
rank
-
1
),
"The Input(X) and Input(Y@Grad) should have the same "
"shape except the last dimension."
);
}
if
(
IsSoftLabel
(
ctx
))
{
if
(
check
)
{
PADDLE_ENFORCE_EQ
(
x_dims
[
rank
-
1
],
label_dims
[
rank
-
1
],
"When Attr(soft_label) == true, the last dimension of "
"Input(X) and Input(Label) should be equal."
);
}
}
else
{
PADDLE_ENFORCE_EQ
(
label_dims
[
rank
-
1
],
1
,
"When Attr(soft_label) == false, the last dimension of "
"Input(Label) should be 1."
);
}
ctx
->
SetOutputDim
(
framework
::
GradVarName
(
"X"
),
x_dims
);
PADDLE_ENFORCE_EQ
(
dy_dims
[
rank
-
1
],
1
,
"The last dimension of Input(Y@Grad) should be 1."
);
ctx
->
SetOutputDim
(
framework
::
GradVarName
(
"X"
),
x_dims
);
ctx
->
ShareLoD
(
VarNameWithXLoD
(),
framework
::
GradVarName
(
"X"
));
}
protected:
// Explicitly set that the data type of computation kernel of cross_entropy
// is determined by its input "X".
framework
::
OpKernelType
GetExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
return
framework
::
OpKernelType
(
ctx
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"Y"
))
->
type
(),
ctx
.
device_context
());
}
virtual
framework
::
DDim
GetXDim
(
framework
::
InferShapeContext
*
ctx
)
const
{
return
ctx
->
GetInputDim
(
"X"
);
}
virtual
const
char
*
VarNameWithXLoD
()
const
{
return
"X"
;
}
virtual
bool
IsSoftLabel
(
framework
::
InferShapeContext
*
ctx
)
const
{
return
ctx
->
Attrs
().
Get
<
bool
>
(
"soft_label"
);
}
};
class
CrossEntropyOpInferVarType
:
public
framework
::
PassInDtypeAndVarTypeToOutput
{
protected:
std
::
unordered_map
<
std
::
string
,
std
::
string
>
GetInputOutputWithSameType
()
const
override
{
return
std
::
unordered_map
<
std
::
string
,
std
::
string
>
{{
"X"
,
/*->*/
"Y"
}};
}
};
class
CrossEntropyOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
void
Make
()
override
{
...
...
@@ -87,12 +230,110 @@ class CrossEntropyGradientOp : public CrossEntropyGradientOpBase {
public:
using
CrossEntropyGradientOpBase
::
CrossEntropyGradientOpBase
;
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"X"
),
"Input(X) should be not null."
);
CrossEntropyGradientOpBase
::
InferShape
(
ctx
);
}
};
class
CrossEntropyOp2
:
public
CrossEntropyOpBase
{
public:
using
CrossEntropyOpBase
::
CrossEntropyOpBase
;
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
CrossEntropyOpBase
::
InferShape
(
ctx
);
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
"XShape"
),
"Output(XShape) 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
));
ctx
->
ShareLoD
(
"X"
,
/*->*/
"XShape"
);
}
protected:
bool
IsSoftLabel
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
return
false
;
}
};
class
CrossEntropyGradientOp2
:
public
CrossEntropyGradientOpBase
{
public:
using
CrossEntropyGradientOpBase
::
CrossEntropyGradientOpBase
;
protected:
virtual
framework
::
DDim
GetXDim
(
framework
::
InferShapeContext
*
ctx
)
const
{
auto
x_shape
=
ctx
->
GetInputDim
(
"XShape"
);
return
framework
::
DDim
(
x_shape
.
Get
(),
x_shape
.
size
()
-
1
);
}
virtual
const
char
*
VarNameWithXLoD
()
const
{
return
"XShape"
;
}
virtual
bool
IsSoftLabel
(
framework
::
InferShapeContext
*
ctx
)
const
{
return
false
;
}
};
class
CrossEntropyOpMaker2
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
void
Make
()
override
{
AddInput
(
"X"
,
"(Tensor, default Tensor<float>), a tensor whose last dimension "
"size is equal to the number of classes. This input is a "
"probability computed by the previous operator, which is almost "
"always the result of a softmax operator."
);
AddInput
(
"Label"
,
"(Tensor), the tensor which represents the ground truth. It has the "
"same shape with 'X' except the last dimension. One hot Tensor."
);
AddOutput
(
"Y"
,
"(Tensor, default Tensor<float>), a tensor whose shape is same "
"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."
);
AddAttr
<
int
>
(
"ignore_index"
,
"(int, default -100), Specifies a target value that is"
"ignored and does not contribute to the input gradient."
"Only valid if soft_label is set to False"
)
.
SetDefault
(
-
100
);
AddComment
(
R"DOC(
Hard-label CrossEntropy Operator.
The input 'X' and 'Label' will first be logically flattened to 2-D matrixs.
The matrix's second dimension(row length) is as same as the original last
dimension, and the first dimension(column length) is the product of all other
original dimensions. Then the softmax computation will take palce on each raw
of flattened matrixs.
Only support hard label.
Both the input X and Label can carry the LoD (Level of Details) information,
or not. But the output only shares the LoD information with input X.
)DOC"
);
}
};
class
CrossEntropyGradOpDescMaker2
:
public
framework
::
SingleGradOpDescMaker
{
public:
using
framework
::
SingleGradOpDescMaker
::
SingleGradOpDescMaker
;
protected:
std
::
unique_ptr
<
framework
::
OpDesc
>
Apply
()
const
override
{
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
(
"XShape"
,
Output
(
"XShape"
));
op
->
SetInput
(
framework
::
GradVarName
(
"Y"
),
OutputGrad
(
"Y"
));
op
->
SetOutput
(
framework
::
GradVarName
(
"X"
),
InputGrad
(
"X"
));
op
->
SetAttrMap
(
Attrs
());
return
op
;
}
};
}
// namespace operators
}
// namespace paddle
...
...
@@ -108,3 +349,14 @@ REGISTER_OP_CPU_KERNEL(cross_entropy, ops::CrossEntropyOpKernel<CPUCtx, float>,
REGISTER_OP_CPU_KERNEL
(
cross_entropy_grad
,
ops
::
CrossEntropyGradientOpKernel
<
CPUCtx
,
float
>
,
ops
::
CrossEntropyGradientOpKernel
<
CPUCtx
,
double
>
);
REGISTER_OPERATOR
(
cross_entropy2
,
ops
::
CrossEntropyOp2
,
ops
::
CrossEntropyOpMaker2
,
ops
::
CrossEntropyOpInferVarType
,
ops
::
CrossEntropyGradOpDescMaker2
);
REGISTER_OPERATOR
(
cross_entropy_grad2
,
ops
::
CrossEntropyGradientOp2
);
REGISTER_OP_CPU_KERNEL
(
cross_entropy2
,
ops
::
CrossEntropyOpKernel2
<
CPUCtx
,
float
>
,
ops
::
CrossEntropyOpKernel2
<
CPUCtx
,
double
>
);
REGISTER_OP_CPU_KERNEL
(
cross_entropy_grad2
,
ops
::
CrossEntropyGradientOpKernel2
<
CPUCtx
,
float
>
,
ops
::
CrossEntropyGradientOpKernel2
<
CPUCtx
,
double
>
);
paddle/fluid/operators/cross_entropy_op.cu
浏览文件 @
1e9fd407
...
...
@@ -27,3 +27,13 @@ REGISTER_OP_CUDA_KERNEL(
cross_entropy_grad
,
ops
::
CrossEntropyGradientOpKernel
<
CUDACtx
,
float
>
,
ops
::
CrossEntropyGradientOpKernel
<
CUDACtx
,
double
>
,
ops
::
CrossEntropyGradientOpKernel
<
CUDACtx
,
plat
::
float16
>
);
REGISTER_OP_CUDA_KERNEL
(
cross_entropy2
,
ops
::
CrossEntropyOpKernel2
<
CUDACtx
,
float
>
,
ops
::
CrossEntropyOpKernel2
<
CUDACtx
,
double
>
,
ops
::
CrossEntropyOpKernel2
<
CUDACtx
,
plat
::
float16
>
);
REGISTER_OP_CUDA_KERNEL
(
cross_entropy_grad2
,
ops
::
CrossEntropyGradientOpKernel2
<
CUDACtx
,
float
>
,
ops
::
CrossEntropyGradientOpKernel2
<
CUDACtx
,
double
>
,
ops
::
CrossEntropyGradientOpKernel2
<
CUDACtx
,
plat
::
float16
>
);
paddle/fluid/operators/cross_entropy_op.h
浏览文件 @
1e9fd407
...
...
@@ -15,6 +15,7 @@ limitations under the License. */
#pragma once
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math.h"
#include "paddle/fluid/operators/math/cross_entropy.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/platform/for_range.h"
...
...
@@ -137,5 +138,85 @@ class CrossEntropyGradientOpKernel : public framework::OpKernel<T> {
}
};
template
<
typename
T
>
struct
HardLabelCrossEntropyBackwardFunctor
{
HardLabelCrossEntropyBackwardFunctor
(
T
*
dx
,
const
T
*
y
,
const
T
*
dy
,
const
int64_t
*
label
,
int64_t
ignore_index
,
int64_t
feature_size
)
:
dx_
(
dx
),
y_
(
y
),
dy_
(
dy
),
label_
(
label
),
ignore_index_
(
ignore_index
),
feature_size_
(
feature_size
)
{}
HOSTDEVICE
void
operator
()(
int64_t
idx
)
const
{
auto
row_idx
=
idx
/
feature_size_
;
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
]);
}
else
{
dx_
[
idx
]
=
0
;
}
}
T
*
dx_
;
const
T
*
y_
;
const
T
*
dy_
;
const
int64_t
*
label_
;
int64_t
ignore_index_
;
int64_t
feature_size_
;
};
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
*
y
=
ctx
.
Output
<
Tensor
>
(
"Y"
);
y
->
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
);
}
};
template
<
typename
DeviceContext
,
typename
T
>
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
*
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_label
=
label
->
data
<
int64_t
>
();
int64_t
ignore_index
=
ctx
.
Attr
<
int
>
(
"ignore_index"
);
int
rank
=
dx
->
dims
().
size
();
int64_t
feature_size
=
dx
->
dims
()[
rank
-
1
];
int64_t
batch_size
=
framework
::
product
(
dx
->
dims
())
/
feature_size
;
platform
::
ForRange
<
DeviceContext
>
for_range
(
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
));
}
};
}
// namespace operators
}
// namespace paddle
paddle/fluid/operators/cross_entropy_op_base.h
已删除
100644 → 0
浏览文件 @
b26e9bd2
// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
//
// 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 <string>
#include <unordered_map>
#include "paddle/fluid/framework/op_registry.h"
namespace
paddle
{
namespace
operators
{
class
CrossEntropyOpBase
:
public
framework
::
OperatorWithKernel
{
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"X"
),
"Input(X) should be not null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Label"
),
"Input(Label) should be not null."
);
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
"Y"
),
"Output(Y) should be not null."
);
auto
x_dims
=
ctx
->
GetInputDim
(
"X"
);
auto
label_dims
=
ctx
->
GetInputDim
(
"Label"
);
int
rank
=
x_dims
.
size
();
PADDLE_ENFORCE_EQ
(
rank
,
label_dims
.
size
(),
"Input(X) and Input(Label) shall have the same rank."
);
bool
check
=
true
;
if
((
!
ctx
->
IsRuntime
())
&&
(
framework
::
product
(
x_dims
)
<=
0
||
framework
::
product
(
label_dims
)
<=
0
))
{
check
=
false
;
}
if
(
check
)
{
PADDLE_ENFORCE_EQ
(
framework
::
slice_ddim
(
x_dims
,
0
,
rank
-
1
),
framework
::
slice_ddim
(
label_dims
,
0
,
rank
-
1
),
"Input(X) and Input(Label) shall have the same shape "
"except the last dimension."
);
}
if
(
IsSoftLabel
(
ctx
))
{
if
(
check
)
{
PADDLE_ENFORCE_EQ
(
x_dims
[
rank
-
1
],
label_dims
[
rank
-
1
],
"If Attr(soft_label) == true, the last dimension of "
"Input(X) and Input(Label) should be equal."
);
}
}
else
{
PADDLE_ENFORCE_EQ
(
label_dims
[
rank
-
1
],
1UL
,
"If Attr(softLabel) == false, the last dimension of "
"Input(Label) should be 1."
);
}
auto
y_dims
=
x_dims
;
y_dims
[
rank
-
1
]
=
1
;
ctx
->
SetOutputDim
(
"Y"
,
y_dims
);
ctx
->
ShareLoD
(
"X"
,
/*->*/
"Y"
);
}
protected:
// Explicitly set that the data type of computation kernel of cross_entropy
// is determined by its input "X".
framework
::
OpKernelType
GetExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
return
framework
::
OpKernelType
(
ctx
.
Input
<
Tensor
>
(
"X"
)
->
type
(),
ctx
.
device_context
());
}
virtual
bool
IsSoftLabel
(
framework
::
InferShapeContext
*
ctx
)
const
{
return
ctx
->
Attrs
().
Get
<
bool
>
(
"soft_label"
);
}
};
class
CrossEntropyOpInferVarType
:
public
framework
::
PassInDtypeAndVarTypeToOutput
{
protected:
std
::
unordered_map
<
std
::
string
,
std
::
string
>
GetInputOutputWithSameType
()
const
override
{
return
std
::
unordered_map
<
std
::
string
,
std
::
string
>
{{
"X"
,
/*->*/
"Y"
}};
}
};
class
CrossEntropyGradientOpBase
:
public
framework
::
OperatorWithKernel
{
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
{
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Label"
),
"Input(Label) should be not null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
framework
::
GradVarName
(
"Y"
)),
"Input(Y@GRAD) shoudl be not null."
);
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
framework
::
GradVarName
(
"X"
)),
"Output(X@GRAD) should be not null."
);
auto
x_dims
=
GetXDim
(
ctx
);
auto
label_dims
=
ctx
->
GetInputDim
(
"Label"
);
auto
dy_dims
=
ctx
->
GetInputDim
(
framework
::
GradVarName
(
"Y"
));
int
rank
=
x_dims
.
size
();
PADDLE_ENFORCE_EQ
(
dy_dims
.
size
(),
rank
,
"Input(Y@Grad) and Input(X) should have the same rank."
);
PADDLE_ENFORCE_EQ
(
label_dims
.
size
(),
rank
,
"Input(Label) and Input(X) should have the same rank."
);
bool
check
=
true
;
if
((
!
ctx
->
IsRuntime
())
&&
(
framework
::
product
(
x_dims
)
<=
0
||
framework
::
product
(
label_dims
)
<=
0
))
{
check
=
false
;
}
if
(
check
)
{
PADDLE_ENFORCE_EQ
(
framework
::
slice_ddim
(
x_dims
,
0
,
rank
-
1
),
framework
::
slice_ddim
(
label_dims
,
0
,
rank
-
1
),
"The Input(X) and Input(Label) should have the same "
"shape except the last dimension."
);
PADDLE_ENFORCE_EQ
(
framework
::
slice_ddim
(
x_dims
,
0
,
rank
-
1
),
framework
::
slice_ddim
(
dy_dims
,
0
,
rank
-
1
),
"The Input(X) and Input(Y@Grad) should have the same "
"shape except the last dimension."
);
}
if
(
IsSoftLabel
(
ctx
))
{
if
(
check
)
{
PADDLE_ENFORCE_EQ
(
x_dims
[
rank
-
1
],
label_dims
[
rank
-
1
],
"When Attr(soft_label) == true, the last dimension of "
"Input(X) and Input(Label) should be equal."
);
}
}
else
{
PADDLE_ENFORCE_EQ
(
label_dims
[
rank
-
1
],
1
,
"When Attr(soft_label) == false, the last dimension of "
"Input(Label) should be 1."
);
}
ctx
->
SetOutputDim
(
framework
::
GradVarName
(
"X"
),
x_dims
);
PADDLE_ENFORCE_EQ
(
dy_dims
[
rank
-
1
],
1
,
"The last dimension of Input(Y@Grad) should be 1."
);
ctx
->
SetOutputDim
(
framework
::
GradVarName
(
"X"
),
x_dims
);
ctx
->
ShareLoD
(
VarNameWithXLoD
(),
framework
::
GradVarName
(
"X"
));
}
protected:
// Explicitly set that the data type of computation kernel of cross_entropy
// is determined by its input "X".
framework
::
OpKernelType
GetExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
return
framework
::
OpKernelType
(
ctx
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"Y"
))
->
type
(),
ctx
.
device_context
());
}
virtual
framework
::
DDim
GetXDim
(
framework
::
InferShapeContext
*
ctx
)
const
{
return
ctx
->
GetInputDim
(
"X"
);
}
virtual
const
char
*
VarNameWithXLoD
()
const
{
return
"X"
;
}
virtual
bool
IsSoftLabel
(
framework
::
InferShapeContext
*
ctx
)
const
{
return
ctx
->
Attrs
().
Get
<
bool
>
(
"soft_label"
);
}
};
}
// namespace operators
}
// namespace paddle
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