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
urhero
Paddle
提交
d0521e6f
P
Paddle
项目概览
urhero
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
0
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
d0521e6f
编写于
9月 14, 2017
作者:
L
Luo Tao
浏览文件
操作
浏览文件
下载
差异文件
Merge branch 'develop' into use_op
上级
f657e21f
0e46f5eb
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
231 addition
and
95 deletion
+231
-95
paddle/operators/cos_sim_op.cc
paddle/operators/cos_sim_op.cc
+69
-26
paddle/operators/cos_sim_op.h
paddle/operators/cos_sim_op.h
+92
-55
paddle/operators/onehot_cross_entropy_op.cc
paddle/operators/onehot_cross_entropy_op.cc
+1
-1
python/paddle/v2/framework/tests/test_cos_sim_op.py
python/paddle/v2/framework/tests/test_cos_sim_op.py
+59
-5
python/paddle/v2/framework/tests/test_cross_entropy_op.py
python/paddle/v2/framework/tests/test_cross_entropy_op.py
+9
-7
python/paddle/v2/framework/tests/test_pad_op.py
python/paddle/v2/framework/tests/test_pad_op.py
+1
-1
未找到文件。
paddle/operators/cos_sim_op.cc
浏览文件 @
d0521e6f
...
@@ -25,16 +25,30 @@ class CosSimOp : public framework::OperatorWithKernel {
...
@@ -25,16 +25,30 @@ class CosSimOp : public framework::OperatorWithKernel {
protected:
protected:
void
InferShape
(
const
framework
::
InferShapeContext
&
ctx
)
const
override
{
void
InferShape
(
const
framework
::
InferShapeContext
&
ctx
)
const
override
{
// notnull check
PADDLE_ENFORCE_NOT_NULL
(
ctx
.
InputVar
(
"X"
),
"Input(X) must not be null."
);
PADDLE_ENFORCE_NOT_NULL
(
ctx
.
InputVar
(
"X"
),
"Input(X) must not be null."
);
PADDLE_ENFORCE_NOT_NULL
(
ctx
.
InputVar
(
"Y"
),
"Input(Y) must not be null."
);
PADDLE_ENFORCE_NOT_NULL
(
ctx
.
InputVar
(
"Y"
),
"Input(Y) must not be null."
);
PADDLE_ENFORCE_EQ
(
ctx
.
Input
<
Tensor
>
(
"X"
)
->
dims
(),
ctx
.
Input
<
Tensor
>
(
"Y"
)
->
dims
(),
// shape check
"Dimensions of Input(X) and Input(Y) must be the same."
);
auto
x_dims
=
ctx
.
Input
<
Tensor
>
(
"X"
)
->
dims
();
auto
y_dims
=
ctx
.
Input
<
Tensor
>
(
"Y"
)
->
dims
();
auto
dims
=
ctx
.
Input
<
Tensor
>
(
"X"
)
->
dims
();
ctx
.
Output
<
Tensor
>
(
"Out"
)
->
Resize
({
dims
[
0
],
1
});
PADDLE_ENFORCE_EQ
(
x_dims
.
size
(),
y_dims
.
size
(),
ctx
.
Output
<
Tensor
>
(
"XNorm"
)
->
Resize
({
dims
[
0
],
1
});
"Ranks of Input(X) and Input(Y) must be equal."
);
ctx
.
Output
<
Tensor
>
(
"YNorm"
)
->
Resize
({
dims
[
0
],
1
});
PADDLE_ENFORCE_GE
(
x_dims
.
size
(),
2
,
"Rank of Input(X) must not be less than 2."
);
PADDLE_ENFORCE_EQ
(
framework
::
slice_ddim
(
x_dims
,
1
,
x_dims
.
size
()),
framework
::
slice_ddim
(
y_dims
,
1
,
y_dims
.
size
()),
"All dimensions except the 1st of Input(X) and Input(Y) "
"must be equal."
);
PADDLE_ENFORCE
(
x_dims
[
0
]
==
y_dims
[
0
]
||
y_dims
[
0
]
==
1
,
"The 1st dimension of Input(Y) must be equal to Input(X) or"
" just 1 (which will be broadcasted to match Input(X))."
);
// resize tensor
ctx
.
Output
<
Tensor
>
(
"Out"
)
->
Resize
({
x_dims
[
0
],
1
});
ctx
.
Output
<
Tensor
>
(
"XNorm"
)
->
Resize
({
x_dims
[
0
],
1
});
ctx
.
Output
<
Tensor
>
(
"YNorm"
)
->
Resize
({
y_dims
[
0
],
1
});
}
}
};
};
...
@@ -42,16 +56,27 @@ class CosSimOpMaker : public framework::OpProtoAndCheckerMaker {
...
@@ -42,16 +56,27 @@ class CosSimOpMaker : public framework::OpProtoAndCheckerMaker {
public:
public:
CosSimOpMaker
(
framework
::
OpProto
*
proto
,
framework
::
OpAttrChecker
*
op_checker
)
CosSimOpMaker
(
framework
::
OpProto
*
proto
,
framework
::
OpAttrChecker
*
op_checker
)
:
OpProtoAndCheckerMaker
(
proto
,
op_checker
)
{
:
OpProtoAndCheckerMaker
(
proto
,
op_checker
)
{
AddInput
(
"X"
,
"The
fir
st input of cos_sim op."
);
AddInput
(
"X"
,
"The
1
st input of cos_sim op."
);
AddInput
(
"Y"
,
"The
seco
nd input of cos_sim op."
);
AddInput
(
"Y"
,
"The
2
nd input of cos_sim op."
);
AddOutput
(
"Out"
,
"The output of cos_sim op."
);
AddOutput
(
"Out"
,
"The output of cos_sim op."
);
AddOutput
(
"XNorm"
,
"Row norm of the first input."
).
AsIntermediate
();
AddOutput
(
"XNorm"
,
AddOutput
(
"YNorm"
,
"Row norm of the second input."
).
AsIntermediate
();
"Norm of the first input, reduced along the 1st "
"dimension."
)
.
AsIntermediate
();
AddOutput
(
"YNorm"
,
"Norm of the second input, reduced along the 1st "
"dimension."
)
.
AsIntermediate
();
AddComment
(
R"DOC(
AddComment
(
R"DOC(
Cosine Similarity Operator.
Cosine Similarity Operator.
The equation is: Out = X^T * Y / (sqrt(X^T * X) * sqrt(Y^T * Y))
The equation is: Out = X^T * Y / (sqrt(X^T * X) * sqrt(Y^T * Y)).
Input(X) and Input(Y) must have the same shape, except that the 1st dimension
of Input(Y) could be just 1 (different from Input(X)), which will be
broadcasted to match the shape of Input(X) before computing their cosine
similarity.
)DOC"
);
)DOC"
);
}
}
};
};
...
@@ -62,32 +87,50 @@ class CosSimOpGrad : public framework::OperatorWithKernel {
...
@@ -62,32 +87,50 @@ class CosSimOpGrad : public framework::OperatorWithKernel {
protected:
protected:
void
InferShape
(
const
framework
::
InferShapeContext
&
ctx
)
const
override
{
void
InferShape
(
const
framework
::
InferShapeContext
&
ctx
)
const
override
{
// notnull check
PADDLE_ENFORCE_NOT_NULL
(
ctx
.
InputVar
(
"X"
),
"Input(X) must not be null."
);
PADDLE_ENFORCE_NOT_NULL
(
ctx
.
InputVar
(
"X"
),
"Input(X) must not be null."
);
PADDLE_ENFORCE_NOT_NULL
(
ctx
.
InputVar
(
"Y"
),
"Input(Y) must not be null."
);
PADDLE_ENFORCE_NOT_NULL
(
ctx
.
InputVar
(
"Y"
),
"Input(Y) must not be null."
);
PADDLE_ENFORCE_NOT_NULL
(
ctx
.
InputVar
(
"XNorm"
),
PADDLE_ENFORCE_NOT_NULL
(
ctx
.
InputVar
(
"XNorm"
),
"Input(XNorm) must not be null."
);
"Input(XNorm) must not be null."
);
PADDLE_ENFORCE_NOT_NULL
(
ctx
.
InputVar
(
"YNorm"
),
PADDLE_ENFORCE_NOT_NULL
(
ctx
.
InputVar
(
"YNorm"
),
"Input(YNorm) must not be null."
);
"Input(YNorm) must not be null."
);
PADDLE_ENFORCE_NOT_NULL
(
ctx
.
InputVar
(
"Out"
),
"Input(Out) must not be null."
);
PADDLE_ENFORCE_NOT_NULL
(
ctx
.
InputVar
(
framework
::
GradVarName
(
"Out"
)),
PADDLE_ENFORCE_NOT_NULL
(
ctx
.
InputVar
(
framework
::
GradVarName
(
"Out"
)),
"Input(Out@GRAD) must not be null."
);
"Input(Out@GRAD) must not be null."
);
// shape check
auto
x_dims
=
ctx
.
Input
<
Tensor
>
(
"X"
)
->
dims
();
auto
x_dims
=
ctx
.
Input
<
Tensor
>
(
"X"
)
->
dims
();
auto
y_dims
=
ctx
.
Input
<
Tensor
>
(
"Y"
)
->
dims
();
auto
y_dims
=
ctx
.
Input
<
Tensor
>
(
"Y"
)
->
dims
();
auto
xnorm_dims
=
ctx
.
Input
<
Tensor
>
(
"XNorm"
)
->
dims
();
auto
xnorm_dims
=
ctx
.
Input
<
Tensor
>
(
"XNorm"
)
->
dims
();
auto
ynorm_dims
=
ctx
.
Input
<
Tensor
>
(
"YNorm"
)
->
dims
();
auto
ynorm_dims
=
ctx
.
Input
<
Tensor
>
(
"YNorm"
)
->
dims
();
auto
out_dims
=
ctx
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"Out"
))
->
dims
();
auto
out_dims
=
ctx
.
Input
<
Tensor
>
(
"Out"
)
->
dims
();
PADDLE_ENFORCE_EQ
(
x_dims
,
y_dims
,
auto
out_grad_dims
=
"Dimensions of Input(X) and Input(Y) must be the same."
);
ctx
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"Out"
))
->
dims
();
PADDLE_ENFORCE_EQ
(
xnorm_dims
[
0
],
x_dims
[
0
],
"1st dimension of XNorm must equal that of Input(X)."
);
PADDLE_ENFORCE_GE
(
x_dims
.
size
(),
y_dims
.
size
(),
PADDLE_ENFORCE_EQ
(
xnorm_dims
[
1
],
1
,
"2st dimension of XNorm must be one."
);
"Ranks of Input(X) and Input(Y) must be equal."
);
PADDLE_ENFORCE_EQ
(
ynorm_dims
[
0
],
y_dims
[
0
],
PADDLE_ENFORCE_GE
(
x_dims
.
size
(),
2
,
"1st dimension of YNorm must equal that of Input(Y)."
);
"Rank of Input(X) must not be less than 2."
);
PADDLE_ENFORCE_EQ
(
ynorm_dims
[
1
],
1
,
"2st dimension of YNorm must be one."
);
PADDLE_ENFORCE_EQ
(
framework
::
slice_ddim
(
x_dims
,
1
,
x_dims
.
size
()),
PADDLE_ENFORCE_EQ
(
out_dims
[
0
],
x_dims
[
0
],
framework
::
slice_ddim
(
y_dims
,
1
,
y_dims
.
size
()),
"1st dimension of Out@GRAD must equal that of Input(X)"
);
"All dimensions except the 1st of Input(X) and Input(Y) "
PADDLE_ENFORCE_EQ
(
out_dims
[
1
],
1
,
"1st dimension of Out@GRAD must be one."
);
"must be equal."
);
PADDLE_ENFORCE
(
x_dims
[
0
]
==
y_dims
[
0
]
||
y_dims
[
0
]
==
1
,
"The 1st dimension of Input(Y) must be equal to Input(X) or"
" just 1 (which will be broadcasted to match Input(X))."
);
auto
target_xnorm_dims
=
framework
::
make_ddim
({
x_dims
[
0
],
1
});
auto
target_ynorm_dims
=
framework
::
make_ddim
({
y_dims
[
0
],
1
});
PADDLE_ENFORCE_EQ
(
xnorm_dims
,
target_xnorm_dims
,
"Shape of Input(XNorm) must be [X.Dim(0), 1]."
);
PADDLE_ENFORCE_EQ
(
ynorm_dims
,
target_ynorm_dims
,
"Shape of Input(YNorm) must be [Y.Dim(0), 1]."
);
PADDLE_ENFORCE_EQ
(
out_dims
,
target_xnorm_dims
,
"Shape of Input(Out) must be [X.Dim(0), 1]."
);
PADDLE_ENFORCE_EQ
(
out_grad_dims
,
target_xnorm_dims
,
"Shape of Input(Out@Grad) must be [X.Dim(0), 1]."
);
// resize tensor
auto
*
x_grad
=
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"X"
));
auto
*
x_grad
=
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"X"
));
auto
*
y_grad
=
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"Y"
));
auto
*
y_grad
=
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"Y"
));
if
(
x_grad
)
x_grad
->
Resize
(
x_dims
);
if
(
x_grad
)
x_grad
->
Resize
(
x_dims
);
...
...
paddle/operators/cos_sim_op.h
浏览文件 @
d0521e6f
...
@@ -31,30 +31,38 @@ template <typename Place, typename T>
...
@@ -31,30 +31,38 @@ template <typename Place, typename T>
class
CosSimKernel
:
public
framework
::
OpKernel
{
class
CosSimKernel
:
public
framework
::
OpKernel
{
public:
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
auto
*
input_x
=
context
.
Input
<
Tensor
>
(
"X"
);
// get Tensor
auto
*
input_y
=
context
.
Input
<
Tensor
>
(
"Y"
);
auto
*
in_x
=
context
.
Input
<
Tensor
>
(
"X"
);
auto
*
output_z
=
context
.
Output
<
Tensor
>
(
"Out"
);
auto
*
in_y
=
context
.
Input
<
Tensor
>
(
"Y"
);
auto
*
output_x_norm
=
context
.
Output
<
Tensor
>
(
"XNorm"
);
auto
*
out_z
=
context
.
Output
<
Tensor
>
(
"Out"
);
auto
*
output_y_norm
=
context
.
Output
<
Tensor
>
(
"YNorm"
);
auto
*
out_x_norm
=
context
.
Output
<
Tensor
>
(
"XNorm"
);
auto
*
out_y_norm
=
context
.
Output
<
Tensor
>
(
"YNorm"
);
out_z
->
mutable_data
<
T
>
(
context
.
GetPlace
());
out_x_norm
->
mutable_data
<
T
>
(
context
.
GetPlace
());
out_y_norm
->
mutable_data
<
T
>
(
context
.
GetPlace
());
output_z
->
mutable_data
<
T
>
(
context
.
GetPlace
());
// convert Tensor to Eigen Tensor
output_x_norm
->
mutable_data
<
T
>
(
context
.
GetPlace
());
int
rows_x
=
in_x
->
dims
()[
0
];
output_y_norm
->
mutable_data
<
T
>
(
context
.
GetPlace
());
int
rows_y
=
in_y
->
dims
()[
0
];
auto
x
=
EigenMatrix
<
T
>::
Reshape
(
*
in_x
,
1
);
auto
dims
=
input_x
->
dims
();
auto
y
=
EigenMatrix
<
T
>::
Reshape
(
*
in_y
,
1
);
int64_t
size
=
input_x
->
numel
();
auto
z
=
EigenVector
<
T
>::
Flatten
(
*
out_z
);
auto
new_dims
=
framework
::
make_ddim
({
dims
[
0
],
size
/
dims
[
0
]});
auto
x_norm
=
EigenVector
<
T
>::
Flatten
(
*
out_x_norm
);
auto
x
=
EigenMatrix
<
T
>::
From
(
*
input_x
,
new_dims
);
auto
y_norm
=
EigenVector
<
T
>::
Flatten
(
*
out_y_norm
);
auto
y
=
EigenMatrix
<
T
>::
From
(
*
input_y
,
new_dims
);
auto
z
=
EigenVector
<
T
>::
Flatten
(
*
output_z
);
auto
x_norm
=
EigenVector
<
T
>::
Flatten
(
*
output_x_norm
);
auto
y_norm
=
EigenVector
<
T
>::
Flatten
(
*
output_y_norm
);
// compute
auto
place
=
context
.
GetEigenDevice
<
Place
>
();
auto
place
=
context
.
GetEigenDevice
<
Place
>
();
auto
xy
=
(
x
*
y
).
sum
(
Eigen
::
array
<
int
,
1
>
({{
1
}}));
auto
row_along
=
Eigen
::
array
<
int
,
1
>
({{
1
}});
x_norm
.
device
(
place
)
=
x
.
square
().
sum
(
Eigen
::
array
<
int
,
1
>
({{
1
}})).
sqrt
();
x_norm
.
device
(
place
)
=
x
.
square
().
sum
(
row_along
).
sqrt
();
y_norm
.
device
(
place
)
=
y
.
square
().
sum
(
Eigen
::
array
<
int
,
1
>
({{
1
}})).
sqrt
();
y_norm
.
device
(
place
)
=
y
.
square
().
sum
(
row_along
).
sqrt
();
z
.
device
(
place
)
=
xy
/
x_norm
/
y_norm
;
if
(
rows_x
==
rows_y
)
{
auto
xy
=
(
x
*
y
).
sum
(
Eigen
::
array
<
int
,
1
>
({
1
}));
z
.
device
(
place
)
=
xy
/
x_norm
/
y_norm
;
}
else
{
Eigen
::
DSizes
<
int
,
2
>
bcast
(
rows_x
,
1
);
auto
xy
=
(
x
*
y
.
broadcast
(
bcast
)).
sum
(
row_along
);
z
.
device
(
place
)
=
xy
/
x_norm
/
y_norm
.
broadcast
(
bcast
);
}
}
}
};
};
...
@@ -62,43 +70,72 @@ template <typename Place, typename T>
...
@@ -62,43 +70,72 @@ template <typename Place, typename T>
class
CosSimGradKernel
:
public
framework
::
OpKernel
{
class
CosSimGradKernel
:
public
framework
::
OpKernel
{
public:
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
auto
*
input_x
=
context
.
Input
<
Tensor
>
(
"X"
);
// get Tensor
auto
*
input_y
=
context
.
Input
<
Tensor
>
(
"Y"
);
auto
*
in_x
=
context
.
Input
<
Tensor
>
(
"X"
);
auto
*
input_z
=
context
.
Input
<
Tensor
>
(
"Out"
);
auto
*
in_y
=
context
.
Input
<
Tensor
>
(
"Y"
);
auto
*
input_x_norm
=
context
.
Input
<
Tensor
>
(
"XNorm"
);
auto
*
in_z
=
context
.
Input
<
Tensor
>
(
"Out"
);
auto
*
input_y_norm
=
context
.
Input
<
Tensor
>
(
"YNorm"
);
auto
*
in_x_norm
=
context
.
Input
<
Tensor
>
(
"XNorm"
);
auto
*
output_grad_x
=
context
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"X"
));
auto
*
in_y_norm
=
context
.
Input
<
Tensor
>
(
"YNorm"
);
auto
*
output_grad_y
=
context
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"Y"
));
auto
*
out_grad_x
=
context
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"X"
));
auto
*
input_grad_z
=
context
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"Out"
));
auto
*
out_grad_y
=
context
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"Y"
));
auto
*
in_grad_z
=
context
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"Out"
));
auto
dims
=
input_x
->
dims
();
// convert Tensor to Eigen Tensor
int64_t
size
=
input_x
->
numel
();
auto
x
=
EigenMatrix
<
T
>::
Reshape
(
*
in_x
,
1
);
auto
new_dims
=
framework
::
make_ddim
({
dims
[
0
],
size
/
dims
[
0
]});
auto
y
=
EigenMatrix
<
T
>::
Reshape
(
*
in_y
,
1
);
auto
x
=
EigenMatrix
<
T
>::
From
(
*
input_x
,
new_dims
);
auto
z
=
EigenMatrix
<
T
>::
Reshape
(
*
in_z
,
1
);
auto
y
=
EigenMatrix
<
T
>::
From
(
*
input_y
,
new_dims
);
auto
x_norm
=
EigenMatrix
<
T
>::
Reshape
(
*
in_x_norm
,
1
);
auto
z
=
EigenMatrix
<
T
>::
From
(
*
input_z
);
auto
y_norm
=
EigenMatrix
<
T
>::
Reshape
(
*
in_y_norm
,
1
);
auto
x_norm
=
EigenMatrix
<
T
>::
From
(
*
input_x_norm
);
auto
dz
=
EigenMatrix
<
T
>::
Reshape
(
*
in_grad_z
,
1
);
auto
y_norm
=
EigenMatrix
<
T
>::
From
(
*
input_y_norm
);
auto
dz
=
EigenMatrix
<
T
>::
From
(
*
input_grad_z
);
Eigen
::
DSizes
<
int
,
2
>
bcast
(
1
,
new_dims
[
1
]);
// compute gradident
auto
z_bcast
=
z
.
broadcast
(
bcast
);
int
rows_x
=
in_x
->
dims
()[
0
];
auto
dz_bcast
=
dz
.
broadcast
(
bcast
);
int
rows_y
=
in_y
->
dims
()[
0
];
int
cols
=
framework
::
product
(
in_x
->
dims
())
/
rows_x
;
Eigen
::
DSizes
<
int
,
2
>
bcast_cols
(
1
,
cols
);
auto
z_bcast
=
z
.
broadcast
(
bcast_cols
);
auto
dz_bcast
=
dz
.
broadcast
(
bcast_cols
);
auto
x_snorm_bcast
=
x_norm
.
square
().
eval
().
broadcast
(
bcast_cols
);
auto
place
=
context
.
GetEigenDevice
<
Place
>
();
auto
place
=
context
.
GetEigenDevice
<
Place
>
();
auto
x_snorm_bcast
=
x_norm
.
square
().
eval
().
broadcast
(
bcast
);
if
(
rows_x
==
rows_y
)
{
auto
y_snorm_bcast
=
y_norm
.
square
().
eval
().
broadcast
(
bcast
);
auto
y_snorm_bcast
=
y_norm
.
square
().
eval
().
broadcast
(
bcast_cols
);
auto
norm_prod_bcast
=
(
x_norm
*
y_norm
).
eval
().
broadcast
(
bcast
);
auto
norm_prod_bcast
=
(
x_norm
*
y_norm
).
eval
().
broadcast
(
bcast_cols
);
if
(
output_grad_x
)
{
// compute dx
output_grad_x
->
mutable_data
<
T
>
(
context
.
GetPlace
());
if
(
out_grad_x
)
{
auto
dx
=
EigenMatrix
<
T
>::
From
(
*
output_grad_x
,
new_dims
);
out_grad_x
->
mutable_data
<
T
>
(
context
.
GetPlace
());
dx
.
device
(
place
)
=
auto
dx
=
EigenMatrix
<
T
>::
Reshape
(
*
out_grad_x
,
1
);
dz_bcast
*
(
y
/
norm_prod_bcast
-
z_bcast
*
x
/
x_snorm_bcast
);
auto
grad
=
y
/
norm_prod_bcast
-
z_bcast
*
x
/
x_snorm_bcast
;
}
dx
.
device
(
place
)
=
dz_bcast
*
grad
;
if
(
output_grad_y
)
{
}
output_grad_y
->
mutable_data
<
T
>
(
context
.
GetPlace
());
// compute dy
auto
dy
=
EigenMatrix
<
T
>::
From
(
*
output_grad_y
,
new_dims
);
if
(
out_grad_y
)
{
dy
.
device
(
place
)
=
out_grad_y
->
mutable_data
<
T
>
(
context
.
GetPlace
());
dz_bcast
*
(
x
/
norm_prod_bcast
-
z_bcast
*
y
/
y_snorm_bcast
);
auto
dy
=
EigenMatrix
<
T
>::
Reshape
(
*
out_grad_y
,
1
);
auto
grad
=
x
/
norm_prod_bcast
-
z_bcast
*
y
/
y_snorm_bcast
;
dy
.
device
(
place
)
=
dz_bcast
*
grad
;
}
}
else
{
Eigen
::
DSizes
<
int
,
2
>
bcast_rows
(
rows_x
,
1
);
Eigen
::
DSizes
<
int
,
2
>
bcast_rows_cols
(
rows_x
,
cols
);
auto
y_bcast
=
y
.
broadcast
(
bcast_rows
);
auto
y_snorm_bcast
=
y_norm
.
square
().
eval
().
broadcast
(
bcast_rows_cols
);
auto
norm_prod_bcast
=
(
x_norm
*
y_norm
.
eval
().
broadcast
(
bcast_rows
))
.
eval
()
.
broadcast
(
bcast_cols
);
// compute dx
if
(
out_grad_x
)
{
out_grad_x
->
mutable_data
<
T
>
(
context
.
GetPlace
());
auto
dx
=
EigenMatrix
<
T
>::
Reshape
(
*
out_grad_x
,
1
);
auto
grad
=
y_bcast
/
norm_prod_bcast
-
z_bcast
*
x
/
x_snorm_bcast
;
dx
.
device
(
place
)
=
dz_bcast
*
grad
;
}
// compute dy
if
(
out_grad_y
)
{
out_grad_y
->
mutable_data
<
T
>
(
context
.
GetPlace
());
auto
dy
=
EigenMatrix
<
T
>::
Reshape
(
*
out_grad_y
,
1
);
auto
grad
=
x
/
norm_prod_bcast
-
z_bcast
*
y_bcast
/
y_snorm_bcast
;
dy
.
device
(
place
)
=
(
dz_bcast
*
grad
).
sum
(
Eigen
::
array
<
int
,
1
>
({
0
}));
}
}
}
}
}
};
};
...
...
paddle/operators/onehot_cross_entropy_op.cc
浏览文件 @
d0521e6f
...
@@ -29,7 +29,7 @@ class OnehotCrossEntropyOp : public framework::OperatorWithKernel {
...
@@ -29,7 +29,7 @@ class OnehotCrossEntropyOp : public framework::OperatorWithKernel {
PADDLE_ENFORCE_EQ
(
X
->
dims
().
size
(),
2
,
"X's dimension must be 2."
);
PADDLE_ENFORCE_EQ
(
X
->
dims
().
size
(),
2
,
"X's dimension must be 2."
);
PADDLE_ENFORCE_EQ
(
label
->
dims
().
size
(),
1
,
"label's dimension must be 1."
);
PADDLE_ENFORCE_EQ
(
label
->
dims
().
size
(),
1
,
"label's dimension must be 1."
);
PADDLE_ENFORCE_EQ
(
X
->
dims
()[
0
],
label
->
dims
()[
0
]);
PADDLE_ENFORCE_EQ
(
X
->
dims
()[
0
],
label
->
dims
()[
0
]);
ctx
.
Output
<
Tensor
>
(
"Y"
)
->
Resize
({
X
->
dims
()[
0
]});
ctx
.
Output
<
Tensor
>
(
"Y"
)
->
Resize
({
X
->
dims
()[
0
]
,
1
});
}
}
};
};
...
...
python/paddle/v2/framework/tests/test_cos_sim_op.py
浏览文件 @
d0521e6f
...
@@ -7,8 +7,8 @@ class TestCosSimOp(OpTest):
...
@@ -7,8 +7,8 @@ class TestCosSimOp(OpTest):
def
setUp
(
self
):
def
setUp
(
self
):
self
.
op_type
=
"cos_sim"
self
.
op_type
=
"cos_sim"
self
.
inputs
=
{
self
.
inputs
=
{
'X'
:
np
.
random
.
random
((
10
,
5
)).
astype
(
"float32"
),
'X'
:
np
.
random
.
random
((
6
,
5
)).
astype
(
"float32"
),
'Y'
:
np
.
random
.
random
((
10
,
5
)).
astype
(
"float32"
)
'Y'
:
np
.
random
.
random
((
6
,
5
)).
astype
(
"float32"
)
}
}
expect_x_norm
=
np
.
linalg
.
norm
(
self
.
inputs
[
'X'
],
axis
=
1
)
expect_x_norm
=
np
.
linalg
.
norm
(
self
.
inputs
[
'X'
],
axis
=
1
)
expect_y_norm
=
np
.
linalg
.
norm
(
self
.
inputs
[
'Y'
],
axis
=
1
)
expect_y_norm
=
np
.
linalg
.
norm
(
self
.
inputs
[
'Y'
],
axis
=
1
)
...
@@ -28,12 +28,66 @@ class TestCosSimOp(OpTest):
...
@@ -28,12 +28,66 @@ class TestCosSimOp(OpTest):
def
test_check_grad_ingore_x
(
self
):
def
test_check_grad_ingore_x
(
self
):
self
.
check_grad
(
self
.
check_grad
(
[
'Y'
],
'Out'
,
max_relative_error
=
0.05
,
no_grad_set
=
set
(
'X'
))
[
'Y'
],
'Out'
,
max_relative_error
=
0.05
,
no_grad_set
=
set
(
"X"
))
def
test_check_grad_i
gn
ore_y
(
self
):
def
test_check_grad_i
ng
ore_y
(
self
):
self
.
check_grad
(
self
.
check_grad
(
[
'X'
],
'Out'
,
max_relative_error
=
0.05
,
no_grad_set
=
set
(
'Y'
))
[
'X'
],
'Out'
,
max_relative_error
=
0.05
,
no_grad_set
=
set
(
'Y'
))
if
__name__
==
"__main__"
:
class
TestCosSimOp2
(
TestCosSimOp
):
def
setUp
(
self
):
self
.
op_type
=
"cos_sim"
self
.
inputs
=
{
'X'
:
np
.
random
.
random
((
6
,
5
)).
astype
(
"float32"
),
'Y'
:
np
.
random
.
random
((
1
,
5
)).
astype
(
"float32"
)
}
expect_x_norm
=
np
.
linalg
.
norm
(
self
.
inputs
[
'X'
],
axis
=
1
)
expect_y_norm
=
np
.
linalg
.
norm
(
self
.
inputs
[
'Y'
],
axis
=
1
)
expect_out
=
(
self
.
inputs
[
'X'
]
*
self
.
inputs
[
'Y'
]).
sum
(
axis
=
1
)
/
\
expect_x_norm
/
expect_y_norm
self
.
outputs
=
{
'XNorm'
:
np
.
expand_dims
(
expect_x_norm
,
1
),
'YNorm'
:
np
.
expand_dims
(
expect_y_norm
,
1
),
'Out'
:
np
.
expand_dims
(
expect_out
,
1
)
}
class
TestCosSimOp3
(
TestCosSimOp
):
def
setUp
(
self
):
self
.
op_type
=
"cos_sim"
self
.
inputs
=
{
'X'
:
np
.
random
.
random
((
6
,
5
,
2
)).
astype
(
"float32"
),
'Y'
:
np
.
random
.
random
((
6
,
5
,
2
)).
astype
(
"float32"
)
}
expect_x_norm
=
np
.
linalg
.
norm
(
self
.
inputs
[
'X'
],
axis
=
(
1
,
2
))
expect_y_norm
=
np
.
linalg
.
norm
(
self
.
inputs
[
'Y'
],
axis
=
(
1
,
2
))
expect_out
=
(
self
.
inputs
[
'X'
]
*
self
.
inputs
[
'Y'
]).
sum
(
axis
=
(
1
,
2
))
/
\
expect_x_norm
/
expect_y_norm
self
.
outputs
=
{
'XNorm'
:
np
.
expand_dims
(
expect_x_norm
,
1
),
'YNorm'
:
np
.
expand_dims
(
expect_y_norm
,
1
),
'Out'
:
np
.
expand_dims
(
expect_out
,
1
)
}
class
TestCosSimOp4
(
TestCosSimOp
):
def
setUp
(
self
):
self
.
op_type
=
"cos_sim"
self
.
inputs
=
{
'X'
:
np
.
random
.
random
((
6
,
5
,
2
)).
astype
(
"float32"
),
'Y'
:
np
.
random
.
random
((
1
,
5
,
2
)).
astype
(
"float32"
)
}
expect_x_norm
=
np
.
linalg
.
norm
(
self
.
inputs
[
'X'
],
axis
=
(
1
,
2
))
expect_y_norm
=
np
.
linalg
.
norm
(
self
.
inputs
[
'Y'
],
axis
=
(
1
,
2
))
expect_out
=
(
self
.
inputs
[
'X'
]
*
self
.
inputs
[
'Y'
]).
sum
(
axis
=
(
1
,
2
))
/
\
expect_x_norm
/
expect_y_norm
self
.
outputs
=
{
'XNorm'
:
np
.
expand_dims
(
expect_x_norm
,
1
),
'YNorm'
:
np
.
expand_dims
(
expect_y_norm
,
1
),
'Out'
:
np
.
expand_dims
(
expect_out
,
1
)
}
if
__name__
==
'__main__'
:
unittest
.
main
()
unittest
.
main
()
python/paddle/v2/framework/tests/test_cross_entropy_op.py
浏览文件 @
d0521e6f
...
@@ -8,20 +8,22 @@ class TestCrossEntropy(OpTest):
...
@@ -8,20 +8,22 @@ class TestCrossEntropy(OpTest):
self
.
op_type
=
"onehot_cross_entropy"
self
.
op_type
=
"onehot_cross_entropy"
batch_size
=
30
batch_size
=
30
class_num
=
10
class_num
=
10
X
=
numpy
.
random
.
uniform
(
0.1
,
1.0
,
X
=
numpy
.
random
.
uniform
(
0.1
,
1.0
,
[
batch_size
,
class_num
]).
astype
(
"float32"
)
[
batch_size
,
class_num
]).
astype
(
"float32"
)
label
=
(
class_num
/
2
)
*
numpy
.
ones
(
batch_size
).
astype
(
"int32"
)
labels
=
numpy
.
random
.
randint
(
0
,
class_num
,
batch_size
,
dtype
=
"int32"
)
self
.
inputs
=
{
'X'
:
X
,
'label'
:
label
}
Y
=
[]
cross_entropy
=
numpy
.
asmatrix
(
for
i
in
range
(
0
,
batch_size
):
[[
-
numpy
.
log
(
X
[
i
][
labels
[
i
]])]
for
i
in
range
(
X
.
shape
[
0
])],
Y
.
append
(
-
numpy
.
log
(
X
[
i
][
label
[
i
]]))
dtype
=
"float32"
)
self
.
outputs
=
{
'Y'
:
numpy
.
array
(
Y
).
astype
(
"float32"
)}
self
.
inputs
=
{
"X"
:
X
,
"label"
:
labels
}
self
.
outputs
=
{
"Y"
:
cross_entropy
}
def
test_check_output
(
self
):
def
test_check_output
(
self
):
self
.
check_output
()
self
.
check_output
()
def
test_check_grad
(
self
):
def
test_check_grad
(
self
):
self
.
check_grad
([
'X'
],
'Y'
)
self
.
check_grad
([
"X"
],
"Y"
)
if
__name__
==
"__main__"
:
if
__name__
==
"__main__"
:
...
...
python/paddle/v2/framework/tests/test_pad_op.py
浏览文件 @
d0521e6f
...
@@ -22,7 +22,7 @@ class TestPadOp(OpTest):
...
@@ -22,7 +22,7 @@ class TestPadOp(OpTest):
self
.
check_output
()
self
.
check_output
()
def
test_check_grad_normal
(
self
):
def
test_check_grad_normal
(
self
):
self
.
check_grad
([
'X'
],
'Out'
)
self
.
check_grad
([
'X'
],
'Out'
,
max_relative_error
=
0.006
)
def
initTestCase
(
self
):
def
initTestCase
(
self
):
self
.
shape
=
(
16
,
16
)
self
.
shape
=
(
16
,
16
)
...
...
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