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
f8395631
P
Paddle
项目概览
PaddlePaddle
/
Paddle
12 个月 前同步成功
通知
2292
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
f8395631
编写于
12月 14, 2017
作者:
Y
Yancey1989
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
fix invalid dims
上级
1f9426fd
变更
4
显示空白变更内容
内联
并排
Showing
4 changed file
with
67 addition
and
32 deletion
+67
-32
paddle/operators/hierarchical_sigmoid_op.cc
paddle/operators/hierarchical_sigmoid_op.cc
+14
-14
paddle/operators/hierarchical_sigmoid_op.h
paddle/operators/hierarchical_sigmoid_op.h
+13
-13
paddle/operators/math/matrix_bit_code.cc
paddle/operators/math/matrix_bit_code.cc
+6
-5
python/paddle/v2/fluid/tests/test_hsigmoid_op.py
python/paddle/v2/fluid/tests/test_hsigmoid_op.py
+34
-0
未找到文件。
paddle/operators/hierarchical_sigmoid_op.cc
浏览文件 @
f8395631
...
@@ -60,12 +60,11 @@ class HierarchicalSigmoidOp : public framework::OperatorWithKernel {
...
@@ -60,12 +60,11 @@ class HierarchicalSigmoidOp : public framework::OperatorWithKernel {
public:
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
PADDLE_ENFORCE
(
ctx
->
HasInputs
(
"X"
),
"Inputs
(X) should not be null."
);
PADDLE_ENFORCE
(
ctx
->
hasInput
(
"X"
),
"Input
(X) should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Label"
),
"Input(Label) should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Label"
),
"Input(Label) should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
"Out"
),
"Output(Out) should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
"Out"
),
"Output(Out) should not be null."
);
const
int64_t
batch_size
=
ctx
->
GetInputsDim
(
"X"
)[
0
][
0
];
const
int64_t
batch_size
=
ctx
->
GetInputDim
(
"X"
)[
0
];
const
int64_t
size
=
ctx
->
GetInputsDim
(
"X"
).
size
();
std
::
vector
<
int64_t
>
output_shape
({
batch_size
,
num_classes_
-
1
});
std
::
vector
<
int64_t
>
output_shape
({
batch_size
,
size
});
ctx
->
SetOutputDim
(
"Out"
,
framework
::
make_ddim
(
output_shape
));
ctx
->
SetOutputDim
(
"Out"
,
framework
::
make_ddim
(
output_shape
));
}
}
};
};
...
@@ -82,22 +81,23 @@ class HierarchicalSigmoidOpMaker : public framework::OpProtoAndCheckerMaker {
...
@@ -82,22 +81,23 @@ class HierarchicalSigmoidOpMaker : public framework::OpProtoAndCheckerMaker {
framework
::
OpAttrChecker
*
op_checker
)
framework
::
OpAttrChecker
*
op_checker
)
:
OpProtoAndCheckerMaker
(
proto
,
op_checker
)
{
:
OpProtoAndCheckerMaker
(
proto
,
op_checker
)
{
AddInput
(
"X"
,
AddInput
(
"X"
,
"(Tensor
Array, required) The input array. Each Tensor has the
"
"(Tensor
, required) The input Tensor, which the shape is
"
"
same shape with [N * D]."
)
"
[N * D], which N is the size of mini-batch,"
.
AsDuplicable
(
);
"D is the embded size"
);
AddInput
(
"Parameters"
,
AddInput
(
"Parameters"
,
"(Tensor, required), The parameters of hierarchical "
"(Tensor, required), The parameters of hierarchical "
"sigmoid operator, each of them is s a
2-D tensor."
)
"sigmoid operator, each of them is s a
3-D tensor, the shape is"
.
AsDuplicable
(
);
"[N, num_classes - 1, D]"
);
AddInput
(
"Label"
,
AddInput
(
"Label"
,
"(Tensor, required), The labels of training data. It's a"
"(Tensor, required), The labels of training data. It's a"
"1-D tensor
.
"
);
"1-D tensor
, which the shape is [1, N]
"
);
AddInput
(
"Bias"
,
AddInput
(
"Bias"
,
"(Tensor, optional), The bias is a 1-D tensor, "
"(Tensor, optional), The bias is a 1-D tensor, "
"which is applied to the output."
);
"which is applied to the output, the shape is"
AddOutput
(
"[1, num_classes -1]"
);
"Out"
,
AddOutput
(
"Out"
,
"(Tensor, required) The output of hierarchical sigmoid operator."
);
"(Tensor, required) The output of hierarchical sigmoid operator."
"the shape is [N, 1]"
);
AddAttr
<
int
>
(
"num_classes"
,
"(int, required)"
,
"The number of classes"
);
AddAttr
<
int
>
(
"num_classes"
,
"(int, required)"
,
"The number of classes"
);
AddComment
(
R"DOC(
AddComment
(
R"DOC(
The hierarchical sigmoid operator organize the classes into a binary tree.
The hierarchical sigmoid operator organize the classes into a binary tree.
...
...
paddle/operators/hierarchical_sigmoid_op.h
浏览文件 @
f8395631
...
@@ -28,8 +28,8 @@ template <typename Place, typename T>
...
@@ -28,8 +28,8 @@ template <typename Place, typename T>
class
HierarchicalSigmoidOpKernel
:
public
framework
::
OpKernel
<
T
>
{
class
HierarchicalSigmoidOpKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
ins
=
ctx
.
Multi
Input
<
framework
::
Tensor
>
(
"X"
);
auto
*
in
=
ctx
.
Input
<
framework
::
Tensor
>
(
"X"
);
auto
params
=
ctx
.
MultiInput
<
framework
::
Tensor
>
(
"Parameters
"
);
auto
*
param
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Parameter
"
);
auto
*
label
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Label"
);
auto
*
label
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Label"
);
auto
*
bias
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Bias"
);
auto
*
bias
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Bias"
);
auto
*
out
=
ctx
.
Output
<
framework
::
Tensor
>
(
"Out"
);
auto
*
out
=
ctx
.
Output
<
framework
::
Tensor
>
(
"Out"
);
...
@@ -56,8 +56,9 @@ class HierarchicalSigmoidOpKernel : public framework::OpKernel<T> {
...
@@ -56,8 +56,9 @@ class HierarchicalSigmoidOpKernel : public framework::OpKernel<T> {
math
::
AddByBitCode
<
T
>
(
num_classes
,
*
label
,
pre_out
,
*
bias
);
math
::
AddByBitCode
<
T
>
(
num_classes
,
*
label
,
pre_out
,
*
bias
);
}
}
for
(
size_t
i
=
0
;
i
<
ins
.
size
();
++
i
)
{
for
(
size_t
i
=
0
;
i
<
in
.
dims
()[
0
];
++
i
)
{
math
::
MulByBitCode
<
T
>
(
num_classes
,
*
label
,
pre_out
,
*
params
[
i
],
*
ins
[
i
]);
math
::
MulByBitCode
<
T
>
(
num_classes
,
*
label
,
pre_out
,
*
params
->
Slice
(
i
,
i
+
1
),
*
in
->
Slice
(
i
,
i
+
1
));
}
}
// clip the matrix with (-40, 40)
// clip the matrix with (-40, 40)
pre_out_mat
.
device
(
place
)
=
pre_out_mat
.
device
(
place
)
=
...
@@ -79,11 +80,10 @@ template <typename Place, typename T>
...
@@ -79,11 +80,10 @@ template <typename Place, typename T>
class
HierarchicalSigmoidGradOpKernel
:
public
framework
::
OpKernel
<
T
>
{
class
HierarchicalSigmoidGradOpKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
ins
=
ctx
.
MultiInput
<
framework
::
Tensor
>
(
"X"
);
auto
*
in
=
ctx
.
Input
<
framework
::
Tensor
>
(
"X"
);
auto
ins_grad
=
auto
*
in_grad
=
ctx
.
Output
<
framework
::
Tensor
>
(
framework
::
GradVarName
(
"X"
));
ctx
.
MultiOutput
<
framework
::
Tensor
>
(
framework
::
GradVarName
(
"X"
));
auto
*
params
=
auto
params
=
ctx
.
MultiOutput
<
framework
::
Tensor
>
(
ctx
.
Output
<
framework
::
Tensor
>
(
framework
::
GradVarName
(
"Parameters"
));
framework
::
GradVarName
(
"Parameters"
));
auto
*
bias
=
ctx
.
Output
<
framework
::
Tensor
>
(
framework
::
GradVarName
(
"Bias"
));
auto
*
bias
=
ctx
.
Output
<
framework
::
Tensor
>
(
framework
::
GradVarName
(
"Bias"
));
auto
*
label
=
auto
*
label
=
ctx
.
Output
<
framework
::
Tensor
>
(
framework
::
GradVarName
(
"Label"
));
ctx
.
Output
<
framework
::
Tensor
>
(
framework
::
GradVarName
(
"Label"
));
...
@@ -92,7 +92,7 @@ class HierarchicalSigmoidGradOpKernel : public framework::OpKernel<T> {
...
@@ -92,7 +92,7 @@ class HierarchicalSigmoidGradOpKernel : public framework::OpKernel<T> {
framework
::
Tensor
pre_out
;
framework
::
Tensor
pre_out
;
auto
place
=
ctx
.
GetEigenDevice
<
Place
>
();
auto
place
=
ctx
.
GetEigenDevice
<
Place
>
();
auto
&
dev_ctx
=
ctx
.
device_context
();
auto
&
dev_ctx
=
ctx
.
device_context
();
int64_t
batch_size
=
in
s_grad
.
size
()
;
int64_t
batch_size
=
in
_grad
.
dims
()[
0
]
;
int64_t
code_length
=
math
::
FindLastSet
(
num_classes
-
1
);
int64_t
code_length
=
math
::
FindLastSet
(
num_classes
-
1
);
auto
pre_out_mat
=
EigenMatrix
<
T
>::
From
(
pre_out
);
auto
pre_out_mat
=
EigenMatrix
<
T
>::
From
(
pre_out
);
...
@@ -111,11 +111,11 @@ class HierarchicalSigmoidGradOpKernel : public framework::OpKernel<T> {
...
@@ -111,11 +111,11 @@ class HierarchicalSigmoidGradOpKernel : public framework::OpKernel<T> {
math
::
AddByBitCodeGrad
<
T
>
(
num_classes
,
*
label
,
pre_out
,
*
bias
);
math
::
AddByBitCodeGrad
<
T
>
(
num_classes
,
*
label
,
pre_out
,
*
bias
);
}
}
for
(
size_t
i
=
0
;
i
<
in
s_grad
.
size
()
;
++
i
)
{
for
(
size_t
i
=
0
;
i
<
in
_grad
.
dims
()[
0
]
;
++
i
)
{
math
::
MulByBitCodeGradWeight
<
T
>
(
num_classes
,
*
label
,
pre_out
,
*
params
[
i
],
math
::
MulByBitCodeGradWeight
<
T
>
(
num_classes
,
*
label
,
pre_out
,
*
params
[
i
],
*
in
s
[
i
]
);
*
in
[
i
]
->
Slice
(
i
,
i
+
1
)
);
math
::
MulByBitCodeGradError
<
T
>
(
num_classes
,
*
label
,
pre_out
,
*
params
[
i
],
math
::
MulByBitCodeGradError
<
T
>
(
num_classes
,
*
label
,
pre_out
,
*
params
[
i
],
*
ins_grad
[
i
]);
*
ins_grad
[
i
]
->
Slice
(
i
,
i
+
1
)
);
}
}
}
}
};
};
...
...
paddle/operators/math/matrix_bit_code.cc
浏览文件 @
f8395631
...
@@ -52,19 +52,20 @@ namespace math {
...
@@ -52,19 +52,20 @@ namespace math {
*/
*/
template
<
class
CodeTable
,
class
Op
,
typename
T
>
template
<
class
CodeTable
,
class
Op
,
typename
T
>
static
void
AddByBitCodeT
(
Op
op
,
CodeTable
code_table
,
static
void
AddByBitCodeT
(
Op
op
,
CodeTable
code_table
,
const
framework
::
Tensor
&
codes
,
framework
::
Tensor
&
a
,
const
framework
::
Tensor
&
codes
,
const
framework
::
Tensor
&
b
)
{
framework
::
Tensor
&
tmat
,
const
framework
::
Tensor
&
vec
)
{
size_t
num_classes
=
code_table
.
size
();
size_t
num_classes
=
code_table
.
size
();
size_t
max_code_length
=
code_table
.
get_max_code_length
();
size_t
max_code_length
=
code_table
.
get_max_code_length
();
size_t
num_sample
=
a
.
dims
()[
0
];
size_t
num_sample
=
tmat
.
dims
()[
0
];
size_t
width
=
a
.
dims
()[
1
];
size_t
width
=
vec
.
dims
()[
1
];
for
(
size_t
i
=
0
;
i
<
num_sample
;
++
i
)
{
for
(
size_t
i
=
0
;
i
<
num_sample
;
++
i
)
{
auto
code
=
code_table
(
codes
.
data
<
T
>
()[
i
]);
auto
code
=
code_table
(
codes
.
data
<
T
>
()[
i
]);
int
code_length
=
code
.
get_length
();
int
code_length
=
code
.
get_length
();
for
(
int
j
=
0
;
j
<
code_length
;
+
j
)
{
for
(
int
j
=
0
;
j
<
code_length
;
+
j
)
{
size_t
index
=
code
.
calc_index
(
j
);
size_t
index
=
code
.
calc_index
(
j
);
op
(
a
.
data
<
T
>
()[
i
*
width
+
j
],
b
.
data
<
T
>
()[
index
]);
op
(
tmat
.
data
<
T
>
()[
i
*
width
+
j
],
vec
.
data
<
T
>
()[
index
]);
}
}
}
}
}
}
...
...
python/paddle/v2/fluid/tests/test_hsigmoid_op.py
0 → 100644
浏览文件 @
f8395631
import
unittest
import
numpy
as
np
from
op_test
import
OpTest
class
TestHSigmoidOp
(
OpTest
):
def
setUp
(
self
):
self
.
op_type
=
"hierarchical_sigmoid_op"
num_classes
=
6
embded_size
=
10
batch_size
=
5
x
=
np
.
random
.
random
((
batch_size
,
embded_size
)).
astype
(
"float32"
)
parameter
=
np
.
random
.
random
(
(
batch_size
,
num_classes
-
1
,
embded_size
)).
astype
(
"float32"
)
label
=
np
.
random
.
randint
(
0
,
num_classes
,
batch_size
).
astype
(
"int64"
)
bias
=
np
.
random
.
random
((
1
,
num_classes
-
1
))
self
.
inputs
=
{
'X'
:
x
,
'Parameters'
:
parameter
,
'Label'
:
label
,
'Bias'
:
bias
}
self
.
attrs
=
{
'num_classes'
:
num_classes
}
self
.
outputs
=
{
'Out'
:
label
}
def
test_check_output
(
self
):
self
.
check_output
()
def
test_check_grad
(
self
):
self
.
check_grad
([
'x0'
],
'Out'
)
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