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
六月暴雪飞梨花
Paddle
提交
069d0004
P
Paddle
项目概览
六月暴雪飞梨花
/
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
069d0004
编写于
11月 04, 2016
作者:
H
Haonan
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
multi_binary_cross_entropy when ids vector is provided
上级
ef5e483c
变更
10
隐藏空白更改
内联
并排
Showing
10 changed file
with
263 addition
and
4 deletion
+263
-4
paddle/cuda/include/hl_matrix.h
paddle/cuda/include/hl_matrix.h
+30
-0
paddle/cuda/include/stub/hl_matrix_stub.h
paddle/cuda/include/stub/hl_matrix_stub.h
+12
-0
paddle/cuda/src/hl_cuda_matrix.cu
paddle/cuda/src/hl_cuda_matrix.cu
+78
-0
paddle/gserver/layers/CostLayer.cpp
paddle/gserver/layers/CostLayer.cpp
+4
-0
paddle/gserver/tests/test_LayerGrad.cpp
paddle/gserver/tests/test_LayerGrad.cpp
+4
-3
paddle/math/Matrix.cpp
paddle/math/Matrix.cpp
+36
-0
paddle/math/Matrix.h
paddle/math/Matrix.h
+4
-0
paddle/math/tests/test_matrixCompare.cpp
paddle/math/tests/test_matrixCompare.cpp
+65
-1
paddle/parameter/Argument.cpp
paddle/parameter/Argument.cpp
+22
-0
paddle/parameter/Argument.h
paddle/parameter/Argument.h
+8
-0
未找到文件。
paddle/cuda/include/hl_matrix.h
浏览文件 @
069d0004
...
...
@@ -126,6 +126,36 @@ extern void hl_matrix_cross_entropy_bp(real* grad_d,
int
dimM
,
int
dimN
);
/**
* @brief Matrix multi-binary label cross entropy
*
* @param[in] output input matrix (M x N).
* @param[out] entropy output matrix (M x 1).
* @param[in] mat input sparse matrix.
* @param[in] dimM matrix height.
* @param[in] dimN matrix width.
*/
extern
void
hl_matrix_multi_binary_cross_entropy
(
real
*
output
,
real
*
entropy
,
hl_sparse_matrix_s
mat
,
int
dimM
,
int
dimN
);
/**
* @brief Matrix multi-binary label cross entropy backprop
*
* @param[in] output input matrix (M x N).
* @param[out] grad output matrix (M x N).
* @param[in] mat input sparse matrix.
* @param[in] dimM matrix height.
* @param[in] dimN matrix width.
*/
extern
void
hl_matrix_multi_binary_cross_entropy_bp
(
real
*
output
,
real
*
grad
,
hl_sparse_matrix_s
mat
,
int
dimM
,
int
dimN
);
/**
* @brief Matrix zero memory.
*
...
...
paddle/cuda/include/stub/hl_matrix_stub.h
浏览文件 @
069d0004
...
...
@@ -57,6 +57,18 @@ inline void hl_matrix_cross_entropy_bp(real* grad_d,
int
dimM
,
int
dimN
)
{}
inline
void
hl_matrix_multi_binary_cross_entropy
(
real
*
output
,
real
*
entropy
,
hl_sparse_matrix_s
mat
,
int
dimM
,
int
dimN
)
{}
inline
void
hl_matrix_multi_binary_cross_entropy_bp
(
real
*
output
,
real
*
grad
,
hl_sparse_matrix_s
mat
,
int
dimM
,
int
dimN
)
{}
inline
void
hl_matrix_zero_mem
(
real
*
data
,
int
num
)
{}
inline
void
hl_param_relu_forward
(
real
*
output
,
...
...
paddle/cuda/src/hl_cuda_matrix.cu
浏览文件 @
069d0004
...
...
@@ -18,6 +18,7 @@ limitations under the License. */
#include "hl_matrix_ops.cuh"
#include "hl_matrix_apply.cuh"
#include "hl_sequence.h"
#include "hl_sparse.ph"
#include "paddle/utils/Logging.h"
#include "hl_device_functions.cuh"
#include "hl_gpu_matrix_kernel.cuh"
...
...
@@ -317,6 +318,83 @@ void hl_matrix_classification_error(real* A_d,
CHECK_SYNC
(
"hl_matrix_classification_error"
);
}
__global__
void
KeMatrixMultiBinaryCrossEntropy
(
real
*
output
,
real
*
entropy
,
int
*
row
,
int
*
col
,
int
dimM
,
int
dimN
)
{
int
index
=
blockIdx
.
x
*
blockDim
.
x
+
threadIdx
.
x
;
if
(
index
<
dimM
)
{
for
(
int
i
=
0
;
i
<
dimN
;
i
++
)
{
entropy
[
index
]
-=
log
(
1
-
output
[
index
*
dimN
+
i
]);
}
int
*
row_col
=
col
+
row
[
index
];
int
col_num
=
row
[
index
+
1
]
-
row
[
index
];
for
(
int
i
=
0
;
i
<
col_num
;
i
++
)
{
real
o
=
output
[
index
*
dimN
+
row_col
[
i
]];
entropy
[
index
]
-=
log
(
o
/
(
1
-
o
));
}
}
}
void
hl_matrix_multi_binary_cross_entropy
(
real
*
output
,
real
*
entropy
,
hl_sparse_matrix_s
csr_mat
,
int
dimM
,
int
dimN
)
{
CHECK_NOTNULL
(
output
);
CHECK_NOTNULL
(
entropy
);
CHECK_NOTNULL
(
csr_mat
);
int
n_threads
=
1024
;
int
blocks
=
(
dimM
+
n_threads
-
1
)
/
n_threads
;
dim3
threads
(
n_threads
);
dim3
grid
(
blocks
);
hl_csr_matrix
mat
=
(
hl_csr_matrix
)(
csr_mat
->
matrix
);
KeMatrixMultiBinaryCrossEntropy
<<<
grid
,
threads
,
0
,
STREAM_DEFAULT
>>>
(
output
,
entropy
,
mat
->
csr_row
,
mat
->
csr_col
,
dimM
,
dimN
);
CHECK_SYNC
(
"hl_matrix_multi_binary_cross_entropy failed"
);
}
__global__
void
KeMatrixMultiBinaryCrossEntropyBp
(
real
*
output
,
real
*
grad
,
int
*
row
,
int
*
col
,
int
dimM
,
int
dimN
)
{
int
row_idx
=
blockIdx
.
x
*
blockDim
.
x
+
threadIdx
.
x
;
if
(
row_idx
<
dimM
)
{
for
(
int
i
=
0
;
i
<
dimN
;
i
++
)
{
int
index
=
row_idx
*
dimN
+
i
;
grad
[
index
]
+=
1.0
/
(
1
-
output
[
index
]);
}
int
col_num
=
row
[
row_idx
+
1
]
-
row
[
row_idx
];
int
*
row_col
=
col
+
row
[
row_idx
];
for
(
int
i
=
0
;
i
<
col_num
;
i
++
)
{
int
index
=
row_idx
*
dimN
+
row_col
[
i
];
grad
[
index
]
-=
1.0
/
(
output
[
index
]
*
(
1
-
output
[
index
]));
}
}
}
void
hl_matrix_multi_binary_cross_entropy_bp
(
real
*
output
,
real
*
grad
,
hl_sparse_matrix_s
csr_mat
,
int
dimM
,
int
dimN
)
{
CHECK_NOTNULL
(
output
);
CHECK_NOTNULL
(
grad
);
CHECK_NOTNULL
(
csr_mat
);
int
n_threads
=
1024
;
int
blocks
=
(
dimM
+
n_threads
-
1
)
/
n_threads
;
dim3
threads
(
n_threads
);
dim3
grid
(
blocks
);
hl_csr_matrix
mat
=
(
hl_csr_matrix
)(
csr_mat
->
matrix
);
KeMatrixMultiBinaryCrossEntropyBp
<<<
grid
,
threads
,
0
,
STREAM_DEFAULT
>>>
(
output
,
grad
,
mat
->
csr_row
,
mat
->
csr_col
,
dimM
,
dimN
);
CHECK_SYNC
(
"hl_matrix_multi_binary_cross_entropy_bp failed"
);
}
__global__
void
KeMatrixCrossEntropy
(
real
*
O
,
real
*
E
,
int
*
label
,
...
...
paddle/gserver/layers/CostLayer.cpp
浏览文件 @
069d0004
...
...
@@ -462,6 +462,8 @@ bool MultiBinaryLabelCrossEntropy::init(const LayerMap& layerMap,
void
MultiBinaryLabelCrossEntropy
::
forwardImp
(
Matrix
&
output
,
Argument
&
label
,
Matrix
&
target
)
{
label
.
idsToSparseMatrix
(
output
.
getWidth
(),
useGpu_
);
if
(
dynamic_cast
<
CpuSparseMatrix
*>
(
label
.
value
.
get
())
||
dynamic_cast
<
GpuSparseMatrix
*>
(
label
.
value
.
get
()))
{
target
.
multiBinaryLabelCrossEntropy
(
output
,
*
label
.
value
);
...
...
@@ -476,6 +478,8 @@ void MultiBinaryLabelCrossEntropy::forwardImp(Matrix& output, Argument& label,
void
MultiBinaryLabelCrossEntropy
::
backwardImp
(
Matrix
&
output
,
Argument
&
label
,
Matrix
&
outputG
)
{
label
.
idsToSparseMatrix
(
output
.
getWidth
(),
useGpu_
);
if
(
dynamic_cast
<
CpuSparseMatrix
*>
(
label
.
value
.
get
())
||
dynamic_cast
<
GpuSparseMatrix
*>
(
label
.
value
.
get
()))
{
outputG
.
multiBinaryLabelCrossEntropyBp
(
output
,
*
label
.
value
);
...
...
paddle/gserver/tests/test_LayerGrad.cpp
浏览文件 @
069d0004
...
...
@@ -538,9 +538,10 @@ TEST(Layer, multi_binary_label) {
config
.
layerConfig
.
add_inputs
();
config
.
layerConfig
.
add_inputs
();
// Not support GPU now
testLayerGrad
(
config
,
"multi_binary_label_cross_entropy"
,
100
,
/* trans */
false
,
/* useGpu */
false
);
for
(
auto
useGpu
:
{
false
,
true
})
{
testLayerGrad
(
config
,
"multi_binary_label_cross_entropy"
,
100
,
/* trans */
false
,
useGpu
);
}
}
TEST
(
Layer
,
multi_cross_with_selfnorm
)
{
...
...
paddle/math/Matrix.cpp
浏览文件 @
069d0004
...
...
@@ -1268,6 +1268,42 @@ void GpuMatrix::bilinearBackward(const Matrix& out,
}
}
void
GpuMatrix
::
multiBinaryLabelCrossEntropy
(
Matrix
&
output
,
Matrix
&
label
)
{
GpuMatrix
*
output_ptr
=
dynamic_cast
<
GpuMatrix
*>
(
&
output
);
auto
label_ptr
=
dynamic_cast
<
GpuSparseMatrix
*>
(
&
label
);
CHECK
(
output_ptr
&&
label_ptr
)
<<
"Invalid argument pointer"
;
CHECK
(
label_ptr
->
format_
==
SPARSE_CSR
)
<<
"Matrix format not supported"
;
CHECK
(
height_
==
output_ptr
->
height_
&&
width_
==
1
&&
output_ptr
->
width_
==
label_ptr
->
getWidth
()
&&
output_ptr
->
height_
==
label_ptr
->
getHeight
())
<<
"Matrix dimensions are not equal"
;
real
*
output_d
=
output_ptr
->
data_
;
real
*
entropy_d
=
data_
;
hl_sparse_matrix_s
mat_d
=
label_ptr
->
sMatrix_
.
get
();
hl_matrix_multi_binary_cross_entropy
(
output_d
,
entropy_d
,
mat_d
,
height_
,
output_ptr
->
width_
);
}
void
GpuMatrix
::
multiBinaryLabelCrossEntropyBp
(
Matrix
&
output
,
Matrix
&
label
)
{
GpuMatrix
*
output_ptr
=
dynamic_cast
<
GpuMatrix
*>
(
&
output
);
auto
label_ptr
=
dynamic_cast
<
GpuSparseMatrix
*>
(
&
label
);
CHECK
(
output_ptr
&&
label_ptr
)
<<
"Invalid argument pointer"
;
CHECK
(
label_ptr
->
format_
==
SPARSE_CSR
)
<<
"Matrix format not supported"
;
CHECK
(
height_
==
output_ptr
->
height_
&&
width_
==
output_ptr
->
width_
&&
output_ptr
->
width_
==
label_ptr
->
getWidth
()
&&
output_ptr
->
height_
==
label_ptr
->
getHeight
())
<<
"Matrix dimensions are not equal"
;
real
*
output_d
=
output_ptr
->
data_
;
real
*
grad_d
=
data_
;
hl_sparse_matrix_s
mat_d
=
label_ptr
->
sMatrix_
.
get
();
hl_matrix_multi_binary_cross_entropy_bp
(
output_d
,
grad_d
,
mat_d
,
height_
,
width_
);
}
/**
* CpuMatrix
*/
...
...
paddle/math/Matrix.h
浏览文件 @
069d0004
...
...
@@ -1303,6 +1303,10 @@ public:
const
size_t
numChannels
,
const
real
ratioH
,
const
real
ratioW
);
void
multiBinaryLabelCrossEntropy
(
Matrix
&
output
,
Matrix
&
label
);
void
multiBinaryLabelCrossEntropyBp
(
Matrix
&
output
,
Matrix
&
label
);
};
class
CpuMatrix
:
public
Matrix
{
...
...
paddle/math/tests/test_matrixCompare.cpp
浏览文件 @
069d0004
...
...
@@ -2208,7 +2208,6 @@ void testCollectSharedBias(int numSamples, int dim, int channel) {
MatrixCheckErr
(
*
cpuBias
,
*
check
);
}
TEST
(
Matrix
,
sharedBias
)
{
for
(
auto
numSamples
:
{
1
,
100
,
520
})
{
for
(
auto
dim
:
{
100
*
16
,
100
*
32
})
{
...
...
@@ -2222,6 +2221,71 @@ TEST(Matrix, sharedBias) {
}
}
void
testMultiBinaryLabelCrossEntropy
(
int
numSamples
,
int
dim
)
{
MatrixPtr
output
=
std
::
make_shared
<
CpuMatrix
>
(
numSamples
,
dim
);
MatrixPtr
cpuOutput
=
std
::
make_shared
<
CpuMatrix
>
(
numSamples
,
dim
);
MatrixPtr
gpuOutput
=
std
::
make_shared
<
GpuMatrix
>
(
numSamples
,
dim
);
MatrixPtr
cpuEntropy
=
std
::
make_shared
<
CpuMatrix
>
(
numSamples
,
1
);
MatrixPtr
gpuEntropy
=
std
::
make_shared
<
GpuMatrix
>
(
numSamples
,
1
);
MatrixPtr
cpuGrad
=
std
::
make_shared
<
CpuMatrix
>
(
numSamples
,
dim
);
MatrixPtr
gpuGrad
=
std
::
make_shared
<
GpuMatrix
>
(
numSamples
,
dim
);
auto
cpuRows
=
IVector
::
create
(
numSamples
+
1
,
false
);
auto
cpuCols
=
IVector
::
create
(
numSamples
,
false
);
auto
gpuRows
=
IVector
::
create
(
numSamples
+
1
,
true
);
auto
gpuCols
=
IVector
::
create
(
numSamples
,
true
);
cpuRows
->
setElement
(
0
,
0
);
gpuRows
->
setElement
(
0
,
0
);
for
(
int
i
=
0
;
i
<
numSamples
;
i
++
)
{
int
id
=
rand
()
%
dim
;
// NOLINT
cpuRows
->
setElement
(
i
+
1
,
i
+
1
);
gpuRows
->
setElement
(
i
+
1
,
i
+
1
);
cpuCols
->
setElement
(
i
,
id
);
gpuCols
->
setElement
(
i
,
id
);
}
MatrixPtr
cpuLabel
=
std
::
make_shared
<
CpuSparseMatrix
>
(
nullptr
,
cpuRows
->
getData
(),
cpuCols
->
getData
(),
numSamples
,
dim
,
numSamples
,
NO_VALUE
,
SPARSE_CSR
,
false
);
MatrixPtr
gpuLabel
=
std
::
make_shared
<
GpuSparseMatrix
>
(
nullptr
,
gpuRows
->
getData
(),
gpuCols
->
getData
(),
numSamples
,
dim
,
numSamples
,
NO_VALUE
,
SPARSE_CSR
,
false
);
output
->
randomizeUniform
();
cpuOutput
->
zeroMem
();
output
->
softmax
(
*
cpuOutput
);
gpuOutput
->
copyFrom
(
*
cpuOutput
);
cpuEntropy
->
zeroMem
();
gpuEntropy
->
zeroMem
();
cpuEntropy
->
multiBinaryLabelCrossEntropy
(
*
cpuOutput
,
*
cpuLabel
);
gpuEntropy
->
multiBinaryLabelCrossEntropy
(
*
gpuOutput
,
*
gpuLabel
);
MatrixPtr
check1
=
std
::
make_shared
<
CpuMatrix
>
(
numSamples
,
1
);
check1
->
copyFrom
(
*
gpuEntropy
);
MatrixCheckErr
(
*
cpuEntropy
,
*
check1
);
cpuGrad
->
zeroMem
();
gpuGrad
->
zeroMem
();
cpuGrad
->
multiBinaryLabelCrossEntropyBp
(
*
cpuOutput
,
*
cpuLabel
);
gpuGrad
->
multiBinaryLabelCrossEntropyBp
(
*
gpuOutput
,
*
gpuLabel
);
MatrixPtr
check2
=
std
::
make_shared
<
CpuMatrix
>
(
numSamples
,
dim
);
check2
->
copyFrom
(
*
gpuGrad
);
MatrixCheckErr
(
*
cpuGrad
,
*
check2
);
}
TEST
(
Matrix
,
multiBinaryCrossEntropy
)
{
for
(
auto
numSamples
:
{
1
,
100
,
500
})
{
for
(
auto
dim
:
{
1000
,
10000
,
100000
})
{
VLOG
(
3
)
<<
" numSamples="
<<
numSamples
<<
" dim="
<<
dim
;
testMultiBinaryLabelCrossEntropy
(
numSamples
,
dim
);
}
}
}
int
main
(
int
argc
,
char
**
argv
)
{
testing
::
InitGoogleTest
(
&
argc
,
argv
);
initMain
(
argc
,
argv
);
...
...
paddle/parameter/Argument.cpp
浏览文件 @
069d0004
...
...
@@ -572,4 +572,26 @@ void Argument::subArgFrom(const Argument& input, size_t offset, size_t height,
}
}
void
Argument
::
idsToSparseMatrix
(
int
width
,
bool
useGpu
)
{
if
(
ids
)
{
CHECK
(
!
value
);
int
height
=
ids
->
getSize
();
int
nnz
=
height
;
auto
rows
=
IVector
::
create
(
height
+
1
,
useGpu
);
auto
cols
=
IVector
::
create
(
nnz
,
useGpu
);
rows
->
setElement
(
0
,
0
);
for
(
int
i
=
0
;
i
<
height
;
i
++
)
{
int
id
=
ids
->
getElement
(
i
);
CHECK_LT
(
id
,
width
);
rows
->
setElement
(
i
+
1
,
i
+
1
);
cols
->
setElement
(
i
,
id
);
}
value
=
Matrix
::
createSparseMatrix
(
nullptr
,
rows
->
getData
(),
cols
->
getData
(),
height
,
width
,
nnz
,
NO_VALUE
,
SPARSE_CSR
,
false
,
useGpu
);
}
else
{
CHECK
(
value
);
}
}
}
// namespace paddle
paddle/parameter/Argument.h
浏览文件 @
069d0004
...
...
@@ -286,6 +286,14 @@ struct Argument {
sequence has sub-sequence degrades to a sequence.
*/
void
degradeSequence
(
const
Argument
&
input
,
bool
useGpu
);
/*
@brief convert the ids vector to value as a sparse matrix
the ids vector keeps valid
@param the matrix width (id range)
@useGpu
*/
void
idsToSparseMatrix
(
int
width
,
bool
useGpu
);
};
}
// 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