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
兽拳
Paddle
提交
6490bb27
P
Paddle
项目概览
兽拳
/
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
1
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
1
Issue
1
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 >>
未验证
提交
6490bb27
编写于
11月 13, 2018
作者:
T
Tao Luo
提交者:
GitHub
11月 13, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #14337 from jczaja/prv-dam-softmax
Softmax op optimization for inference
上级
9f68e9a7
03299ed4
变更
7
隐藏空白更改
内联
并排
Showing
7 changed file
with
68 addition
and
15 deletion
+68
-15
paddle/fluid/operators/math/softmax.cc
paddle/fluid/operators/math/softmax.cc
+4
-2
paddle/fluid/operators/math/softmax.cu
paddle/fluid/operators/math/softmax.cu
+8
-3
paddle/fluid/operators/math/softmax.h
paddle/fluid/operators/math/softmax.h
+1
-1
paddle/fluid/operators/math/softmax_impl.h
paddle/fluid/operators/math/softmax_impl.h
+37
-4
paddle/fluid/operators/softmax_op.h
paddle/fluid/operators/softmax_op.h
+8
-2
paddle/fluid/operators/softmax_with_cross_entropy_op.h
paddle/fluid/operators/softmax_with_cross_entropy_op.h
+2
-2
python/paddle/fluid/tests/unittests/test_softmax_op.py
python/paddle/fluid/tests/unittests/test_softmax_op.py
+8
-1
未找到文件。
paddle/fluid/operators/math/softmax.cc
浏览文件 @
6490bb27
...
...
@@ -19,8 +19,10 @@ namespace paddle {
namespace
operators
{
namespace
math
{
template
class
SoftmaxFunctor
<
platform
::
CPUDeviceContext
,
float
>;
template
class
SoftmaxFunctor
<
platform
::
CPUDeviceContext
,
double
>;
template
class
SoftmaxFunctor
<
platform
::
CPUDeviceContext
,
float
,
true
>;
template
class
SoftmaxFunctor
<
platform
::
CPUDeviceContext
,
float
,
false
>;
template
class
SoftmaxFunctor
<
platform
::
CPUDeviceContext
,
double
,
true
>;
template
class
SoftmaxFunctor
<
platform
::
CPUDeviceContext
,
double
,
false
>;
template
class
SoftmaxGradFunctor
<
platform
::
CPUDeviceContext
,
float
>;
template
class
SoftmaxGradFunctor
<
platform
::
CPUDeviceContext
,
double
>;
...
...
paddle/fluid/operators/math/softmax.cu
浏览文件 @
6490bb27
...
...
@@ -98,9 +98,14 @@ template class SoftmaxGradCUDNNFunctor<float>;
template
class
SoftmaxGradCUDNNFunctor
<
double
>;
template
class
SoftmaxGradCUDNNFunctor
<
platform
::
float16
>;
template
class
SoftmaxFunctor
<
platform
::
CUDADeviceContext
,
platform
::
float16
>;
template
class
SoftmaxFunctor
<
platform
::
CUDADeviceContext
,
float
>;
template
class
SoftmaxFunctor
<
platform
::
CUDADeviceContext
,
double
>;
template
class
SoftmaxFunctor
<
platform
::
CUDADeviceContext
,
platform
::
float16
,
false
>;
template
class
SoftmaxFunctor
<
platform
::
CUDADeviceContext
,
platform
::
float16
,
true
>;
template
class
SoftmaxFunctor
<
platform
::
CUDADeviceContext
,
float
,
false
>;
template
class
SoftmaxFunctor
<
platform
::
CUDADeviceContext
,
double
,
false
>;
template
class
SoftmaxFunctor
<
platform
::
CUDADeviceContext
,
float
,
true
>;
template
class
SoftmaxFunctor
<
platform
::
CUDADeviceContext
,
double
,
true
>;
template
class
SoftmaxGradFunctor
<
platform
::
CUDADeviceContext
,
float
>;
template
class
SoftmaxGradFunctor
<
platform
::
CUDADeviceContext
,
double
>;
template
class
SoftmaxGradFunctor
<
platform
::
CUDADeviceContext
,
...
...
paddle/fluid/operators/math/softmax.h
浏览文件 @
6490bb27
...
...
@@ -19,7 +19,7 @@ namespace paddle {
namespace
operators
{
namespace
math
{
template
<
typename
DeviceContext
,
typename
T
>
template
<
typename
DeviceContext
,
typename
T
,
bool
is_test
>
class
SoftmaxFunctor
{
public:
void
operator
()(
const
DeviceContext
&
context
,
const
framework
::
Tensor
*
X
,
...
...
paddle/fluid/operators/math/softmax_impl.h
浏览文件 @
6490bb27
...
...
@@ -32,10 +32,10 @@ struct ValueClip {
}
};
template
<
typename
DeviceContext
,
typename
T
>
void
SoftmaxFunctor
<
DeviceContext
,
T
>::
operator
()(
const
DeviceContext
&
context
,
const
framework
::
Tensor
*
X
,
framework
::
Tensor
*
Y
)
{
template
<
typename
DeviceContext
,
typename
T
,
bool
is_test
>
void
SoftmaxFunctor
<
DeviceContext
,
T
,
is_test
>::
operator
()(
const
DeviceContext
&
context
,
const
framework
::
Tensor
*
X
,
framework
::
Tensor
*
Y
)
{
auto
logits
=
EigenMatrix
<
T
>::
From
(
*
X
);
auto
softmax
=
EigenMatrix
<
T
>::
From
(
*
Y
);
...
...
@@ -65,6 +65,39 @@ void SoftmaxFunctor<DeviceContext, T>::operator()(const DeviceContext& context,
.
broadcast
(
one_by_class
));
}
template
<
typename
DeviceContext
,
typename
T
>
class
SoftmaxFunctor
<
DeviceContext
,
T
,
true
>
{
void
operator
()(
const
DeviceContext
&
context
,
const
framework
::
Tensor
*
X
,
framework
::
Tensor
*
Y
)
{
auto
logits
=
EigenMatrix
<
T
>::
From
(
*
X
);
auto
softmax
=
EigenMatrix
<
T
>::
From
(
*
Y
);
const
int
kBatchDim
=
0
;
const
int
kClassDim
=
1
;
const
int
batch_size
=
logits
.
dimension
(
kBatchDim
);
const
int
num_classes
=
logits
.
dimension
(
kClassDim
);
Eigen
::
DSizes
<
int
,
1
>
along_class
(
kClassDim
);
Eigen
::
DSizes
<
int
,
2
>
batch_by_one
(
batch_size
,
1
);
Eigen
::
DSizes
<
int
,
2
>
one_by_class
(
1
,
num_classes
);
auto
shifted_logits
=
(
logits
-
logits
.
maximum
(
along_class
)
.
eval
()
.
reshape
(
batch_by_one
)
.
broadcast
(
one_by_class
));
softmax
.
device
(
*
context
.
eigen_device
())
=
shifted_logits
.
exp
();
softmax
.
device
(
*
context
.
eigen_device
())
=
(
softmax
*
softmax
.
sum
(
along_class
)
.
inverse
()
.
eval
()
.
reshape
(
batch_by_one
)
.
broadcast
(
one_by_class
));
}
};
template
<
typename
DeviceContext
,
typename
T
>
void
SoftmaxGradFunctor
<
DeviceContext
,
T
>::
operator
()(
const
DeviceContext
&
context
,
const
framework
::
Tensor
*
y
,
...
...
paddle/fluid/operators/softmax_op.h
浏览文件 @
6490bb27
...
...
@@ -35,8 +35,14 @@ class SoftmaxKernel : public framework::OpKernel<T> {
Tensor
X_2d
=
framework
::
ReshapeToMatrix
(
*
X
,
rank
-
1
);
Tensor
Out_2d
=
framework
::
ReshapeToMatrix
(
*
Out
,
rank
-
1
);
math
::
SoftmaxFunctor
<
DeviceContext
,
T
>
()(
context
.
template
device_context
<
DeviceContext
>(),
&
X_2d
,
&
Out_2d
);
const
bool
is_test
=
context
.
Attr
<
bool
>
(
"is_test"
);
if
(
is_test
==
true
)
{
math
::
SoftmaxFunctor
<
DeviceContext
,
T
,
true
>
()(
context
.
template
device_context
<
DeviceContext
>(),
&
X_2d
,
&
Out_2d
);
}
else
{
math
::
SoftmaxFunctor
<
DeviceContext
,
T
,
false
>
()(
context
.
template
device_context
<
DeviceContext
>(),
&
X_2d
,
&
Out_2d
);
}
}
};
...
...
paddle/fluid/operators/softmax_with_cross_entropy_op.h
浏览文件 @
6490bb27
...
...
@@ -42,8 +42,8 @@ class SoftmaxWithCrossEntropyKernel : public framework::OpKernel<T> {
auto
&
dev_ctx
=
context
.
template
device_context
<
platform
::
CPUDeviceContext
>();
math
::
SoftmaxFunctor
<
platform
::
CPUDeviceContext
,
T
>
()(
dev_ctx
,
logits
,
softmax
);
math
::
SoftmaxFunctor
<
platform
::
CPUDeviceContext
,
T
,
false
>
()(
dev_ctx
,
logits
,
softmax
);
math
::
CrossEntropyFunctor
<
platform
::
CPUDeviceContext
,
T
>
()(
dev_ctx
,
loss
,
softmax
,
labels
,
context
.
Attr
<
bool
>
(
"soft_label"
),
context
.
Attr
<
int
>
(
"ignore_index"
));
...
...
python/paddle/fluid/tests/unittests/test_softmax_op.py
浏览文件 @
6490bb27
...
...
@@ -35,6 +35,7 @@ class TestSoftmaxOp(OpTest):
self
.
op_type
=
"softmax"
self
.
use_cudnn
=
False
self
.
use_mkldnn
=
False
self
.
is_test
=
False
self
.
dtype
=
np
.
float32
self
.
init_kernel_type
()
self
.
shape
=
self
.
get_x_shape
()
...
...
@@ -48,7 +49,8 @@ class TestSoftmaxOp(OpTest):
self
.
outputs
=
{
'Out'
:
out
}
self
.
attrs
=
{
'use_cudnn'
:
self
.
use_cudnn
,
'use_mkldnn'
:
self
.
use_mkldnn
'use_mkldnn'
:
self
.
use_mkldnn
,
'is_test'
:
self
.
is_test
}
def
init_kernel_type
(
self
):
...
...
@@ -144,6 +146,11 @@ class TestSoftmaxFP16CUDNNOp2(TestSoftmaxFP16CUDNNOp):
return
[
2
,
3
,
4
,
5
]
class
TestSoftmaxInference
(
TestSoftmaxOp
):
def
init_kernel_type
(
self
):
self
.
is_test
=
True
class
TestSoftmaxMKLDNNOp
(
TestSoftmaxOp
):
def
init_kernel_type
(
self
):
self
.
use_mkldnn
=
True
...
...
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