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
cdf2579d
P
Paddle
项目概览
PaddlePaddle
/
Paddle
11 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 >>
未验证
提交
cdf2579d
编写于
10月 31, 2018
作者:
T
Tao Luo
提交者:
GitHub
10月 31, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #14053 from jczaja/prv-seqpool-max
Max Sequence pool optimization
上级
35915fc5
458b16f4
变更
8
隐藏空白更改
内联
并排
Showing
8 changed file
with
77 addition
and
15 deletion
+77
-15
paddle/fluid/API.spec
paddle/fluid/API.spec
+1
-1
paddle/fluid/operators/math/sequence_pooling.cc
paddle/fluid/operators/math/sequence_pooling.cc
+44
-4
paddle/fluid/operators/math/sequence_pooling.cu
paddle/fluid/operators/math/sequence_pooling.cu
+1
-1
paddle/fluid/operators/math/sequence_pooling.h
paddle/fluid/operators/math/sequence_pooling.h
+1
-1
paddle/fluid/operators/sequence_pool_op.cc
paddle/fluid/operators/sequence_pool_op.cc
+1
-0
paddle/fluid/operators/sequence_pool_op.h
paddle/fluid/operators/sequence_pool_op.h
+11
-6
python/paddle/fluid/layers/nn.py
python/paddle/fluid/layers/nn.py
+4
-2
python/paddle/fluid/tests/unittests/test_seq_pool.py
python/paddle/fluid/tests/unittests/test_seq_pool.py
+14
-0
未找到文件。
paddle/fluid/API.spec
浏览文件 @
cdf2579d
...
...
@@ -64,7 +64,7 @@ paddle.fluid.layers.chunk_eval ArgSpec(args=['input', 'label', 'chunk_scheme', '
paddle.fluid.layers.sequence_conv ArgSpec(args=['input', 'num_filters', 'filter_size', 'filter_stride', 'padding', 'bias_attr', 'param_attr', 'act', 'name'], varargs=None, keywords=None, defaults=(3, 1, None, None, None, None, None))
paddle.fluid.layers.conv2d ArgSpec(args=['input', 'num_filters', 'filter_size', 'stride', 'padding', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(1, 0, 1, None, None, None, True, None, None))
paddle.fluid.layers.conv3d ArgSpec(args=['input', 'num_filters', 'filter_size', 'stride', 'padding', 'dilation', 'groups', 'param_attr', 'bias_attr', 'use_cudnn', 'act', 'name'], varargs=None, keywords=None, defaults=(1, 0, 1, None, None, None, True, None, None))
paddle.fluid.layers.sequence_pool ArgSpec(args=['input', 'pool_type'
], varargs=None, keywords=None, defaults=None
)
paddle.fluid.layers.sequence_pool ArgSpec(args=['input', 'pool_type'
, 'is_test'], varargs=None, keywords=None, defaults=(False,)
)
paddle.fluid.layers.sequence_softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(False, None))
paddle.fluid.layers.softmax ArgSpec(args=['input', 'use_cudnn', 'name'], varargs=None, keywords=None, defaults=(True, None))
paddle.fluid.layers.pool2d ArgSpec(args=['input', 'pool_size', 'pool_type', 'pool_stride', 'pool_padding', 'global_pooling', 'use_cudnn', 'ceil_mode', 'name'], varargs=None, keywords=None, defaults=(-1, 'max', 1, 0, False, True, False, None))
...
...
paddle/fluid/operators/math/sequence_pooling.cc
浏览文件 @
cdf2579d
...
...
@@ -31,7 +31,7 @@ template <typename T, int MajorType = Eigen::RowMajor,
typename
IndexType
=
Eigen
::
DenseIndex
>
using
EigenMatrix
=
framework
::
EigenMatrix
<
T
,
MajorType
,
IndexType
>
;
template
<
typename
T
>
template
<
typename
T
,
bool
is_test
>
class
MaxSeqPoolFunctor
{
public:
void
operator
()(
const
platform
::
CPUDeviceContext
&
context
,
...
...
@@ -70,7 +70,41 @@ class MaxSeqPoolFunctor {
}
}
};
// Instantisation of Max Sequence Pooling for test phase eg. no need to fill
// index buffer
template
<
typename
T
>
class
MaxSeqPoolFunctor
<
T
,
true
>
{
public:
void
operator
()(
const
platform
::
CPUDeviceContext
&
context
,
const
framework
::
LoDTensor
&
input
,
framework
::
Tensor
*
output
,
framework
::
Tensor
*
index
)
{
auto
in_dims
=
input
.
dims
();
auto
out_dims
=
output
->
dims
();
PADDLE_ENFORCE_GT
(
in_dims
.
size
(),
1
);
PADDLE_ENFORCE_GT
(
out_dims
.
size
(),
1
);
for
(
int64_t
i
=
1
;
i
<
in_dims
.
size
();
++
i
)
{
PADDLE_ENFORCE_EQ
(
in_dims
[
i
],
out_dims
[
i
]);
}
auto
starts
=
input
.
lod
()[
0
];
const
T
*
in_data
=
input
.
data
<
T
>
();
T
*
out_data
=
output
->
data
<
T
>
();
int64_t
num_seq
=
out_dims
[
0
];
int64_t
dim
=
output
->
numel
()
/
num_seq
;
for
(
int64_t
i
=
0
;
i
<
num_seq
;
++
i
)
{
std
::
memcpy
(
&
out_data
[
i
*
dim
],
&
in_data
[
starts
[
i
]
*
dim
],
dim
*
sizeof
(
T
));
for
(
size_t
j
=
starts
[
i
]
+
1
;
j
<
starts
[
i
+
1
];
++
j
)
{
for
(
int64_t
k
=
0
;
k
<
dim
;
++
k
)
{
if
(
in_data
[
j
*
dim
+
k
]
>
out_data
[
i
*
dim
+
k
])
{
out_data
[
i
*
dim
+
k
]
=
in_data
[
j
*
dim
+
k
];
}
}
}
}
}
};
template
<
typename
T
>
class
MaxSeqPoolGradFunctor
{
public:
...
...
@@ -188,11 +222,16 @@ class SequencePoolFunctor<platform::CPUDeviceContext, T> {
/* max pool has index output */
void
operator
()(
const
platform
::
CPUDeviceContext
&
context
,
const
std
::
string
pooltype
,
const
framework
::
LoDTensor
&
input
,
framework
::
Tensor
*
output
,
framework
::
Tensor
*
output
,
bool
is_test
,
framework
::
Tensor
*
index
=
nullptr
)
{
if
(
pooltype
==
"MAX"
)
{
math
::
MaxSeqPoolFunctor
<
T
>
max_pool
;
max_pool
(
context
,
input
,
output
,
index
);
if
(
is_test
)
{
math
::
MaxSeqPoolFunctor
<
T
,
true
>
max_pool
;
max_pool
(
context
,
input
,
output
,
index
);
}
else
{
math
::
MaxSeqPoolFunctor
<
T
,
false
>
max_pool
;
max_pool
(
context
,
input
,
output
,
index
);
}
return
;
}
if
(
pooltype
==
"LAST"
)
{
...
...
@@ -200,6 +239,7 @@ class SequencePoolFunctor<platform::CPUDeviceContext, T> {
last_pool
(
context
,
input
,
output
);
return
;
}
if
(
pooltype
==
"FIRST"
)
{
math
::
FirstSeqPoolFunctor
<
T
>
first_pool
;
first_pool
(
context
,
input
,
output
);
...
...
paddle/fluid/operators/math/sequence_pooling.cu
浏览文件 @
cdf2579d
...
...
@@ -133,7 +133,7 @@ class SequencePoolFunctor<platform::CUDADeviceContext, T> {
public:
void
operator
()(
const
platform
::
CUDADeviceContext
&
context
,
const
std
::
string
pooltype
,
const
framework
::
LoDTensor
&
input
,
framework
::
Tensor
*
output
,
framework
::
Tensor
*
output
,
bool
is_test
,
framework
::
Tensor
*
index
=
nullptr
)
{
auto
&
lod
=
input
.
lod
()[
0
];
const
size_t
item_dim
=
output
->
numel
()
/
output
->
dims
()[
0
];
...
...
paddle/fluid/operators/math/sequence_pooling.h
浏览文件 @
cdf2579d
...
...
@@ -28,7 +28,7 @@ class SequencePoolFunctor {
/* max pool has index output */
void
operator
()(
const
DeviceContext
&
context
,
const
std
::
string
pooltype
,
const
framework
::
LoDTensor
&
input
,
framework
::
Tensor
*
output
,
framework
::
Tensor
*
index
=
nullptr
);
bool
is_test
=
false
,
framework
::
Tensor
*
index
=
nullptr
);
};
template
<
typename
DeviceContext
,
typename
T
>
...
...
paddle/fluid/operators/sequence_pool_op.cc
浏览文件 @
cdf2579d
...
...
@@ -47,6 +47,7 @@ class SequencePoolOpMaker : public framework::OpProtoAndCheckerMaker {
"(Tensor<int>) This tensor is used for the sequence max-pooling "
"to record the max indexes."
)
.
AsIntermediate
();
AddAttr
<
bool
>
(
"is_test"
,
""
).
SetDefault
(
false
);
AddAttr
<
std
::
string
>
(
"pooltype"
,
"(string, default 'AVERAGE') the pooling pooltype of SequencePoolOp."
)
...
...
paddle/fluid/operators/sequence_pool_op.h
浏览文件 @
cdf2579d
...
...
@@ -32,10 +32,6 @@ class SequencePoolKernel : public framework::OpKernel<T> {
auto
*
in
=
context
.
Input
<
LoDTensor
>
(
"X"
);
auto
*
out
=
context
.
Output
<
Tensor
>
(
"Out"
);
std
::
string
pooltype
=
context
.
Attr
<
std
::
string
>
(
"pooltype"
);
Tensor
*
index
=
nullptr
;
if
(
pooltype
==
"MAX"
)
{
index
=
context
.
Output
<
Tensor
>
(
"MaxIndex"
);
}
auto
dims
=
in
->
dims
();
auto
lod
=
in
->
lod
();
...
...
@@ -48,13 +44,22 @@ class SequencePoolKernel : public framework::OpKernel<T> {
dims
[
0
]
=
lod
[
0
].
size
()
-
1
;
out
->
Resize
({
dims
});
out
->
mutable_data
<
T
>
(
context
.
GetPlace
());
if
(
pooltype
==
"MAX"
)
{
Tensor
*
index
=
nullptr
;
const
bool
is_test
=
context
.
Attr
<
bool
>
(
"is_test"
);
// Do not create index buffer for inference (is_test) mode
// TODO(jczaja): Skip index buffer creation for other devices eg. GPU
if
(
pooltype
==
"MAX"
&&
(
is_test
==
false
||
platform
::
is_cpu_place
(
context
.
GetPlace
())
==
false
))
{
index
=
context
.
Output
<
Tensor
>
(
"MaxIndex"
);
index
->
Resize
({
dims
});
index
->
mutable_data
<
int
>
(
context
.
GetPlace
());
}
math
::
SequencePoolFunctor
<
DeviceContext
,
T
>
pool
;
pool
(
context
.
template
device_context
<
DeviceContext
>(),
pooltype
,
*
in
,
out
,
index
);
i
s_test
,
i
ndex
);
}
};
...
...
python/paddle/fluid/layers/nn.py
浏览文件 @
cdf2579d
...
...
@@ -1825,7 +1825,7 @@ def conv3d(input,
return
helper
.
append_activation
(
pre_act
)
def
sequence_pool
(
input
,
pool_type
):
def
sequence_pool
(
input
,
pool_type
,
is_test
=
False
):
"""
This function add the operator for sequence pooling.
It pools features of all time-steps of each instance, and is applied
...
...
@@ -1862,6 +1862,7 @@ def sequence_pool(input, pool_type):
input(variable): The input variable which is a LoDTensor.
pool_type (string): The pooling type of sequence_pool.
It supports average, sum, sqrt and max.
is_test(bool, Default False): Used distinguish training from scoring mode.
Returns:
The sequence pooling variable which is a Tensor.
...
...
@@ -1889,7 +1890,8 @@ def sequence_pool(input, pool_type):
inputs
=
{
"X"
:
input
},
outputs
=
{
"Out"
:
pool_out
,
"MaxIndex"
:
max_index
},
attrs
=
{
"pooltype"
:
pool_type
.
upper
()})
attrs
=
{
"pooltype"
:
pool_type
.
upper
(),
"is_test"
:
is_test
})
# when pool_type is max, variable max_index is initialized,
# so we stop the gradient explicitly here
...
...
python/paddle/fluid/tests/unittests/test_seq_pool.py
浏览文件 @
cdf2579d
...
...
@@ -184,6 +184,20 @@ class TestSeqMaxPool2D(TestSeqAvgPool2D):
out
[
i
]
=
np
.
reshape
(
np
.
amax
(
sub_x
,
axis
=
0
),
(
3
,
11
))
class
TestSeqMaxPool2DInference
(
TestSeqMaxPool2D
):
def
compute
(
self
,
x
,
offset
,
out
):
self
.
attrs
=
{
'pooltype'
:
"MAX"
,
'is_test'
:
True
}
for
i
in
range
(
len
(
offset
[
0
])
-
1
):
sub_x
=
np
.
reshape
(
x
[
offset
[
0
][
i
]:
offset
[
0
][
i
+
1
],
:],
(
-
1
,
3
*
11
))
out
[
i
]
=
np
.
reshape
(
np
.
amax
(
sub_x
,
axis
=
0
),
(
3
,
11
))
def
test_check_grad
(
self
):
"""Grad computation does not apply to Sequence MAX
Pool executed when is_test is true """
return
class
TestSeqLastPool2D
(
TestSeqAvgPool2D
):
def
compute
(
self
,
x
,
offset
,
out
):
self
.
attrs
=
{
'pooltype'
:
"LAST"
}
...
...
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