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
d189d4db
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2298
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
d189d4db
编写于
8月 24, 2018
作者:
Z
Zeng Jinle
提交者:
GitHub
8月 24, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #12884 from sneaxiy/sequence_mask_op
Add sequence_mask_op for DAM model
上级
7240e721
7df74a5e
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
366 addition
and
0 deletion
+366
-0
paddle/fluid/API.spec
paddle/fluid/API.spec
+1
-0
paddle/fluid/operators/sequence_mask_op.cc
paddle/fluid/operators/sequence_mask_op.cc
+26
-0
paddle/fluid/operators/sequence_mask_op.cu
paddle/fluid/operators/sequence_mask_op.cu
+22
-0
paddle/fluid/operators/sequence_mask_op.h
paddle/fluid/operators/sequence_mask_op.h
+154
-0
python/paddle/fluid/layers/nn.py
python/paddle/fluid/layers/nn.py
+69
-0
python/paddle/fluid/tests/unittests/test_sequence_mask.py
python/paddle/fluid/tests/unittests/test_sequence_mask.py
+94
-0
未找到文件。
paddle/fluid/API.spec
浏览文件 @
d189d4db
...
...
@@ -162,6 +162,7 @@ paddle.fluid.layers.crop ArgSpec(args=['x', 'shape', 'offsets', 'name'], varargs
paddle.fluid.layers.rank_loss ArgSpec(args=['label', 'left', 'right', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.prelu ArgSpec(args=['x', 'mode', 'param_attr', 'name'], varargs=None, keywords=None, defaults=(None, None))
paddle.fluid.layers.flatten ArgSpec(args=['x', 'axis', 'name'], varargs=None, keywords=None, defaults=(1, None))
paddle.fluid.layers.sequence_mask ArgSpec(args=['x', 'maxlen', 'dtype', 'name'], varargs=None, keywords=None, defaults=(None, 'int64', None))
paddle.fluid.layers.stack ArgSpec(args=['x', 'axis'], varargs=None, keywords=None, defaults=(0,))
paddle.fluid.layers.data ArgSpec(args=['name', 'shape', 'append_batch_size', 'dtype', 'lod_level', 'type', 'stop_gradient'], varargs=None, keywords=None, defaults=(True, 'float32', 0, VarType.LOD_TENSOR, True))
paddle.fluid.layers.open_recordio_file ArgSpec(args=['filename', 'shapes', 'lod_levels', 'dtypes', 'pass_num', 'for_parallel'], varargs=None, keywords=None, defaults=(1, True))
...
...
paddle/fluid/operators/sequence_mask_op.cc
0 → 100644
浏览文件 @
d189d4db
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/operators/sequence_mask_op.h"
REGISTER_OPERATOR
(
sequence_mask
,
paddle
::
operators
::
SequenceMaskOp
,
paddle
::
operators
::
SequenceMaskOpMaker
,
paddle
::
framework
::
EmptyGradOpMaker
);
REGISTER_OP_CPU_KERNEL
(
sequence_mask
,
paddle
::
operators
::
SequenceMaskKernel
<
paddle
::
platform
::
CPUDeviceContext
,
int
>
,
paddle
::
operators
::
SequenceMaskKernel
<
paddle
::
platform
::
CPUDeviceContext
,
int64_t
>
);
paddle/fluid/operators/sequence_mask_op.cu
0 → 100644
浏览文件 @
d189d4db
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/fluid/operators/sequence_mask_op.h"
REGISTER_OP_CUDA_KERNEL
(
sequence_mask
,
paddle
::
operators
::
SequenceMaskKernel
<
paddle
::
platform
::
CUDADeviceContext
,
int
>
,
paddle
::
operators
::
SequenceMaskKernel
<
paddle
::
platform
::
CUDADeviceContext
,
int64_t
>
);
paddle/fluid/operators/sequence_mask_op.h
0 → 100644
浏览文件 @
d189d4db
// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#ifdef __NVCC__
#include <thrust/device_ptr.h>
#include <thrust/functional.h>
#include <thrust/reduce.h>
#else
#include <algorithm>
#endif
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/platform/for_range.h"
namespace
paddle
{
namespace
operators
{
class
SequenceMaskOp
:
public
framework
::
OperatorWithKernel
{
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"X"
),
"Input(X) must exist"
);
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
"Y"
),
"Output(Y) must exist"
);
auto
maxlen
=
ctx
->
Attrs
().
Get
<
int
>
(
"maxlen"
);
if
(
maxlen
>
0
)
{
// We can only infershape when maxlen > 0
auto
dim
=
framework
::
vectorize2int
(
ctx
->
GetInputDim
(
"X"
));
dim
.
push_back
(
maxlen
);
ctx
->
SetOutputDim
(
"Y"
,
framework
::
make_ddim
(
dim
));
}
}
};
class
SequenceMaskOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
void
Make
()
override
{
AddInput
(
"X"
,
"The input tensor of sequence_mask op."
);
AddOutput
(
"Y"
,
"The output mask of sequence_mask op."
);
AddAttr
<
int
>
(
"maxlen"
,
"The maximum length of the sequence. If maxlen < 0, maxlen "
"= max(Input(X))."
)
.
SetDefault
(
-
1
)
.
AddCustomChecker
([](
int
&
v
)
{
PADDLE_ENFORCE
(
v
<
0
||
v
>=
1
,
"Attr(maxlen) must be less than 0 or larger than 1"
);
});
AddAttr
<
int
>
(
"out_dtype"
,
"Output data type"
);
AddComment
(
R"DOC(
SequenceMask Operator
This operator outputs a Mask according to Input(X) and Attr(maxlen).
Supposing Input(X) is a Tensor with shape [d_1, d_2, ..., d_n], the
Output(Y) is a mask with shape [d_1, d_2, ..., d_n, maxlen], where:
Y(i_1, i_2, ..., i_n, j) = (j < X(i_1, i_2, ..., i_n))
If maxlen < 0, maxlen = max(X)
)DOC"
);
}
};
template
<
typename
Tx
,
typename
Ty
>
struct
SequenceMaskForRangeFunctor
{
HOSTDEVICE
SequenceMaskForRangeFunctor
(
const
Tx
*
x
,
Ty
*
y
,
int
maxlen
)
:
x_
(
x
),
y_
(
y
),
maxlen_
(
maxlen
)
{}
HOSTDEVICE
void
operator
()(
int
y_idx
)
const
{
int
x_idx
=
y_idx
/
maxlen_
;
int
j
=
y_idx
%
maxlen_
;
y_
[
y_idx
]
=
static_cast
<
Ty
>
(
j
<
x_
[
x_idx
]
?
1
:
0
);
}
private:
const
Tx
*
x_
;
Ty
*
y_
;
int
maxlen_
;
};
template
<
typename
DeviceContext
,
typename
Tx
>
struct
SequenceMaskFunctor
{
using
Tensor
=
framework
::
LoDTensor
;
SequenceMaskFunctor
(
const
DeviceContext
&
ctx
,
const
Tx
*
x
,
Tensor
*
y
,
int
limits
,
int
maxlen
)
:
ctx_
(
ctx
),
x_
(
x
),
y_
(
y
),
limits_
(
limits
),
maxlen_
(
maxlen
)
{}
template
<
typename
Ty
>
void
operator
()()
const
{
auto
*
y_data
=
y_
->
mutable_data
<
Ty
>
(
ctx_
.
GetPlace
());
platform
::
ForRange
<
DeviceContext
>
for_range
(
ctx_
,
limits_
);
for_range
(
SequenceMaskForRangeFunctor
<
Tx
,
Ty
>
(
x_
,
y_data
,
maxlen_
));
}
private:
const
DeviceContext
&
ctx_
;
const
Tx
*
x_
;
Tensor
*
y_
;
int
limits_
;
int
maxlen_
;
};
template
<
typename
DeviceContext
,
typename
Tx
>
class
SequenceMaskKernel
:
public
framework
::
OpKernel
<
Tx
>
{
using
Tensor
=
framework
::
LoDTensor
;
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
*
x
=
ctx
.
Input
<
Tensor
>
(
"X"
);
auto
*
y
=
ctx
.
Output
<
Tensor
>
(
"Y"
);
auto
maxlen
=
ctx
.
Attr
<
int
>
(
"maxlen"
);
auto
*
x_data
=
x
->
data
<
Tx
>
();
auto
x_numel
=
x
->
numel
();
if
(
maxlen
<
0
)
{
#ifdef __NVCC__
VLOG
(
10
)
<<
"SequenceMaskOp on GPU may be slow when maxlen is not provided."
;
maxlen
=
static_cast
<
int
>
(
thrust
::
reduce
(
thrust
::
device_pointer_cast
(
x_data
),
thrust
::
device_pointer_cast
(
x_data
)
+
x_numel
,
static_cast
<
Tx
>
(
0
),
thrust
::
maximum
<
Tx
>
()));
#else
maxlen
=
static_cast
<
int
>
(
*
std
::
max_element
(
x_data
,
x_data
+
x_numel
));
#endif
auto
y_dim
=
framework
::
vectorize2int
(
x
->
dims
());
y_dim
.
push_back
(
maxlen
);
y
->
Resize
(
framework
::
make_ddim
(
y_dim
));
}
auto
out_dtype
=
static_cast
<
framework
::
proto
::
VarType
::
Type
>
(
ctx
.
Attr
<
int
>
(
"out_dtype"
));
auto
&
dev_ctx
=
ctx
.
template
device_context
<
DeviceContext
>();
framework
::
VisitDataType
(
out_dtype
,
SequenceMaskFunctor
<
DeviceContext
,
Tx
>
(
dev_ctx
,
x_data
,
y
,
x_numel
*
maxlen
,
maxlen
));
}
};
}
// namespace operators
}
// namespace paddle
python/paddle/fluid/layers/nn.py
浏览文件 @
d189d4db
...
...
@@ -103,6 +103,7 @@ __all__ = [
'rank_loss'
,
'prelu'
,
'flatten'
,
'sequence_mask'
,
'stack'
,
]
...
...
@@ -5520,7 +5521,75 @@ def flatten(x, axis=1, name=None):
return
out
def
sequence_mask
(
x
,
maxlen
=
None
,
dtype
=
'int64'
,
name
=
None
):
"""
**SequenceMask Layer**
This layer outputs a mask according to the input :code:`x` and
:code:`maxlen` with data type of :code:`dtype`.
Supposing :code:`x` is a Tensor with shape [d_1, d_2, ..., d_n], the
:code:`y` is a mask with shape [d_1, d_2, ..., d_n, maxlen], where:
.. math::
y(i_1, i_2,..., i_n, j) = (j < x(i_1, i_2,..., i_n))
Args:
x (Variable): Input tensor of sequence_mask layer,
whose elements are integers less than :code:`maxlen`.
maxlen (int|None): Maximum length of the sequence. If :code:`maxlen`
is None, it would be replace with :math:`max(x)`.
dtype (np.dtype|core.VarDesc.VarType|str): Data type of the output.
name (str|None): A name for this layer(optional). If set None, the
layer will be named automatically.
Returns:
Variable: The output sequence mask.
"""
helper
=
LayerHelper
(
'sequence_mask'
,
**
locals
())
if
name
is
None
:
out
=
helper
.
create_tmp_variable
(
dtype
=
dtype
)
else
:
out
=
helper
.
create_tmp_variable
(
dtype
=
dtype
,
name
=
name
)
helper
.
append_op
(
type
=
'sequence_mask'
,
inputs
=
{
'X'
:
[
x
]},
outputs
=
{
'Y'
:
out
},
attrs
=
{
'max_len'
:
maxlen
if
maxlen
is
not
None
else
-
1
,
'out_dtype'
:
out
.
dtype
})
return
out
def
stack
(
x
,
axis
=
0
):
"""
**Stack Layer**
This layer stacks all of the input :code:`x` along axis.
Input :code:`x` can be a single variable, a :code:`list` of variables,
or a :code:`tuple` of variables. If :code:`x` is a :code:`list` or
:code:`tuple`, the shapes of all these variables must be the same.
Supposing the shape of each input is :math:`[d_0, d_1, ..., d_{n-1}]`,
the shape of the output variable would be
:math:`[d_0, d_1, ..., d_{axis}=len(x), ..., d_{n-1}]`.
If :code:`axis` < 0, it would be replaced with :code:`axis+rank(x[0])+1`.
If :code:`axis` is None, it would be replaced with 0.
Args:
x (Variable|list(Variable)|tuple(Variable)): Input variables.
axis (int|None): The axis along which all inputs are stacked.
Returns:
Variable: The stacked variable.
"""
helper
=
LayerHelper
(
'stack'
,
**
locals
())
axis
=
0
if
axis
is
None
else
axis
...
...
python/paddle/fluid/tests/unittests/test_sequence_mask.py
0 → 100644
浏览文件 @
d189d4db
# Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from
op_test
import
OpTest
import
paddle.fluid
as
fluid
from
paddle.fluid.framework
import
convert_np_dtype_to_dtype_
import
paddle.fluid.core
as
core
import
numpy
as
np
import
copy
import
unittest
class
SequenceMaskTestBase
(
OpTest
):
def
initDefaultParameters
(
self
):
self
.
op_type
=
'sequence_mask'
self
.
maxlen
=
10
self
.
mask_dtype
=
'int64'
self
.
x
=
[[
0
,
3
,
4
],
[
5
,
7
,
9
]]
def
initParameters
(
self
):
pass
def
setUp
(
self
):
self
.
initDefaultParameters
()
self
.
initParameters
()
if
not
isinstance
(
self
.
x
,
np
.
ndarray
):
self
.
x
=
np
.
array
(
self
.
x
)
self
.
inputs
=
{
'X'
:
self
.
x
}
self
.
outputs
=
{
'Y'
:
self
.
calc_ground_truth_mask
()}
self
.
attrs
=
{
'maxlen'
:
self
.
maxlen
,
'out_dtype'
:
convert_np_dtype_to_dtype_
(
self
.
mask_dtype
)
}
def
calc_ground_truth_mask
(
self
):
maxlen
=
np
.
max
(
self
.
x
)
if
self
.
maxlen
<
0
else
self
.
maxlen
shape
=
self
.
x
.
shape
+
(
maxlen
,
)
index_broadcast
=
np
.
broadcast_to
(
np
.
reshape
(
range
(
maxlen
),
newshape
=
[
1
]
*
self
.
x
.
ndim
+
[
-
1
]),
shape
=
shape
)
x_broadcast
=
np
.
broadcast_to
(
np
.
reshape
(
self
.
x
,
newshape
=
self
.
x
.
shape
+
(
-
1
,
)),
shape
=
shape
)
return
(
index_broadcast
<
x_broadcast
).
astype
(
self
.
mask_dtype
)
def
test_check_output
(
self
):
self
.
check_output
()
class
SequenceMaskTest1
(
SequenceMaskTestBase
):
def
initParameters
(
self
):
self
.
mask_dtype
=
'bool'
class
SequenceMaskTest2
(
SequenceMaskTestBase
):
def
initParameters
(
self
):
self
.
mask_dtype
=
'uint8'
class
SequenceMaskTest3
(
SequenceMaskTestBase
):
def
initParameters
(
self
):
self
.
mask_dtype
=
'int32'
class
SequenceMaskTest4
(
SequenceMaskTestBase
):
def
initParameters
(
self
):
self
.
mask_dtype
=
'float32'
class
SequenceMaskTest5
(
SequenceMaskTestBase
):
def
initParameters
(
self
):
self
.
mask_dtype
=
'float64'
class
SequenceMaskTest6
(
SequenceMaskTestBase
):
def
initParameters
(
self
):
self
.
maxlen
=
-
1
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