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
21ec93aa
P
Paddle
项目概览
PaddlePaddle
/
Paddle
1 年多 前同步成功
通知
2302
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
未验证
提交
21ec93aa
编写于
9月 19, 2018
作者:
Q
Qingsheng Li
提交者:
GitHub
9月 19, 2018
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[WIP]Sequence Scatter Op (#12625)
Sequence Scatter Op
上级
103deb11
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
415 addition
and
0 deletion
+415
-0
paddle/fluid/API.spec
paddle/fluid/API.spec
+1
-0
paddle/fluid/operators/sequence_scatter_op.cc
paddle/fluid/operators/sequence_scatter_op.cc
+156
-0
paddle/fluid/operators/sequence_scatter_op.h
paddle/fluid/operators/sequence_scatter_op.h
+122
-0
python/paddle/fluid/layers/nn.py
python/paddle/fluid/layers/nn.py
+61
-0
python/paddle/fluid/tests/unittests/test_layers.py
python/paddle/fluid/tests/unittests/test_layers.py
+24
-0
python/paddle/fluid/tests/unittests/test_sequence_scatter_op.py
.../paddle/fluid/tests/unittests/test_sequence_scatter_op.py
+51
-0
未找到文件。
paddle/fluid/API.spec
浏览文件 @
21ec93aa
...
...
@@ -154,6 +154,7 @@ paddle.fluid.layers.image_resize_short ArgSpec(args=['input', 'out_short_len', '
paddle.fluid.layers.resize_bilinear ArgSpec(args=['input', 'out_shape', 'scale', 'name'], varargs=None, keywords=None, defaults=(None, None, None))
paddle.fluid.layers.gather ArgSpec(args=['input', 'index'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.scatter ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.sequence_scatter ArgSpec(args=['input', 'index', 'updates', 'name'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.random_crop ArgSpec(args=['x', 'shape', 'seed'], varargs=None, keywords=None, defaults=(None,))
paddle.fluid.layers.mean_iou ArgSpec(args=['input', 'label', 'num_classes'], varargs=None, keywords=None, defaults=None)
paddle.fluid.layers.relu ArgSpec(args=['x', 'name'], varargs=None, keywords=None, defaults=(None,))
...
...
paddle/fluid/operators/sequence_scatter_op.cc
0 → 100644
浏览文件 @
21ec93aa
/* 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_scatter_op.h"
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/gather.h"
#include "paddle/fluid/operators/scatter.h"
namespace
paddle
{
namespace
operators
{
using
Tensor
=
framework
::
Tensor
;
using
LoDTensor
=
framework
::
LoDTensor
;
class
SequenceScatterOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
void
Make
()
override
{
AddInput
(
"X"
,
"(Tensor) The source input of sequence scatter op"
);
AddInput
(
"Ids"
,
"(LoDTensor) The index input of sequence scatter op where X"
" will be updated, must be a LoDTensor"
);
AddInput
(
"Updates"
,
"(LoDTensor) The values to scatter to the input tensor "
"X, must be a LoDTensor with the same LoD information as Ids"
);
AddOutput
(
"Out"
,
"(Tensor) The output tensor of sequence scatter op, which "
"has the same dims as X"
);
AddComment
(
R"DOC(
Sequence Scatter Operator.
This operator scatters the Updates tensor to the input X. It uses the LoD
information of Ids to select the rows to update, and use the values in Ids as
the columns to update in each row of X.
Following are cases to better explain how this works:
Example 1:
Given an all-ones Tensor input(X)
X.data = [[1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
[1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
[1.0, 1.0, 1.0, 1.0, 1.0, 1.0]]
X.dims = [3, 6]
a LoDTensor input(Ids)
Ids.data = [[0], [1], [2], [5], [4], [3], [2], [1], [3], [2], [5], [4]]
Ids.lod = [[0, 3, 8, 12]]
and a Tensor input(Updates)
Updates.data = [[0.3], [0.3], [0.4], [0.1], [0.2], [0.3], [0.4], [0.0], [0.2], [0.3], [0.1], [0.4]]
Updates.lod = [[ 0, 3, 8, 12]]
then we get an output Tensor
Out.data = [[1.3, 1.3, 1.4, 1.0, 1.0, 1.0],
[1.0, 1.0, 1.4, 1.3, 1.2, 1.1],
[1.0, 1.0, 1.3, 1.2, 1.4, 1.1]]
Out.dims = X.dims = [3, 6]
)DOC"
);
}
};
class
SequenceScatterOp
:
public
framework
::
OperatorWithKernel
{
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
// Enforce has inputs and outputs
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"X"
),
"Input(X) of SequenceScatterOp should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Ids"
),
"Input(Ids) of SequenceScatterOp should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Updates"
),
"Input(Updates) of SequenceScatterOp should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
"Out"
),
"Output(Out) of SequenceScatterOp should not be null."
);
// Set output dim the same as input
auto
ref_dims
=
ctx
->
GetInputDim
(
"X"
);
ctx
->
SetOutputDim
(
"Out"
,
ref_dims
);
// Enforce the Updates and Ids are the same shape
PADDLE_ENFORCE_EQ
(
ctx
->
GetInputDim
(
"Updates"
)[
0
],
ctx
->
GetInputDim
(
"Ids"
)[
0
],
"Updates and Ids should have same shape."
);
// Enforce LoD of ids and updates be the same
if
(
ctx
->
IsRuntime
())
{
framework
::
Variable
*
ids_var
=
boost
::
get
<
framework
::
Variable
*>
(
ctx
->
GetInputVarPtrs
(
"Ids"
)[
0
]);
framework
::
Variable
*
updates_var
=
boost
::
get
<
framework
::
Variable
*>
(
ctx
->
GetInputVarPtrs
(
"Updates"
)[
0
]);
auto
&
ids_lod
=
ids_var
->
Get
<
LoDTensor
>
().
lod
();
auto
&
updates_lod
=
updates_var
->
Get
<
LoDTensor
>
().
lod
();
PADDLE_ENFORCE_EQ
(
ids_lod
.
size
(),
1
,
"Currently only level 1 LoD could be"
" processed by sequence scatter op."
);
PADDLE_ENFORCE_EQ
(
updates_lod
.
size
(),
1
,
"Currently only level 1 LoD "
"could be processed by sequence scatter op."
);
}
}
protected:
framework
::
OpKernelType
GetExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
return
framework
::
OpKernelType
(
framework
::
ToDataType
(
ctx
.
Input
<
Tensor
>
(
"X"
)
->
type
()),
platform
::
CPUPlace
());
}
};
class
SequenceScatterGradOp
:
public
framework
::
OperatorWithKernel
{
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
ctx
->
SetOutputDim
(
framework
::
GradVarName
(
"Updates"
),
ctx
->
GetInputDim
(
"Updates"
));
ctx
->
SetOutputDim
(
framework
::
GradVarName
(
"X"
),
ctx
->
GetInputDim
(
"X"
));
}
protected:
framework
::
OpKernelType
GetExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
return
framework
::
OpKernelType
(
framework
::
ToDataType
(
ctx
.
Input
<
Tensor
>
(
"X"
)
->
type
()),
platform
::
CPUPlace
());
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OPERATOR
(
sequence_scatter
,
ops
::
SequenceScatterOp
,
ops
::
SequenceScatterOpMaker
,
paddle
::
framework
::
DefaultGradOpDescMaker
<
true
>
);
REGISTER_OPERATOR
(
sequence_scatter_grad
,
ops
::
SequenceScatterGradOp
);
REGISTER_OP_CPU_KERNEL
(
sequence_scatter
,
ops
::
SequenceScatterOpKernel
<
float
>
,
ops
::
SequenceScatterOpKernel
<
double
>
,
ops
::
SequenceScatterOpKernel
<
int
>
,
ops
::
SequenceScatterOpKernel
<
int64_t
>
);
REGISTER_OP_CPU_KERNEL
(
sequence_scatter_grad
,
ops
::
SequenceScatterGradientOpKernel
<
float
>
,
ops
::
SequenceScatterGradientOpKernel
<
double
>
,
ops
::
SequenceScatterGradientOpKernel
<
int
>
,
ops
::
SequenceScatterGradientOpKernel
<
int64_t
>
);
paddle/fluid/operators/sequence_scatter_op.h
0 → 100644
浏览文件 @
21ec93aa
/* 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
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/gather.h"
#include "paddle/fluid/operators/scatter.h"
namespace
paddle
{
namespace
operators
{
using
Tensor
=
framework
::
Tensor
;
using
LoDTensor
=
framework
::
LoDTensor
;
template
<
typename
T
>
class
SequenceScatterOpKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
*
x
=
ctx
.
Input
<
Tensor
>
(
"X"
);
auto
*
ids
=
ctx
.
Input
<
LoDTensor
>
(
"Ids"
);
auto
*
updates
=
ctx
.
Input
<
LoDTensor
>
(
"Updates"
);
auto
*
out
=
ctx
.
Output
<
Tensor
>
(
"Out"
);
auto
&
ids_lod
=
ids
->
lod
();
// Initialize out as same as x
out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
framework
::
TensorCopySync
(
*
x
,
ctx
.
GetPlace
(),
out
);
auto
x_dims
=
x
->
dims
();
auto
out_dims
=
out
->
dims
();
for
(
int
i
=
0
;
i
<
x_dims
.
size
();
++
i
)
PADDLE_ENFORCE
(
x_dims
[
i
]
==
out_dims
[
i
],
"Input and output shape of "
"sequence scatter op must exactly be the same."
);
size_t
slice_size
=
1
;
for
(
int
i
=
1
;
i
<
x_dims
.
size
();
++
i
)
slice_size
*=
x_dims
[
i
];
auto
lod_vec
=
ids_lod
[
0
];
unsigned
int
seg
=
0
;
for
(
int
i
=
0
;
i
<
ids
->
dims
()[
0
];
++
i
)
{
PADDLE_ENFORCE_LT
(
seg
,
lod_vec
.
size
()
-
1
,
"Segment num must not exceed batch size.
\n
"
);
int
lower_bound
=
lod_vec
[
seg
];
int
upper_bound
=
lod_vec
[
seg
+
1
];
if
(
i
>=
lower_bound
&&
i
<
upper_bound
)
{
T
*
p_out
=
out
->
data
<
T
>
();
const
T
*
p_updates
=
updates
->
data
<
T
>
();
const
int64_t
*
p_index
=
ids
->
data
<
int64_t
>
();
p_out
[
seg
*
slice_size
+
p_index
[
i
]]
+=
p_updates
[
i
];
}
else
{
++
seg
;
--
i
;
}
}
}
};
template
<
typename
T
>
class
SequenceScatterGradientOpKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
PADDLE_ENFORCE
(
platform
::
is_cpu_place
(
ctx
.
GetPlace
()),
"This kernel only runs on CPU."
);
auto
*
dX
=
ctx
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"X"
));
auto
*
dUpdates
=
ctx
.
Output
<
LoDTensor
>
(
framework
::
GradVarName
(
"Updates"
));
auto
*
ids
=
ctx
.
Input
<
LoDTensor
>
(
"Ids"
);
auto
*
dOut
=
ctx
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"Out"
));
auto
&
ids_lod
=
ids
->
lod
();
dX
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
framework
::
TensorCopySync
(
*
dOut
,
ctx
.
GetPlace
(),
dX
);
dUpdates
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
dx_dims
=
dX
->
dims
();
auto
dout_dims
=
dOut
->
dims
();
for
(
int
i
=
0
;
i
<
dx_dims
.
size
();
++
i
)
PADDLE_ENFORCE
(
dx_dims
[
i
]
==
dout_dims
[
i
],
"Input and output shape of "
"sequence scatter grad op must exactly be the same."
);
size_t
slice_size
=
1
;
for
(
int
i
=
1
;
i
<
dx_dims
.
size
();
++
i
)
slice_size
*=
dx_dims
[
i
];
auto
lod_vec
=
ids_lod
[
0
];
unsigned
int
seg
=
0
;
for
(
int
i
=
0
;
i
<
ids
->
dims
()[
0
];
++
i
)
{
PADDLE_ENFORCE_LT
(
seg
,
lod_vec
.
size
()
-
1
,
"Segment num must not exceed batch size.
\n
"
);
int
lower_bound
=
lod_vec
[
seg
];
int
upper_bound
=
lod_vec
[
seg
+
1
];
if
(
i
>=
lower_bound
&&
i
<
upper_bound
)
{
const
T
*
p_dOut
=
dOut
->
data
<
T
>
();
const
int64_t
*
p_index
=
ids
->
data
<
int64_t
>
();
T
*
p_dUpdates
=
dUpdates
->
data
<
T
>
();
p_dUpdates
[
i
]
=
p_dOut
[
seg
*
slice_size
+
p_index
[
i
]];
}
else
{
++
seg
;
--
i
;
}
}
}
};
}
// namespace operators
}
// namespace paddle
python/paddle/fluid/layers/nn.py
浏览文件 @
21ec93aa
...
...
@@ -100,6 +100,7 @@ __all__ = [
'resize_bilinear'
,
'gather'
,
'scatter'
,
'sequence_scatter'
,
'random_crop'
,
'mean_iou'
,
'relu'
,
...
...
@@ -5425,6 +5426,66 @@ def scatter(input, index, updates, name=None):
return
out
def
sequence_scatter
(
input
,
index
,
updates
,
name
=
None
):
"""
**Sequence Scatter Layer**
This operator scatters the Updates tensor to the input X. It uses the LoD
information of Ids to select the rows to update, and use the values in Ids as
the columns to update in each row of X.
Here is an example:
Given the following input:
.. code-block:: text
input.data = [[1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
[1.0, 1.0, 1.0, 1.0, 1.0, 1.0],
[1.0, 1.0, 1.0, 1.0, 1.0, 1.0]]
input.dims = [3, 6]
index.data = [[0], [1], [2], [5], [4], [3], [2], [1], [3], [2], [5], [4]]
index.lod = [[0, 3, 8, 12]]
updates.data = [[0.3], [0.3], [0.4], [0.1], [0.2], [0.3], [0.4], [0.0], [0.2], [0.3], [0.1], [0.4]]
updates.lod = [[ 0, 3, 8, 12]]
Then we have the output:
.. code-block:: text
out.data = [[1.3, 1.3, 1.4, 1.0, 1.0, 1.0],
[1.0, 1.0, 1.4, 1.3, 1.2, 1.1],
[1.0, 1.0, 1.3, 1.2, 1.4, 1.1]]
out.dims = X.dims = [3, 6]
Args:
input (Variable): The source input with rank>=1.
index (Variable): A LoD Tensor. The index input of sequence scatter op
where input will be updated. The index input with rank=1. Its dtype
should be int32 or int64 as it is used as indexes.
updates (Variable): A LoD Tensor. The values to scatter to the input
tensor X, must be a LoDTensor with the same LoD information as index.
name (str|None): The output variable name. Default None.
Returns:
output (Variable): The output is a tensor with the same shape as input.
Examples:
.. code-block:: python
output = fluid.layers.sequence_scatter(input, index, updates)
"""
helper
=
LayerHelper
(
'sequence_scatter'
,
**
locals
())
dtype
=
helper
.
input_dtype
()
out
=
helper
.
create_tmp_variable
(
dtype
)
helper
.
append_op
(
type
=
"sequence_scatter"
,
inputs
=
{
"X"
:
input
,
"Ids"
:
index
,
"Updates"
:
updates
},
outputs
=
{
"Out"
:
out
})
return
out
@
templatedoc
()
def
random_crop
(
x
,
shape
,
seed
=
None
):
"""
...
...
python/paddle/fluid/tests/unittests/test_layers.py
浏览文件 @
21ec93aa
...
...
@@ -382,6 +382,30 @@ class TestBook(unittest.TestCase):
self
.
assertIsNotNone
(
out
)
print
(
str
(
program
))
def
test_sequence_scatter
(
self
):
program
=
Program
()
with
program_guard
(
program
):
x
=
layers
.
data
(
name
=
'x'
,
shape
=
[
3
,
6
],
append_batch_size
=
False
,
dtype
=
'float32'
)
idx
=
layers
.
data
(
name
=
'idx'
,
shape
=
[
12
,
1
],
append_batch_size
=
False
,
dtype
=
'int32'
,
lod_level
=
1
)
updates
=
layers
.
data
(
name
=
'updates'
,
shape
=
[
12
,
1
],
append_batch_size
=
False
,
dtype
=
'float32'
,
lod_level
=
1
)
out
=
layers
.
sequence_scatter
(
input
=
x
,
index
=
idx
,
updates
=
updates
)
self
.
assertIsNotNone
(
out
)
print
(
str
(
program
))
def
test_lod_reset
(
self
):
program
=
Program
()
with
program_guard
(
program
):
...
...
python/paddle/fluid/tests/unittests/test_sequence_scatter_op.py
0 → 100644
浏览文件 @
21ec93aa
# 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.
import
unittest
import
numpy
as
np
from
op_test
import
OpTest
class
TestSequenceScatterOp
(
OpTest
):
def
setUp
(
self
):
self
.
op_type
=
"sequence_scatter"
X_data
=
np
.
random
.
uniform
(
0.1
,
1.0
,
[
3
,
6
]).
astype
(
'float32'
)
Ids_data
=
np
.
array
([[
0
],
[
1
],
[
2
],
[
5
],
[
4
],
[
3
],
[
2
],
[
1
],
[
3
],
[
2
],
[
5
],
[
4
]]).
astype
(
'int64'
)
Ids_lod
=
[[
3
,
5
,
4
]]
Updates_data
=
np
.
random
.
uniform
(
0.1
,
1.0
,
[
12
,
1
]).
astype
(
'float32'
)
Updates_lod
=
Ids_lod
Out_data
=
np
.
copy
(
X_data
)
Out_data
[
0
][
Ids_data
[
0
:
3
]]
+=
Updates_data
[
0
:
3
]
Out_data
[
1
][
Ids_data
[
3
:
8
]]
+=
Updates_data
[
3
:
8
]
Out_data
[
2
][
Ids_data
[
8
:]]
+=
Updates_data
[
8
:]
self
.
inputs
=
{
'X'
:
X_data
,
'Ids'
:
(
Ids_data
,
Ids_lod
),
'Updates'
:
(
Updates_data
,
Updates_lod
)
}
self
.
outputs
=
{
'Out'
:
Out_data
}
def
test_check_output
(
self
):
self
.
check_output
()
def
test_check_grad
(
self
):
self
.
check_grad
([
'Updates'
],
'Out'
,
in_place
=
True
)
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