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
wzpei9
Paddle
提交
7c426be9
P
Paddle
项目概览
wzpei9
/
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
7c426be9
编写于
6月 13, 2018
作者:
Q
Qiao Longfei
提交者:
GitHub
6月 13, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #11342 from jacquesqiao/add-merge-splited-ids
Add merge_ids_op
上级
d3b0129c
e6f54d5a
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
297 addition
and
9 deletion
+297
-9
paddle/fluid/framework/executor.cc
paddle/fluid/framework/executor.cc
+5
-1
paddle/fluid/framework/operator.cc
paddle/fluid/framework/operator.cc
+21
-0
paddle/fluid/operators/merge_ids_op.cc
paddle/fluid/operators/merge_ids_op.cc
+128
-0
paddle/fluid/operators/merge_ids_op.h
paddle/fluid/operators/merge_ids_op.h
+92
-0
python/paddle/fluid/tests/unittests/test_merge_ids_op.py
python/paddle/fluid/tests/unittests/test_merge_ids_op.py
+38
-0
python/paddle/fluid/transpiler/distribute_transpiler.py
python/paddle/fluid/transpiler/distribute_transpiler.py
+13
-8
未找到文件。
paddle/fluid/framework/executor.cc
浏览文件 @
7c426be9
...
...
@@ -330,8 +330,12 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
}
for
(
auto
&
op
:
ctx
->
ops_
)
{
VLOG
(
3
)
<<
place_
<<
" "
<<
op
->
DebugStringEx
(
local_scope
);
VLOG
(
4
)
<<
place_
<<
" "
<<
op
->
DebugStringEx
(
local_scope
);
op
->
Run
(
*
local_scope
,
place_
);
// NOTE! Please do not delete this line, it's usefull because the debug
// string before and after op.run are different, after run the output
// will have right shape which is usefull for debug.
VLOG
(
3
)
<<
place_
<<
" "
<<
op
->
DebugStringEx
(
local_scope
);
if
(
FLAGS_benchmark
)
{
VLOG
(
2
)
<<
"Memory used after operator "
+
op
->
Type
()
+
" running: "
...
...
paddle/fluid/framework/operator.cc
浏览文件 @
7c426be9
...
...
@@ -69,6 +69,19 @@ static DDim GetDims(const Scope& scope, const std::string& name,
}
}
static
int
GetRowSize
(
const
Scope
&
scope
,
const
std
::
string
&
name
)
{
Variable
*
var
=
scope
.
FindVar
(
name
);
if
(
var
==
nullptr
)
{
return
-
1
;
}
if
(
var
->
IsType
<
SelectedRows
>
())
{
return
var
->
Get
<
SelectedRows
>
().
rows
().
size
();
}
return
-
1
;
}
static
LoD
GetLoD
(
const
Scope
&
scope
,
const
std
::
string
&
name
)
{
Variable
*
var
=
scope
.
FindVar
(
name
);
auto
default_lod
=
LoD
({{}});
...
...
@@ -153,6 +166,10 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const {
for
(
size_t
i
=
0
;
i
<
input
.
second
.
size
();
++
i
)
{
ss
<<
input
.
second
[
i
];
if
(
scope
)
{
int
row_size
=
GetRowSize
(
*
scope
,
input
.
second
[
i
]);
if
(
row_size
>=
0
)
{
ss
<<
"[row_size="
<<
row_size
<<
"]"
;
}
ss
<<
"["
<<
GetDims
(
*
scope
,
input
.
second
[
i
],
true
)
<<
"]"
;
ss
<<
"("
<<
GetLoD
(
*
scope
,
input
.
second
[
i
])
<<
")"
;
}
...
...
@@ -173,6 +190,10 @@ std::string OperatorBase::DebugStringEx(const Scope* scope) const {
for
(
size_t
i
=
0
;
i
<
output
.
second
.
size
();
++
i
)
{
ss
<<
output
.
second
[
i
];
if
(
scope
)
{
int
row_size
=
GetRowSize
(
*
scope
,
output
.
second
[
i
]);
if
(
row_size
>=
0
)
{
ss
<<
"[row_size="
<<
row_size
<<
"]"
;
}
ss
<<
"["
<<
GetDims
(
*
scope
,
output
.
second
[
i
],
true
)
<<
"]"
;
ss
<<
"("
<<
GetLoD
(
*
scope
,
output
.
second
[
i
])
<<
")"
;
}
...
...
paddle/fluid/operators/merge_ids_op.cc
0 → 100644
浏览文件 @
7c426be9
/* 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/merge_ids_op.h"
namespace
paddle
{
namespace
operators
{
class
MergeIdsOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
void
Make
()
override
{
AddInput
(
"Ids"
,
"(LoDTensor) the input ids with shape{batch_num, 1}"
);
AddInput
(
"X"
,
"(LoDTensors) multi input tensor with shape{batch_num, N}, N is the "
"size of embedding table"
)
.
AsDuplicable
();
AddOutput
(
"Out"
,
"(LoDTensor) The merged outputs of the input tensors."
);
AddComment
(
R"DOC(
Merge multi LoDTensor's into one according to Ids's shard num.
split_ids_op -> prefetch_op -> merge_ids_op
merge_ids_op should be used after split_ids_op and prefetch_op, split_ids_op
will split input Ids into multiple tensors according to Id's shard number.
prefetch_op will send them to parameter server to prefetch embedding value
back. During split, the order of ids is disordered. In merge_ids_op we use
the original Ids to restore the order of the fetched embedding value and
also pass the lod information to the merged output.
Example:
Ids = [1,2,3,4,5,6] # 3 shared
split_ids_op ->
Id0 = [3, 6] # id % 3 == 0
Id1 = [1, 4] # id % 3 == 1
Id2 = [2, 5] # id % 3 == 2
prefetch_op ->
X0 = [[0.3 0.3] # 3
[0.6 0.6]] # 6
X1 = [[0.1 0.1] # 1
[0.4 0.4]] # 4
X2 = [[0.2 0.2] # 2
[0.5 0.5]] # 5
merge_ids_op ->
Out = [[0.1 0.1] # 1
[0.2 0.2] # 2
[0.3 0.3] # 3
[0.4 0.4] # 4
[0.5 0.5] # 5
[0.6 0.6]] # 6
)DOC"
);
}
};
class
MergeIdsOp
:
public
framework
::
OperatorWithKernel
{
public:
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Ids"
),
"MergeIdsOp must has input Ids."
);
PADDLE_ENFORCE
(
ctx
->
HasInputs
(
"X"
),
"MergeIdsOp must has input X."
);
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
"Out"
),
"MergeIdsOp must has output Out."
);
auto
ids_var_type
=
ctx
->
GetInputsVarType
(
"Ids"
).
front
();
auto
ids_dims
=
ctx
->
GetInputDim
(
"Ids"
);
if
(
ids_var_type
==
framework
::
proto
::
VarType
::
LOD_TENSOR
)
{
PADDLE_ENFORCE_EQ
(
ids_dims
.
size
(),
2
);
PADDLE_ENFORCE_EQ
(
ids_dims
[
1
],
1
);
}
auto
x_var_type
=
ctx
->
GetInputsVarType
(
"X"
);
for
(
auto
&
var_type
:
x_var_type
)
{
PADDLE_ENFORCE_EQ
(
var_type
,
framework
::
proto
::
VarType
::
LOD_TENSOR
,
"input X only support lod tensors"
);
}
ctx
->
ShareLoD
(
"Ids"
,
"Out"
);
}
private:
framework
::
OpKernelType
GetExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
return
framework
::
OpKernelType
(
framework
::
ToDataType
(
ctx
.
MultiInput
<
framework
::
Tensor
>
(
"X"
).
front
()
->
type
()),
ctx
.
GetPlace
());
}
};
class
MergeIdsOpInferVarType
:
public
framework
::
VarTypeInference
{
public:
void
operator
()(
const
framework
::
OpDesc
&
op_desc
,
framework
::
BlockDesc
*
block
)
const
override
{
auto
*
input_var
=
block
->
Var
(
op_desc
.
Input
(
"Ids"
)[
0
]);
for
(
auto
&
out_var
:
op_desc
.
Output
(
"Out"
))
{
block
->
Var
(
out_var
)
->
SetType
(
input_var
->
GetType
());
}
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OPERATOR
(
merge_ids
,
ops
::
MergeIdsOp
,
ops
::
MergeIdsOpMaker
,
ops
::
MergeIdsOpInferVarType
);
REGISTER_OP_CPU_KERNEL
(
merge_ids
,
ops
::
MergeIdsOpKernel
<
paddle
::
platform
::
CPUPlace
,
float
>
);
paddle/fluid/operators/merge_ids_op.h
0 → 100644
浏览文件 @
7c426be9
/* 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 <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/tensor_util.h"
#include "paddle/fluid/operators/math/selected_rows_functor.h"
namespace
paddle
{
namespace
operators
{
template
<
typename
DeviceContext
,
typename
T
>
class
MergeIdsOpKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
place
=
ctx
.
GetPlace
();
if
(
!
platform
::
is_cpu_place
(
place
))
{
PADDLE_THROW
(
"MergeIds do not support GPU kernel"
);
}
VLOG
(
3
)
<<
"run in MergeIdsOpKernel"
;
const
auto
*
ids_var
=
ctx
.
InputVar
(
"Ids"
);
PADDLE_ENFORCE
(
ids_var
->
IsType
<
framework
::
LoDTensor
>
(),
"only support to merge Ids of LoDTensor"
);
const
auto
&
ids_tensor
=
ids_var
->
Get
<
framework
::
LoDTensor
>
();
const
auto
&
ids_dims
=
ids_tensor
.
dims
();
const
int64_t
*
ids
=
ids_tensor
.
data
<
int64_t
>
();
auto
x_tensors
=
ctx
.
MultiInput
<
framework
::
LoDTensor
>
(
"X"
);
auto
*
out
=
ctx
.
Output
<
framework
::
LoDTensor
>
(
"Out"
);
int
batch_size
=
0
;
int
embedding_size
=
0
;
for
(
auto
&
input
:
x_tensors
)
{
if
(
framework
::
product
(
input
->
dims
())
!=
0
)
{
if
(
embedding_size
==
0
)
{
embedding_size
=
input
->
dims
()[
1
];
}
PADDLE_ENFORCE_EQ
(
embedding_size
,
input
->
dims
()[
1
],
"embedding size of all input should be the same"
);
batch_size
+=
input
->
dims
()[
0
];
}
}
PADDLE_ENFORCE_EQ
(
batch_size
,
ids_dims
[
0
],
"the batch size of ids and merged embedding value should be the same"
);
const
size_t
shard_num
=
x_tensors
.
size
();
if
(
shard_num
==
1
)
{
VLOG
(
3
)
<<
"only one shard, we can copy the data directly"
;
TensorCopy
(
*
x_tensors
[
0
],
place
,
out
);
}
else
{
std
::
vector
<
int
>
in_indexs
(
shard_num
,
0
);
auto
*
out_data
=
out
->
mutable_data
<
T
>
(
framework
::
make_ddim
({
batch_size
,
embedding_size
}),
place
);
// copy data from ins[shard_num] to out.
for
(
int
i
=
0
;
i
<
ids_dims
[
0
];
++
i
)
{
int64_t
id
=
ids
[
i
];
size_t
shard_id
=
static_cast
<
size_t
>
(
id
)
%
shard_num
;
int
index
=
in_indexs
[
shard_id
];
memcpy
(
out_data
+
embedding_size
*
i
,
x_tensors
[
shard_id
]
->
data
<
T
>
()
+
index
*
embedding_size
,
sizeof
(
T
)
*
embedding_size
);
in_indexs
[
shard_id
]
+=
1
;
}
for
(
size_t
i
=
0
;
i
<
shard_num
;
++
i
)
{
PADDLE_ENFORCE_EQ
(
in_indexs
[
i
],
x_tensors
[
i
]
->
dims
()[
0
],
"after merge, all data in x_tensor should be used"
);
}
}
}
};
}
// namespace operators
}
// namespace paddle
python/paddle/fluid/tests/unittests/test_merge_ids_op.py
0 → 100644
浏览文件 @
7c426be9
# 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
TestMergeIdsOp
(
OpTest
):
def
setUp
(
self
):
self
.
op_type
=
"merge_ids"
ids
=
np
.
array
([[
0
],
[
2
],
[
2
],
[
3
],
[
5
],
[
5
],
[
6
]]).
astype
(
'int64'
)
x0
=
np
.
array
([[
0.1
,
0.2
],
[
0.2
,
0.3
],
[
0.3
,
0.4
]]).
astype
(
'float32'
)
x1
=
np
.
array
([]).
astype
(
'float32'
)
x2
=
np
.
array
([[
0.4
,
0.5
],
[
0.4
,
0.5
],
[
0.5
,
0.6
],
[
0.5
,
0.6
]]).
astype
(
'float32'
)
out
=
np
.
array
([[
0.1
,
0.2
],
[
0.4
,
0.5
],
[
0.4
,
0.5
],
[
0.2
,
0.3
],
[
0.5
,
0.6
],
[
0.5
,
0.6
],
[
0.3
,
0.4
]]).
astype
(
'float32'
)
self
.
inputs
=
{
'Ids'
:
ids
,
"X"
:
[(
'x0'
,
x0
),
(
'x1'
,
x1
),
(
'x2'
,
x2
)]}
self
.
outputs
=
{
'Out'
:
out
}
def
test_check_output
(
self
):
self
.
check_output
()
if
__name__
==
'__main__'
:
unittest
.
main
()
python/paddle/fluid/transpiler/distribute_transpiler.py
浏览文件 @
7c426be9
...
...
@@ -629,7 +629,7 @@ class DistributeTranspiler:
if
op
.
type
==
LOOKUP_TABLE_TYPE
:
continue_search_lookup_table_op
=
True
op_index
=
list
(
all_ops
).
index
(
op
)
lookup_table_
op_index
=
list
(
all_ops
).
index
(
op
)
ids_name
=
op
.
input
(
"Ids"
)
out_name
=
op
.
output
(
"Out"
)
...
...
@@ -649,7 +649,7 @@ class DistributeTranspiler:
# insert split_ids_op
program
.
global_block
().
insert_op
(
index
=
op_index
,
index
=
lookup_table_
op_index
,
type
=
"split_ids"
,
inputs
=
{
'Ids'
:
[
...
...
@@ -661,7 +661,7 @@ class DistributeTranspiler:
# insert prefetch_op
program
.
global_block
().
insert_op
(
index
=
op_index
+
1
,
index
=
lookup_table_
op_index
+
1
,
type
=
"prefetch"
,
inputs
=
{
'X'
:
prefetch_input_vars
},
outputs
=
{
"Out"
:
prefetch_output_vars
},
...
...
@@ -672,16 +672,21 @@ class DistributeTranspiler:
# insert concat_op
program
.
global_block
().
insert_op
(
index
=
op_index
+
2
,
type
=
"concat"
,
inputs
=
{
'X'
:
prefetch_output_vars
},
index
=
lookup_table_op_index
+
2
,
type
=
"merge_ids"
,
inputs
=
{
'Ids'
:
[
program
.
global_block
().
vars
[
varname
]
for
varname
in
ids_name
],
'X'
:
prefetch_output_vars
},
outputs
=
{
"Out"
:
[
program
.
global_block
().
vars
[
varname
]
for
varname
in
out_name
]
},
attrs
=
{
"axis"
:
0
})
})
# delete lookup_table_op
delete_ops
(
program
.
global_block
(),
[
op
])
...
...
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