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
9ff558a4
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2297
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
9ff558a4
编写于
4月 22, 2020
作者:
H
hutuxian
提交者:
GitHub
4月 22, 2020
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Optimize DataFeed (#23957)
* Make batch_float_feasigns & batch_uint64_feasigns as member variable
上级
5cccc69f
变更
2
显示空白变更内容
内联
并排
Showing
2 changed file
with
59 addition
and
42 deletion
+59
-42
paddle/fluid/framework/data_feed.cc
paddle/fluid/framework/data_feed.cc
+55
-42
paddle/fluid/framework/data_feed.h
paddle/fluid/framework/data_feed.h
+4
-0
未找到文件。
paddle/fluid/framework/data_feed.cc
浏览文件 @
9ff558a4
...
...
@@ -798,6 +798,19 @@ void MultiSlotInMemoryDataFeed::Init(
}
}
feed_vec_
.
resize
(
use_slots_
.
size
());
const
int
kEstimatedFeasignNumPerSlot
=
5
;
// Magic Number
for
(
size_t
i
=
0
;
i
<
all_slot_num
;
i
++
)
{
batch_float_feasigns_
.
push_back
(
std
::
vector
<
float
>
());
batch_uint64_feasigns_
.
push_back
(
std
::
vector
<
uint64_t
>
());
batch_float_feasigns_
[
i
].
reserve
(
default_batch_size_
*
kEstimatedFeasignNumPerSlot
);
batch_uint64_feasigns_
[
i
].
reserve
(
default_batch_size_
*
kEstimatedFeasignNumPerSlot
);
offset_
.
push_back
(
std
::
vector
<
size_t
>
());
offset_
[
i
].
reserve
(
default_batch_size_
+
1
);
// Each lod info will prepend a zero
}
visit_
.
resize
(
all_slot_num
,
false
);
pipe_command_
=
data_feed_desc
.
pipe_command
();
finish_init_
=
true
;
}
...
...
@@ -989,13 +1002,12 @@ bool MultiSlotInMemoryDataFeed::ParseOneInstance(Record* instance) {
void
MultiSlotInMemoryDataFeed
::
PutToFeedVec
(
const
std
::
vector
<
Record
>&
ins_vec
)
{
#ifdef _LINUX
std
::
vector
<
std
::
vector
<
float
>>
batch_float_feasigns
(
use_slots_
.
size
(),
std
::
vector
<
float
>
());
std
::
vector
<
std
::
vector
<
uint64_t
>>
batch_uint64_feasigns
(
use_slots_
.
size
(),
std
::
vector
<
uint64_t
>
());
std
::
vector
<
std
::
vector
<
size_t
>>
offset
(
use_slots_
.
size
(),
std
::
vector
<
size_t
>
{
0
});
std
::
vector
<
bool
>
visit
(
use_slots_
.
size
(),
false
);
for
(
size_t
i
=
0
;
i
<
batch_float_feasigns_
.
size
();
++
i
)
{
batch_float_feasigns_
[
i
].
clear
();
batch_uint64_feasigns_
[
i
].
clear
();
offset_
[
i
].
clear
();
offset_
[
i
].
push_back
(
0
);
}
ins_content_vec_
.
clear
();
ins_content_vec_
.
reserve
(
ins_vec
.
size
());
ins_id_vec_
.
clear
();
...
...
@@ -1005,30 +1017,31 @@ void MultiSlotInMemoryDataFeed::PutToFeedVec(
ins_id_vec_
.
push_back
(
r
.
ins_id_
);
ins_content_vec_
.
push_back
(
r
.
content_
);
for
(
auto
&
item
:
r
.
float_feasigns_
)
{
batch_float_feasigns
[
item
.
slot
()].
push_back
(
item
.
sign
().
float_feasign_
);
visit
[
item
.
slot
()]
=
true
;
batch_float_feasigns
_
[
item
.
slot
()].
push_back
(
item
.
sign
().
float_feasign_
);
visit
_
[
item
.
slot
()]
=
true
;
}
for
(
auto
&
item
:
r
.
uint64_feasigns_
)
{
batch_uint64_feasigns
[
item
.
slot
()].
push_back
(
item
.
sign
().
uint64_feasign_
);
visit
[
item
.
slot
()]
=
true
;
batch_uint64_feasigns_
[
item
.
slot
()].
push_back
(
item
.
sign
().
uint64_feasign_
);
visit_
[
item
.
slot
()]
=
true
;
}
for
(
size_t
j
=
0
;
j
<
use_slots_
.
size
();
++
j
)
{
const
auto
&
type
=
all_slots_type_
[
j
];
if
(
visit
[
j
])
{
visit
[
j
]
=
false
;
if
(
visit
_
[
j
])
{
visit
_
[
j
]
=
false
;
}
else
{
// fill slot value with default value 0
if
(
type
[
0
]
==
'f'
)
{
// float
batch_float_feasigns
[
j
].
push_back
(
0.0
);
batch_float_feasigns
_
[
j
].
push_back
(
0.0
);
}
else
if
(
type
[
0
]
==
'u'
)
{
// uint64
batch_uint64_feasigns
[
j
].
push_back
(
0
);
batch_uint64_feasigns
_
[
j
].
push_back
(
0
);
}
}
// get offset of this ins in this slot
if
(
type
[
0
]
==
'f'
)
{
// float
offset
[
j
].
push_back
(
batch_float_feasigns
[
j
].
size
());
offset
_
[
j
].
push_back
(
batch_float_feasigns_
[
j
].
size
());
}
else
if
(
type
[
0
]
==
'u'
)
{
// uint64
offset
[
j
].
push_back
(
batch_uint64_feasigns
[
j
].
size
());
offset
_
[
j
].
push_back
(
batch_uint64_feasigns_
[
j
].
size
());
}
}
}
...
...
@@ -1037,21 +1050,21 @@ void MultiSlotInMemoryDataFeed::PutToFeedVec(
if
(
feed_vec_
[
i
]
==
nullptr
)
{
continue
;
}
int
total_instance
=
offset
[
i
].
back
();
int
total_instance
=
offset
_
[
i
].
back
();
const
auto
&
type
=
all_slots_type_
[
i
];
if
(
type
[
0
]
==
'f'
)
{
// float
float
*
feasign
=
batch_float_feasigns
[
i
].
data
();
float
*
feasign
=
batch_float_feasigns
_
[
i
].
data
();
float
*
tensor_ptr
=
feed_vec_
[
i
]
->
mutable_data
<
float
>
({
total_instance
,
1
},
this
->
place_
);
CopyToFeedTensor
(
tensor_ptr
,
feasign
,
total_instance
*
sizeof
(
float
));
}
else
if
(
type
[
0
]
==
'u'
)
{
// uint64
// no uint64_t type in paddlepaddle
uint64_t
*
feasign
=
batch_uint64_feasigns
[
i
].
data
();
uint64_t
*
feasign
=
batch_uint64_feasigns
_
[
i
].
data
();
int64_t
*
tensor_ptr
=
feed_vec_
[
i
]
->
mutable_data
<
int64_t
>
(
{
total_instance
,
1
},
this
->
place_
);
CopyToFeedTensor
(
tensor_ptr
,
feasign
,
total_instance
*
sizeof
(
int64_t
));
}
auto
&
slot_offset
=
offset
[
i
];
auto
&
slot_offset
=
offset
_
[
i
];
LoD
data_lod
{
slot_offset
};
feed_vec_
[
i
]
->
set_lod
(
data_lod
);
if
(
use_slots_is_dense_
[
i
])
{
...
...
@@ -1427,13 +1440,12 @@ int PaddleBoxDataFeed::GetCurrentPhase() {
void
PaddleBoxDataFeed
::
PutToFeedVec
(
const
std
::
vector
<
Record
*>&
ins_vec
)
{
#ifdef _LINUX
std
::
vector
<
std
::
vector
<
float
>>
batch_float_feasigns
(
use_slots_
.
size
(),
std
::
vector
<
float
>
());
std
::
vector
<
std
::
vector
<
uint64_t
>>
batch_uint64_feasigns
(
use_slots_
.
size
(),
std
::
vector
<
uint64_t
>
());
std
::
vector
<
std
::
vector
<
size_t
>>
offset
(
use_slots_
.
size
(),
std
::
vector
<
size_t
>
{
0
});
std
::
vector
<
bool
>
visit
(
use_slots_
.
size
(),
false
);
for
(
size_t
i
=
0
;
i
<
batch_float_feasigns_
.
size
();
++
i
)
{
batch_float_feasigns_
[
i
].
clear
();
batch_uint64_feasigns_
[
i
].
clear
();
offset_
[
i
].
clear
();
offset_
[
i
].
push_back
(
0
);
}
ins_content_vec_
.
clear
();
ins_content_vec_
.
reserve
(
ins_vec
.
size
());
ins_id_vec_
.
clear
();
...
...
@@ -1443,30 +1455,31 @@ void PaddleBoxDataFeed::PutToFeedVec(const std::vector<Record*>& ins_vec) {
ins_id_vec_
.
push_back
(
r
->
ins_id_
);
ins_content_vec_
.
push_back
(
r
->
content_
);
for
(
auto
&
item
:
r
->
float_feasigns_
)
{
batch_float_feasigns
[
item
.
slot
()].
push_back
(
item
.
sign
().
float_feasign_
);
visit
[
item
.
slot
()]
=
true
;
batch_float_feasigns
_
[
item
.
slot
()].
push_back
(
item
.
sign
().
float_feasign_
);
visit
_
[
item
.
slot
()]
=
true
;
}
for
(
auto
&
item
:
r
->
uint64_feasigns_
)
{
batch_uint64_feasigns
[
item
.
slot
()].
push_back
(
item
.
sign
().
uint64_feasign_
);
visit
[
item
.
slot
()]
=
true
;
batch_uint64_feasigns_
[
item
.
slot
()].
push_back
(
item
.
sign
().
uint64_feasign_
);
visit_
[
item
.
slot
()]
=
true
;
}
for
(
size_t
j
=
0
;
j
<
use_slots_
.
size
();
++
j
)
{
const
auto
&
type
=
all_slots_type_
[
j
];
if
(
visit
[
j
])
{
visit
[
j
]
=
false
;
if
(
visit
_
[
j
])
{
visit
_
[
j
]
=
false
;
}
else
{
// fill slot value with default value 0
if
(
type
[
0
]
==
'f'
)
{
// float
batch_float_feasigns
[
j
].
push_back
(
0.0
);
batch_float_feasigns
_
[
j
].
push_back
(
0.0
);
}
else
if
(
type
[
0
]
==
'u'
)
{
// uint64
batch_uint64_feasigns
[
j
].
push_back
(
0
);
batch_uint64_feasigns
_
[
j
].
push_back
(
0
);
}
}
// get offset of this ins in this slot
if
(
type
[
0
]
==
'f'
)
{
// float
offset
[
j
].
push_back
(
batch_float_feasigns
[
j
].
size
());
offset
_
[
j
].
push_back
(
batch_float_feasigns_
[
j
].
size
());
}
else
if
(
type
[
0
]
==
'u'
)
{
// uint64
offset
[
j
].
push_back
(
batch_uint64_feasigns
[
j
].
size
());
offset
_
[
j
].
push_back
(
batch_uint64_feasigns_
[
j
].
size
());
}
}
}
...
...
@@ -1475,21 +1488,21 @@ void PaddleBoxDataFeed::PutToFeedVec(const std::vector<Record*>& ins_vec) {
if
(
feed_vec_
[
i
]
==
nullptr
)
{
continue
;
}
int
total_instance
=
offset
[
i
].
back
();
int
total_instance
=
offset
_
[
i
].
back
();
const
auto
&
type
=
all_slots_type_
[
i
];
if
(
type
[
0
]
==
'f'
)
{
// float
float
*
feasign
=
batch_float_feasigns
[
i
].
data
();
float
*
feasign
=
batch_float_feasigns
_
[
i
].
data
();
float
*
tensor_ptr
=
feed_vec_
[
i
]
->
mutable_data
<
float
>
({
total_instance
,
1
},
this
->
place_
);
CopyToFeedTensor
(
tensor_ptr
,
feasign
,
total_instance
*
sizeof
(
float
));
}
else
if
(
type
[
0
]
==
'u'
)
{
// uint64
// no uint64_t type in paddlepaddle
uint64_t
*
feasign
=
batch_uint64_feasigns
[
i
].
data
();
uint64_t
*
feasign
=
batch_uint64_feasigns
_
[
i
].
data
();
int64_t
*
tensor_ptr
=
feed_vec_
[
i
]
->
mutable_data
<
int64_t
>
(
{
total_instance
,
1
},
this
->
place_
);
CopyToFeedTensor
(
tensor_ptr
,
feasign
,
total_instance
*
sizeof
(
int64_t
));
}
auto
&
slot_offset
=
offset
[
i
];
auto
&
slot_offset
=
offset
_
[
i
];
LoD
data_lod
{
slot_offset
};
feed_vec_
[
i
]
->
set_lod
(
data_lod
);
if
(
use_slots_is_dense_
[
i
])
{
...
...
paddle/fluid/framework/data_feed.h
浏览文件 @
9ff558a4
...
...
@@ -597,6 +597,10 @@ class MultiSlotInMemoryDataFeed : public InMemoryDataFeed<Record> {
virtual
void
PutToFeedVec
(
const
std
::
vector
<
Record
>&
ins_vec
);
virtual
void
GetMsgFromLogKey
(
const
std
::
string
&
log_key
,
uint64_t
*
search_id
,
uint32_t
*
cmatch
,
uint32_t
*
rank
);
std
::
vector
<
std
::
vector
<
float
>>
batch_float_feasigns_
;
std
::
vector
<
std
::
vector
<
uint64_t
>>
batch_uint64_feasigns_
;
std
::
vector
<
std
::
vector
<
size_t
>>
offset_
;
std
::
vector
<
bool
>
visit_
;
};
class
PaddleBoxDataFeed
:
public
MultiSlotInMemoryDataFeed
{
...
...
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