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
2e9a836c
P
Paddle
项目概览
PaddlePaddle
/
Paddle
12 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 >>
提交
2e9a836c
编写于
3月 06, 2019
作者:
X
xjqbest
提交者:
dongdaxiang
3月 29, 2019
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
add DataSet and InMemoryDataFeed, support load data into memory and shuffle data
上级
8de4d31a
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
131 addition
and
2 deletion
+131
-2
paddle/fluid/framework/CMakeLists.txt
paddle/fluid/framework/CMakeLists.txt
+24
-0
paddle/fluid/framework/data_feed.cc
paddle/fluid/framework/data_feed.cc
+103
-0
paddle/fluid/pybind/data_set_py.cc
paddle/fluid/pybind/data_set_py.cc
+4
-2
未找到文件。
paddle/fluid/framework/CMakeLists.txt
浏览文件 @
2e9a836c
...
...
@@ -190,6 +190,30 @@ cc_library(parallel_executor SRCS parallel_executor.cc DEPS
graph build_strategy
fast_threaded_ssa_graph_executor variable_helper
)
<<<<<<< HEAD
=======
if
(
WITH_PSLIB
)
cc_library
(
async_executor SRCS async_executor.cc data_feed.cc data_feed_factory.cc
executor_thread_worker.cc multi_trainer.cc dist_multi_trainer.cc
trainer_factory.cc trainer.cc device_worker.cc hogwild_worker.cc
downpour_worker.cc pull_dense_worker.cc device_worker_factory.cc
data_set.cc
DEPS op_registry device_context scope framework_proto
trainer_desc_proto glog lod_rank_table fleet_wrapper lodtensor_printer
feed_fetch_method graph_to_program_pass async_executor_proto
variable_helper pslib_brpc pslib timer
)
else
()
cc_library
(
async_executor SRCS async_executor.cc data_feed.cc data_feed_factory.cc
executor_thread_worker.cc multi_trainer.cc dist_multi_trainer.cc
trainer_factory.cc trainer.cc device_worker.cc hogwild_worker.cc
downpour_worker.cc pull_dense_worker.cc device_worker_factory.cc
data_set.cc
DEPS op_registry device_context scope framework_proto
trainer_desc_proto glog lod_rank_table fleet_wrapper lodtensor_printer
feed_fetch_method graph_to_program_pass async_executor_proto
variable_helper timer
)
endif
(
WITH_PSLIB
)
>>>>>>> 870b88bbd7... add DataSet and InMemoryDataFeed, support load data into memory and shuffle data
cc_library
(
async_executor SRCS async_executor.cc data_feed.cc data_feed_factory.cc
executor_thread_worker.cc multi_trainer.cc dist_multi_trainer.cc
...
...
paddle/fluid/framework/data_feed.cc
浏览文件 @
2e9a836c
...
...
@@ -242,6 +242,109 @@ void InMemoryDataFeed<T>::GlobalShuffle(int trainer_num) {
template
class
InMemoryDataFeed
<
std
::
vector
<
MultiSlotType
>
>
;
template
<
typename
T
>
InMemoryDataFeed
<
T
>::
InMemoryDataFeed
()
{
cur_channel_
=
0
;
shuffled_ins_
=
nullptr
;
shuffled_ins_out_
=
nullptr
;
}
template
<
typename
T
>
bool
InMemoryDataFeed
<
T
>::
Start
()
{
DataFeed
::
CheckSetFileList
();
if
(
memory_data_
.
size
()
!=
0
)
{
CHECK_EQ
(
cur_channel_
,
0
);
shuffled_ins_
->
Extend
(
std
::
move
(
memory_data_
));
std
::
vector
<
T
>
().
swap
(
memory_data_
);
}
DataFeed
::
finish_start_
=
true
;
return
true
;
}
template
<
typename
T
>
int
InMemoryDataFeed
<
T
>::
Next
()
{
DataFeed
::
CheckStart
();
std
::
shared_ptr
<
paddle
::
framework
::
BlockingQueue
<
T
>>
in_channel
=
nullptr
;
std
::
shared_ptr
<
paddle
::
framework
::
BlockingQueue
<
T
>>
out_channel
=
nullptr
;
if
(
cur_channel_
==
0
)
{
in_channel
=
shuffled_ins_
;
out_channel
=
shuffled_ins_out_
;
}
else
{
in_channel
=
shuffled_ins_out_
;
out_channel
=
shuffled_ins_
;
}
CHECK
(
in_channel
!=
nullptr
);
CHECK
(
out_channel
!=
nullptr
);
int
index
=
0
;
T
instance
;
T
ins_vec
;
while
(
index
<
DataFeed
::
default_batch_size_
)
{
if
(
in_channel
->
Size
()
==
0
)
{
break
;
}
in_channel
->
Pop
(
instance
);
AddInstanceToInsVec
(
&
ins_vec
,
instance
,
index
++
);
out_channel
->
Push
(
std
::
move
(
instance
));
}
DataFeed
::
batch_size_
=
index
;
if
(
DataFeed
::
batch_size_
!=
0
)
{
PutToFeedVec
(
ins_vec
);
}
else
{
cur_channel_
=
1
-
cur_channel_
;
}
return
DataFeed
::
batch_size_
;
}
template
<
typename
T
>
void
InMemoryDataFeed
<
T
>::
PutInsToChannel
(
const
std
::
string
&
ins_str
)
{
T
ins
;
DeserializeIns
(
ins
,
ins_str
);
shuffled_ins_
->
Push
(
std
::
move
(
ins
));
}
template
<
typename
T
>
void
InMemoryDataFeed
<
T
>::
LoadIntoMemory
()
{
std
::
vector
<
T
>
local_vec
;
std
::
string
filename
;
while
(
DataFeed
::
PickOneFile
(
&
filename
))
{
int
err_no
=
0
;
PrivateQueueDataFeed
<
T
>::
fp_
=
fs_open_read
(
filename
,
&
err_no
,
PrivateQueueDataFeed
<
T
>::
pipe_command_
);
__fsetlocking
(
&*
PrivateQueueDataFeed
<
T
>::
fp_
,
FSETLOCKING_BYCALLER
);
T
instance
;
while
(
ParseOneInstanceFromPipe
(
&
instance
))
{
local_vec
.
push_back
(
instance
);
}
memory_data_
.
insert
(
memory_data_
.
end
(),
local_vec
.
begin
(),
local_vec
.
end
());
std
::
vector
<
T
>
().
swap
(
local_vec
);
}
}
template
<
typename
T
>
void
InMemoryDataFeed
<
T
>::
LocalShuffle
()
{
std
::
random_shuffle
(
memory_data_
.
begin
(),
memory_data_
.
end
());
}
// todo global shuffle
/*
template <typename T>
void InMemoryDataFeed<T>::GlobalShuffle(int trainer_num) {
std::random_shuffle(memory_data_.begin(), memory_data_.end());
for (int64_t i = 0; i < memory_data_.size(); ++i) {
// todo get ins id
//std::string ins_id = memory_data_[i].ins_id;
// todo hash
int64_t hash_id = paddle::ps::local_random_engine()();
//int64_t hash_id = hash(ins_id);
int64_t node_id = hash_id % trainer_num_;
std::string str;
SerializeIns(memory_data_[i], str);
auto fleet_ptr = FleetWrapper::GetInstance();
auto ret = fleet_ptr->send_client2client_msg(0, node_id, str);
}
}
*/
void
MultiSlotDataFeed
::
Init
(
const
paddle
::
framework
::
DataFeedDesc
&
data_feed_desc
)
{
finish_init_
=
false
;
...
...
paddle/fluid/pybind/data_set_py.cc
浏览文件 @
2e9a836c
...
...
@@ -12,6 +12,8 @@ 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 <fcntl.h>
// To avoid conflicting definition in gcc-4.8.2 headers and pyconfig.h (2.7.3)
#ifdef _POSIX_C_SOURCE
#undef _POSIX_C_SOURCE
#endif
...
...
@@ -41,7 +43,7 @@ namespace paddle {
namespace
pybind
{
void
BindDataset
(
py
::
module
*
m
)
{
py
::
class_
<
framework
::
Data
s
et
>
(
*
m
,
"Dataset"
)
py
::
class_
<
framework
::
Data
S
et
>
(
*
m
,
"Dataset"
)
.
def
(
py
::
init
([]()
{
return
std
::
unique_ptr
<
framework
::
Dataset
>
(
new
framework
::
Dataset
());
}))
...
...
@@ -51,7 +53,7 @@ void BindDataset(py::module* m) {
.
def
(
"set_data_feed_desc"
,
&
framework
::
Dataset
::
SetDataFeedDesc
)
.
def
(
"load_into_memory"
,
&
framework
::
Dataset
::
LoadIntoMemory
)
.
def
(
"local_shuffle"
,
&
framework
::
Dataset
::
LocalShuffle
)
.
def
(
"global_shuffle"
,
&
framework
::
Dataset
::
G
lobalShuffle
);
.
def
(
"global_shuffle"
,
&
framework
::
Dataset
::
G
LobalShuffle
)
}
}
// end namespace pybind
...
...
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