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
机器未来
Paddle
提交
e5281b3c
P
Paddle
项目概览
机器未来
/
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
提交
e5281b3c
编写于
5月 14, 2018
作者:
Y
yuyang18
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Clean code & add execution strategy
上级
9923be5d
变更
9
显示空白变更内容
内联
并排
Showing
9 changed file
with
154 addition
and
104 deletion
+154
-104
paddle/fluid/framework/details/execution_strategy.h
paddle/fluid/framework/details/execution_strategy.h
+29
-0
paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
...le/fluid/framework/details/threaded_ssa_graph_executor.cc
+8
-9
paddle/fluid/framework/details/threaded_ssa_graph_executor.h
paddle/fluid/framework/details/threaded_ssa_graph_executor.h
+6
-5
paddle/fluid/framework/parallel_executor.cc
paddle/fluid/framework/parallel_executor.cc
+4
-5
paddle/fluid/framework/parallel_executor.h
paddle/fluid/framework/parallel_executor.h
+19
-17
paddle/fluid/pybind/pybind.cc
paddle/fluid/pybind/pybind.cc
+27
-16
python/paddle/fluid/__init__.py
python/paddle/fluid/__init__.py
+28
-26
python/paddle/fluid/parallel_executor.py
python/paddle/fluid/parallel_executor.py
+29
-22
python/paddle/fluid/tests/unittests/test_parallel_executor.py
...on/paddle/fluid/tests/unittests/test_parallel_executor.py
+4
-4
未找到文件。
paddle/fluid/framework/details/execution_strategy.h
0 → 100644
浏览文件 @
e5281b3c
// 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
namespace
paddle
{
namespace
framework
{
namespace
details
{
struct
ExecutionStrategy
{
size_t
num_threads_
{
0
};
bool
use_event_
{
true
};
bool
allow_op_delay_
{
false
};
};
}
// namespace details
}
// namespace framework
}
// namespace paddle
paddle/fluid/framework/details/threaded_ssa_graph_executor.cc
浏览文件 @
e5281b3c
...
@@ -18,18 +18,17 @@ namespace paddle {
...
@@ -18,18 +18,17 @@ namespace paddle {
namespace
framework
{
namespace
framework
{
namespace
details
{
namespace
details
{
ThreadedSSAGraphExecutor
::
ThreadedSSAGraphExecutor
(
ThreadedSSAGraphExecutor
::
ThreadedSSAGraphExecutor
(
size_t
num_threads
,
bool
use_event
,
const
ExecutionStrategy
&
strategy
,
const
std
::
vector
<
Scope
*>
&
local_scopes
,
const
std
::
vector
<
Scope
*>
&
local_scopes
,
const
std
::
vector
<
platform
::
Place
>
&
places
,
const
std
::
vector
<
platform
::
Place
>
&
places
,
std
::
unique_ptr
<
SSAGraph
>
&&
graph
,
bool
allow_op_delay
)
std
::
unique_ptr
<
SSAGraph
>
&&
graph
)
:
SSAGraphExecutor
(
std
::
move
(
graph
)),
:
SSAGraphExecutor
(
std
::
move
(
graph
)),
pool_
(
num_threads
>=
2
?
new
::
ThreadPool
(
num_threads
)
:
nullptr
),
pool_
(
strategy
.
num_threads_
>=
2
?
new
::
ThreadPool
(
strategy
.
num_threads_
)
:
nullptr
),
local_scopes_
(
local_scopes
),
local_scopes_
(
local_scopes
),
places_
(
places
),
places_
(
places
),
fetch_ctxs_
(
places
),
fetch_ctxs_
(
places
),
use_event_
(
use_event
),
running_ops_
(
0
),
running_ops_
(
0
),
allow_op_delay_
(
allow_op_dela
y
)
{}
strategy_
(
strateg
y
)
{}
FeedFetchList
ThreadedSSAGraphExecutor
::
Run
(
FeedFetchList
ThreadedSSAGraphExecutor
::
Run
(
const
std
::
vector
<
std
::
string
>
&
fetch_tensors
)
{
const
std
::
vector
<
std
::
string
>
&
fetch_tensors
)
{
...
@@ -86,7 +85,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
...
@@ -86,7 +85,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
//
//
// NOTE: DelayedOps have a lower priority. It will be scheduled after all
// NOTE: DelayedOps have a lower priority. It will be scheduled after all
// ready_ops have been performed.
// ready_ops have been performed.
if
(
ready_ops
.
empty
()
&&
allow_op_delay_
&&
running_ops_
==
0
)
{
if
(
ready_ops
.
empty
()
&&
strategy_
.
allow_op_delay_
&&
running_ops_
==
0
)
{
run_all_ops
(
delayed_ops
);
run_all_ops
(
delayed_ops
);
}
else
{
}
else
{
run_all_ops
(
ready_ops
);
run_all_ops
(
ready_ops
);
...
@@ -113,7 +112,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
...
@@ -113,7 +112,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
auto
&
deps
=
pending_ops
[
op
];
auto
&
deps
=
pending_ops
[
op
];
--
deps
;
--
deps
;
if
(
deps
==
0
)
{
if
(
deps
==
0
)
{
if
(
op
->
IsMultiDeviceTransfer
()
&&
allow_op_delay_
)
{
if
(
op
->
IsMultiDeviceTransfer
()
&&
strategy_
.
allow_op_delay_
)
{
delayed_ops
.
insert
(
op
);
delayed_ops
.
insert
(
op
);
}
else
{
}
else
{
ready_ops
.
insert
(
op
);
ready_ops
.
insert
(
op
);
...
@@ -191,7 +190,7 @@ void ThreadedSSAGraphExecutor::RunOp(
...
@@ -191,7 +190,7 @@ void ThreadedSSAGraphExecutor::RunOp(
auto
op_run
=
[
ready_var_q
,
op
,
this
]
{
auto
op_run
=
[
ready_var_q
,
op
,
this
]
{
try
{
try
{
VLOG
(
10
)
<<
op
<<
" "
<<
op
->
Name
()
<<
" : "
<<
op
->
DebugString
();
VLOG
(
10
)
<<
op
<<
" "
<<
op
->
Name
()
<<
" : "
<<
op
->
DebugString
();
op
->
Run
(
use_event_
);
op
->
Run
(
strategy_
.
use_event_
);
VLOG
(
10
)
<<
op
<<
" "
<<
op
->
Name
()
<<
" Done "
;
VLOG
(
10
)
<<
op
<<
" "
<<
op
->
Name
()
<<
" Done "
;
running_ops_
--
;
running_ops_
--
;
ready_var_q
->
Extend
(
op
->
Outputs
());
ready_var_q
->
Extend
(
op
->
Outputs
());
...
...
paddle/fluid/framework/details/threaded_ssa_graph_executor.h
浏览文件 @
e5281b3c
...
@@ -23,6 +23,7 @@
...
@@ -23,6 +23,7 @@
#include <functional>
#include <functional>
#include "ThreadPool.h" // ThreadPool in thrird party
#include "ThreadPool.h" // ThreadPool in thrird party
#include "paddle/fluid/framework/blocking_queue.h"
#include "paddle/fluid/framework/blocking_queue.h"
#include "paddle/fluid/framework/details/execution_strategy.h"
#include "paddle/fluid/framework/details/fetch_op_handle.h"
#include "paddle/fluid/framework/details/fetch_op_handle.h"
#include "paddle/fluid/framework/details/ssa_graph_executor.h"
#include "paddle/fluid/framework/details/ssa_graph_executor.h"
...
@@ -34,11 +35,10 @@ namespace details {
...
@@ -34,11 +35,10 @@ namespace details {
class
ThreadedSSAGraphExecutor
:
public
SSAGraphExecutor
{
class
ThreadedSSAGraphExecutor
:
public
SSAGraphExecutor
{
public:
public:
ThreadedSSAGraphExecutor
(
size_t
num_threads
,
bool
use_event
,
ThreadedSSAGraphExecutor
(
const
ExecutionStrategy
&
strategy
,
const
std
::
vector
<
Scope
*>
&
local_scopes
,
const
std
::
vector
<
Scope
*>
&
local_scopes
,
const
std
::
vector
<
platform
::
Place
>
&
places
,
const
std
::
vector
<
platform
::
Place
>
&
places
,
std
::
unique_ptr
<
SSAGraph
>
&&
graph
,
std
::
unique_ptr
<
SSAGraph
>
&&
graph
);
bool
allow_op_delay
);
// Run a SSAGraph by a thread pool
// Run a SSAGraph by a thread pool
// Use topological sort algorithm
// Use topological sort algorithm
...
@@ -55,10 +55,8 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
...
@@ -55,10 +55,8 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
std
::
vector
<
Scope
*>
local_scopes_
;
std
::
vector
<
Scope
*>
local_scopes_
;
std
::
vector
<
platform
::
Place
>
places_
;
std
::
vector
<
platform
::
Place
>
places_
;
platform
::
DeviceContextPool
fetch_ctxs_
;
platform
::
DeviceContextPool
fetch_ctxs_
;
const
bool
use_event_
;
std
::
unique_ptr
<
platform
::
EnforceNotMet
>
exception_
;
std
::
unique_ptr
<
platform
::
EnforceNotMet
>
exception_
;
std
::
atomic
<
int
>
running_ops_
;
std
::
atomic
<
int
>
running_ops_
;
bool
allow_op_delay_
;
void
InsertPendingOp
(
std
::
unordered_map
<
OpHandleBase
*
,
size_t
>
*
pending_ops
,
void
InsertPendingOp
(
std
::
unordered_map
<
OpHandleBase
*
,
size_t
>
*
pending_ops
,
OpHandleBase
*
op_instance
)
const
;
OpHandleBase
*
op_instance
)
const
;
...
@@ -74,6 +72,9 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
...
@@ -74,6 +72,9 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
std
::
unordered_map
<
OpHandleBase
*
,
size_t
>
*
pending_ops
,
std
::
unordered_map
<
OpHandleBase
*
,
size_t
>
*
pending_ops
,
std
::
unordered_set
<
VarHandleBase
*>
*
pending_vars
,
std
::
unordered_set
<
VarHandleBase
*>
*
pending_vars
,
BlockingQueue
<
VarHandleBase
*>
*
ready_vars
,
FeedFetchList
*
fetch_data
);
BlockingQueue
<
VarHandleBase
*>
*
ready_vars
,
FeedFetchList
*
fetch_data
);
private:
ExecutionStrategy
strategy_
;
};
};
}
// namespace details
}
// namespace details
...
...
paddle/fluid/framework/parallel_executor.cc
浏览文件 @
e5281b3c
...
@@ -52,13 +52,13 @@ std::vector<Scope *> &ParallelExecutor::GetLocalScopes() {
...
@@ -52,13 +52,13 @@ std::vector<Scope *> &ParallelExecutor::GetLocalScopes() {
}
}
ParallelExecutor
::
ParallelExecutor
(
ParallelExecutor
::
ParallelExecutor
(
size_t
num_threads
,
bool
use_event
,
const
std
::
vector
<
platform
::
Place
>
&
places
,
const
std
::
vector
<
platform
::
Place
>
&
places
,
const
std
::
unordered_set
<
std
::
string
>
&
params
,
const
std
::
unordered_set
<
std
::
string
>
&
params
,
const
std
::
unordered_set
<
std
::
string
>
&
bcast_vars
,
const
std
::
unordered_set
<
std
::
string
>
&
bcast_vars
,
const
ProgramDesc
&
main_program
,
const
std
::
string
&
loss_var_name
,
const
ProgramDesc
&
main_program
,
const
std
::
string
&
loss_var_name
,
Scope
*
scope
,
const
std
::
vector
<
Scope
*>
&
local_scopes
,
bool
allow_op_delay
,
Scope
*
scope
,
const
std
::
vector
<
Scope
*>
&
local_scopes
,
bool
use_default_grad_scale
,
bool
balance_parameter_opt_between_cards
)
bool
use_default_grad_scale
,
bool
balance_parameter_opt_between_cards
,
const
ExecutionStrategy
&
exec_strategy
)
:
member_
(
new
ParallelExecutorPrivate
(
places
))
{
:
member_
(
new
ParallelExecutorPrivate
(
places
))
{
member_
->
global_scope_
=
scope
;
member_
->
global_scope_
=
scope
;
...
@@ -103,8 +103,7 @@ ParallelExecutor::ParallelExecutor(
...
@@ -103,8 +103,7 @@ ParallelExecutor::ParallelExecutor(
auto
graph
=
builder
.
Build
(
main_program
);
auto
graph
=
builder
.
Build
(
main_program
);
member_
->
executor_
.
reset
(
new
details
::
ThreadedSSAGraphExecutor
(
member_
->
executor_
.
reset
(
new
details
::
ThreadedSSAGraphExecutor
(
num_threads
,
use_event
,
member_
->
local_scopes_
,
places
,
std
::
move
(
graph
),
exec_strategy
,
member_
->
local_scopes_
,
places
,
std
::
move
(
graph
)));
allow_op_delay
));
// Step 3. Create vars in each scope;
// Step 3. Create vars in each scope;
for
(
auto
*
var
:
main_program
.
Block
(
0
).
AllVars
())
{
for
(
auto
*
var
:
main_program
.
Block
(
0
).
AllVars
())
{
...
...
paddle/fluid/framework/parallel_executor.h
浏览文件 @
e5281b3c
...
@@ -17,53 +17,55 @@ limitations under the License. */
...
@@ -17,53 +17,55 @@ limitations under the License. */
#include <string>
#include <string>
#include <unordered_set>
#include <unordered_set>
#include <vector>
#include <vector>
#include "paddle/fluid/framework/details/execution_strategy.h"
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/executor.h"
#include "paddle/fluid/framework/op_info.h"
#include "paddle/fluid/framework/op_info.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/program_desc.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/scope.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/framework/tensor.h"
#include "paddle/fluid/platform/device_context.h"
#include "paddle/fluid/platform/device_context.h"
namespace
paddle
{
namespace
paddle
{
namespace
framework
{
namespace
framework
{
class
ParallelExecutorPrivate
;
class
ParallelExecutorPrivate
;
using
details
::
ExecutionStrategy
;
class
ParallelExecutor
{
class
ParallelExecutor
{
DISABLE_COPY_AND_ASSIGN
(
ParallelExecutor
);
DISABLE_COPY_AND_ASSIGN
(
ParallelExecutor
);
public:
public:
explicit
ParallelExecutor
(
size_t
num_threads
,
bool
use_event
,
explicit
ParallelExecutor
(
const
std
::
vector
<
platform
::
Place
>
&
places
,
const
std
::
vector
<
platform
::
Place
>&
place
s
,
const
std
::
unordered_set
<
std
::
string
>
&
param
s
,
const
std
::
unordered_set
<
std
::
string
>
&
param
s
,
const
std
::
unordered_set
<
std
::
string
>
&
bcast_var
s
,
const
std
::
unordered_set
<
std
::
string
>&
bcast_vars
,
const
ProgramDesc
&
main_program
,
const
ProgramDesc
&
main_program
,
const
std
::
string
&
loss_var_name
,
Scope
*
scope
,
const
std
::
string
&
loss_var_name
,
Scope
*
scope
,
const
std
::
vector
<
Scope
*>
&
local_scopes
,
const
std
::
vector
<
Scope
*>&
local_scopes
,
bool
use_default_grad_scale
,
bool
allow_op_delay
,
bool
use_default_grad_scale
,
bool
balance_parameter_opt_between_cards
,
bool
balance_parameter_opt_between_cards
);
const
ExecutionStrategy
&
exec_strategy
);
~
ParallelExecutor
();
~
ParallelExecutor
();
std
::
vector
<
Scope
*>&
GetLocalScopes
();
std
::
vector
<
Scope
*>
&
GetLocalScopes
();
/**
/**
* Feed tensors to local scopes. The size of tensors should be equal to the
* Feed tensors to local scopes. The size of tensors should be equal to the
* size of local scopes.
* size of local scopes.
*/
*/
void
FeedTensorsIntoLocalScopes
(
void
FeedTensorsIntoLocalScopes
(
const
std
::
vector
<
std
::
unordered_map
<
std
::
string
,
LoDTensor
>>
&
tensors
);
const
std
::
vector
<
std
::
unordered_map
<
std
::
string
,
LoDTensor
>>
&
tensors
);
void
FeedAndSplitTensorIntoLocalScopes
(
void
FeedAndSplitTensorIntoLocalScopes
(
const
std
::
unordered_map
<
std
::
string
,
LoDTensor
>
&
tensors
);
const
std
::
unordered_map
<
std
::
string
,
LoDTensor
>
&
tensors
);
void
Run
(
const
std
::
vector
<
std
::
string
>
&
fetch_tensors
,
void
Run
(
const
std
::
vector
<
std
::
string
>
&
fetch_tensors
,
const
std
::
string
&
fetched_var_name
);
const
std
::
string
&
fetched_var_name
);
void
BCastParamsToGPUs
(
const
std
::
unordered_set
<
std
::
string
>
&
vars
)
const
;
void
BCastParamsToGPUs
(
const
std
::
unordered_set
<
std
::
string
>
&
vars
)
const
;
private:
private:
ParallelExecutorPrivate
*
member_
;
ParallelExecutorPrivate
*
member_
;
};
};
}
// namespace framework
}
// namespace framework
...
...
paddle/fluid/pybind/pybind.cc
浏览文件 @
e5281b3c
...
@@ -494,22 +494,33 @@ All parameter, weight, gradient are variables in Paddle.
...
@@ -494,22 +494,33 @@ All parameter, weight, gradient are variables in Paddle.
m
.
def
(
"disable_profiler"
,
platform
::
DisableProfiler
);
m
.
def
(
"disable_profiler"
,
platform
::
DisableProfiler
);
m
.
def
(
"reset_profiler"
,
platform
::
ResetProfiler
);
m
.
def
(
"reset_profiler"
,
platform
::
ResetProfiler
);
py
::
class_
<
ParallelExecutor
>
(
m
,
"ParallelExecutor"
)
py
::
class_
<
ParallelExecutor
>
pe
(
m
,
"ParallelExecutor"
);
.
def
(
"__init__"
,
py
::
class_
<
ExecutionStrategy
>
(
pe
,
"ExecutionStrategy"
)
[](
ParallelExecutor
&
self
,
size_t
num_threads
,
bool
use_event
,
.
def
(
py
::
init
())
const
std
::
vector
<
platform
::
Place
>
&
places
,
.
def_property
(
const
std
::
unordered_set
<
std
::
string
>
&
params
,
"num_threads"
,
const
std
::
unordered_set
<
std
::
string
>
&
bcast_vars
,
[](
const
ExecutionStrategy
&
self
)
{
return
self
.
num_threads_
;
},
const
ProgramDesc
&
main_program
,
const
std
::
string
&
loss_var_name
,
[](
ExecutionStrategy
&
self
,
size_t
num_threads
)
{
Scope
*
scope
,
std
::
vector
<
Scope
*>
&
local_scopes
,
self
.
num_threads_
=
num_threads
;
bool
allow_op_delay
,
bool
use_default_grad_scale
,
})
bool
balance_parameter_opt_between_cards
)
{
.
def_property
(
new
(
&
self
)
ParallelExecutor
(
"use_event"
,
num_threads
,
use_event
,
places
,
params
,
bcast_vars
,
[](
const
ExecutionStrategy
&
self
)
{
return
self
.
use_event_
;
},
main_program
,
loss_var_name
,
scope
,
local_scopes
,
[](
ExecutionStrategy
&
self
,
bool
use_event
)
{
allow_op_delay
,
use_default_grad_scale
,
self
.
use_event_
=
use_event
;
balance_parameter_opt_between_cards
);
})
})
.
def_property
(
"allow_op_delay"
,
[](
const
ExecutionStrategy
&
self
)
{
return
self
.
allow_op_delay_
;
},
[](
ExecutionStrategy
&
self
,
bool
allow_op_delay
)
{
self
.
allow_op_delay_
=
allow_op_delay
;
});
pe
.
def
(
py
::
init
<
const
std
::
vector
<
platform
::
Place
>
&
,
const
std
::
unordered_set
<
std
::
string
>
&
,
const
std
::
unordered_set
<
std
::
string
>
&
,
const
ProgramDesc
&
,
const
std
::
string
&
,
Scope
*
,
std
::
vector
<
Scope
*>
&
,
bool
,
bool
,
const
ExecutionStrategy
&>
())
.
def
(
"bcast_params"
,
&
ParallelExecutor
::
BCastParamsToGPUs
)
.
def
(
"bcast_params"
,
&
ParallelExecutor
::
BCastParamsToGPUs
)
// NOTE: even we return a vec<Scope*>* to Python use reference policy.
// NOTE: even we return a vec<Scope*>* to Python use reference policy.
// We still cannot get local_scope from this vector, since the element
// We still cannot get local_scope from this vector, since the element
...
...
python/paddle/fluid/__init__.py
浏览文件 @
e5281b3c
...
@@ -44,18 +44,19 @@ import transpiler
...
@@ -44,18 +44,19 @@ import transpiler
from
param_attr
import
ParamAttr
,
WeightNormParamAttr
from
param_attr
import
ParamAttr
,
WeightNormParamAttr
from
data_feeder
import
DataFeeder
from
data_feeder
import
DataFeeder
from
core
import
LoDTensor
,
CPUPlace
,
CUDAPlace
,
CUDAPinnedPlace
from
core
import
LoDTensor
,
CPUPlace
,
CUDAPlace
,
CUDAPinnedPlace
from
transpiler
import
DistributeTranspiler
,
SimpleDistributeTranspiler
,
InferenceTranspiler
,
memory_optimize
,
release_memory
from
transpiler
import
DistributeTranspiler
,
SimpleDistributeTranspiler
,
\
InferenceTranspiler
,
memory_optimize
,
release_memory
from
concurrency
import
(
Go
,
make_channel
,
channel_send
,
channel_recv
,
from
concurrency
import
(
Go
,
make_channel
,
channel_send
,
channel_recv
,
channel_close
,
Select
)
channel_close
,
Select
)
import
clip
import
clip
import
profiler
import
profiler
import
unique_name
import
unique_name
import
recordio_writer
import
recordio_writer
from
parallel_executor
import
ParallelExecutor
from
parallel_executor
import
ParallelExecutor
,
ExecutionStrategy
Tensor
=
LoDTensor
Tensor
=
LoDTensor
__all__
=
framework
.
__all__
+
executor
.
__all__
+
concurrency
.
__all__
+
\
__all__
=
framework
.
__all__
+
executor
.
__all__
+
concurrency
.
__all__
+
\
trainer
.
__all__
+
inferencer
.
__all__
+
transpiler
.
__all__
+
[
trainer
.
__all__
+
inferencer
.
__all__
+
transpiler
.
__all__
+
[
'io'
,
'io'
,
'initializer'
,
'initializer'
,
...
@@ -79,7 +80,8 @@ __all__ = framework.__all__ + executor.__all__ + concurrency.__all__ +\
...
@@ -79,7 +80,8 @@ __all__ = framework.__all__ + executor.__all__ + concurrency.__all__ +\
'unique_name'
,
'unique_name'
,
'recordio_writer'
,
'recordio_writer'
,
'ParallelExecutor'
,
'ParallelExecutor'
,
]
'ExecutionStrategy'
,
]
def
__bootstrap__
():
def
__bootstrap__
():
...
...
python/paddle/fluid/parallel_executor.py
浏览文件 @
e5281b3c
...
@@ -19,7 +19,9 @@ import executor
...
@@ -19,7 +19,9 @@ import executor
import
warnings
import
warnings
import
sys
import
sys
__all__
=
[
'ParallelExecutor'
]
__all__
=
[
'ParallelExecutor'
,
'ExecutionStrategy'
]
ExecutionStrategy
=
core
.
ParallelExecutor
.
ExecutionStrategy
class
ParallelExecutor
(
object
):
class
ParallelExecutor
(
object
):
...
@@ -27,11 +29,11 @@ class ParallelExecutor(object):
...
@@ -27,11 +29,11 @@ class ParallelExecutor(object):
use_cuda
,
use_cuda
,
loss_name
=
None
,
loss_name
=
None
,
main_program
=
None
,
main_program
=
None
,
num_threads
=
None
,
allow_op_delay
=
False
,
share_vars_from
=
None
,
share_vars_from
=
None
,
use_default_grad_scale
=
True
,
use_default_grad_scale
=
True
,
balance_parameter_opt_between_cards
=
False
):
balance_parameter_opt_between_cards
=
False
,
exec_strategy
=
None
,
**
kwargs
):
"""
"""
ParallelExecutor can run program in parallel.
ParallelExecutor can run program in parallel.
...
@@ -40,11 +42,6 @@ class ParallelExecutor(object):
...
@@ -40,11 +42,6 @@ class ParallelExecutor(object):
loss_name(str, default None): The loss name must set in training.
loss_name(str, default None): The loss name must set in training.
main_program(Program, default None): The program that need to run,
main_program(Program, default None): The program that need to run,
if not provided, then default_main_program will be used.
if not provided, then default_main_program will be used.
num_threads(int, default None): How many threads are used for
training.
allow_op_delay(bool, default False): Whether to delay and buffer
some operators together for scheduling or not, which may
improve performance in some cases, default False.
share_vars_from(ParallelExecutor, default None): If provied,
share_vars_from(ParallelExecutor, default None): If provied,
it will share variables from the specified ParallelExecutor.
it will share variables from the specified ParallelExecutor.
use_default_grad_scale(bool, default True): If set True, a default
use_default_grad_scale(bool, default True): If set True, a default
...
@@ -76,6 +73,16 @@ class ParallelExecutor(object):
...
@@ -76,6 +73,16 @@ class ParallelExecutor(object):
train_loss, = train_exe.run([loss.name], feed=feed_dict)
train_loss, = train_exe.run([loss.name], feed=feed_dict)
test_loss, = test_exe.run([loss.name], feed=feed_dict)
test_loss, = test_exe.run([loss.name], feed=feed_dict)
"""
"""
if
len
(
kwargs
)
!=
0
:
err_msg
=
""
for
key
in
kwargs
:
if
key
in
dir
(
ExecutionStrategy
):
err_msg
+=
\
"Setting {0} by constructor is deprecated. Use "
\
"strategy=ExecutionStrategy(); strategy.{0}=xxx; "
\
"pe=ParallelExecutor(exec_strategy=strategy) "
\
"instead.
\n
"
raise
ValueError
(
err_msg
)
self
.
_places
=
[]
self
.
_places
=
[]
self
.
_act_places
=
[]
self
.
_act_places
=
[]
...
@@ -93,13 +100,20 @@ class ParallelExecutor(object):
...
@@ -93,13 +100,20 @@ class ParallelExecutor(object):
self
.
_places
.
append
(
p
)
self
.
_places
.
append
(
p
)
assert
self
.
_places
,
"no place for execution"
assert
self
.
_places
,
"no place for execution"
if
num_threads
is
None
:
if
exec_strategy
is
None
:
exec_strategy
=
ExecutionStrategy
()
if
use_cuda
:
exec_strategy
.
use_event
=
True
else
:
exec_strategy
.
use_event
=
False
if
exec_strategy
.
num_threads
==
0
:
if
use_cuda
:
if
use_cuda
:
# Experiments on se-resnext shows that too many threads hurt
# Experiments on se-resnext shows that too many threads hurt
# performance. Worth tunning for other models in the future.
# performance. Worth tunning for other models in the future.
num_threads
=
len
(
self
.
_places
)
*
2
exec_strategy
.
num_threads
=
len
(
self
.
_places
)
*
2
else
:
else
:
num_threads
=
min
(
exec_strategy
.
num_threads
=
min
(
len
(
self
.
_places
)
*
2
,
multiprocessing
.
cpu_count
())
len
(
self
.
_places
)
*
2
,
multiprocessing
.
cpu_count
())
main
=
main_program
main
=
main_program
...
@@ -120,21 +134,14 @@ class ParallelExecutor(object):
...
@@ -120,21 +134,14 @@ class ParallelExecutor(object):
]
]
self
.
executor
=
core
.
ParallelExecutor
(
self
.
executor
=
core
.
ParallelExecutor
(
num_threads
,
True
if
use_cuda
else
False
,
# use_event
self
.
_places
,
self
.
_places
,
set
([
set
([
p
.
name
for
p
in
main
.
global_block
().
iter_parameters
()
p
.
name
for
p
in
main
.
global_block
().
iter_parameters
()
if
not
p
.
stop_gradient
if
not
p
.
stop_gradient
]),
]),
set
(
self
.
persistable_vars
),
set
(
self
.
persistable_vars
),
main
.
desc
,
loss_name
main
.
desc
,
if
loss_name
else
''
,
scope
,
local_scopes
,
use_default_grad_scale
,
loss_name
if
loss_name
else
''
,
balance_parameter_opt_between_cards
,
exec_strategy
)
scope
,
local_scopes
,
allow_op_delay
,
use_default_grad_scale
,
balance_parameter_opt_between_cards
)
self
.
scope
=
scope
self
.
scope
=
scope
...
...
python/paddle/fluid/tests/unittests/test_parallel_executor.py
浏览文件 @
e5281b3c
...
@@ -232,14 +232,14 @@ class TestParallelExecutorBase(unittest.TestCase):
...
@@ -232,14 +232,14 @@ class TestParallelExecutorBase(unittest.TestCase):
place
=
fluid
.
CUDAPlace
(
0
)
place
=
fluid
.
CUDAPlace
(
0
)
startup_exe
=
fluid
.
Executor
(
place
)
startup_exe
=
fluid
.
Executor
(
place
)
startup_exe
.
run
(
startup
)
startup_exe
.
run
(
startup
)
exec_strategy
=
fluid
.
ExecutionStrategy
()
exec_strategy
.
allow_op_delay
=
allow_op_delay
if
use_parallel_executor
:
if
use_parallel_executor
:
exe
=
fluid
.
ParallelExecutor
(
exe
=
fluid
.
ParallelExecutor
(
True
,
True
,
loss_name
=
loss
.
name
,
loss_name
=
loss
.
name
,
allow_op_delay
=
allow_op_delay
,
balance_parameter_opt_between_cards
=
balance_parameter_opt_between_cards
,
balance_parameter_opt_between_cards
=
balance_parameter_opt_between_cards
exec_strategy
=
exec_strategy
)
)
else
:
else
:
exe
=
fluid
.
Executor
(
place
=
place
)
exe
=
fluid
.
Executor
(
place
=
place
)
...
...
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