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
PaddleDetection
提交
2c8fe4e5
P
PaddleDetection
项目概览
PaddlePaddle
/
PaddleDetection
12 个月 前同步成功
通知
692
Star
11112
Fork
2696
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
184
列表
看板
标记
里程碑
合并请求
40
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
PaddleDetection
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
184
Issue
184
列表
看板
标记
里程碑
合并请求
40
合并请求
40
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
2c8fe4e5
编写于
4月 24, 2018
作者:
Y
Yu Yang
提交者:
GitHub
4月 24, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #10143 from typhoonzero/fix_multiGPU_dist_train
Fix multi gpu dist train
上级
e5f2cb80
f0341520
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
46 addition
and
4 deletion
+46
-4
paddle/fluid/framework/details/multi_devices_graph_builder.cc
...le/fluid/framework/details/multi_devices_graph_builder.cc
+42
-3
paddle/fluid/framework/details/multi_devices_graph_builder.h
paddle/fluid/framework/details/multi_devices_graph_builder.h
+4
-1
未找到文件。
paddle/fluid/framework/details/multi_devices_graph_builder.cc
浏览文件 @
2c8fe4e5
...
...
@@ -78,6 +78,33 @@ void MultiDevSSAGraphBuilder::CreateOpHandleIOs(SSAGraph *result,
}
}
bool
MultiDevSSAGraphBuilder
::
IsDistTrainOp
(
const
OpDesc
&
op
,
OpDesc
*
send_op
)
const
{
if
(
send_op
==
nullptr
)
{
return
false
;
}
auto
checker
=
[
&
](
const
std
::
vector
<
std
::
string
>
opvars
,
const
std
::
vector
<
std
::
string
>
sendvars
)
->
bool
{
bool
is_dist_train_op
=
false
;
for
(
auto
&
var
:
opvars
)
{
if
(
var
.
find
(
".block"
)
!=
std
::
string
::
npos
&&
std
::
find
(
sendvars
.
begin
(),
sendvars
.
end
(),
var
)
!=
sendvars
.
end
())
{
is_dist_train_op
=
true
;
break
;
}
}
return
is_dist_train_op
;
};
if
(
op
.
Type
()
==
"split"
)
{
return
checker
(
op
.
OutputArgumentNames
(),
send_op
->
InputArgumentNames
());
}
else
if
(
op
.
Type
()
==
"concat"
)
{
return
checker
(
op
.
InputArgumentNames
(),
send_op
->
OutputArgumentNames
());
}
return
false
;
}
std
::
unique_ptr
<
SSAGraph
>
MultiDevSSAGraphBuilder
::
Build
(
const
ProgramDesc
&
program
)
const
{
auto
graph
=
new
SSAGraph
();
...
...
@@ -89,19 +116,30 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
std
::
unordered_map
<
std
::
string
,
std
::
vector
<
std
::
unique_ptr
<
VarHandle
>>>>
(
places_
.
size
());
// Find "send" op first for split is in front of send.
OpDesc
*
send_op
=
nullptr
;
for
(
auto
*
op
:
program
.
Block
(
0
).
AllOps
())
{
if
(
op
->
Type
()
==
"send"
)
{
send_op
=
op
;
break
;
}
}
bool
is_forwarding
=
true
;
for
(
auto
*
op
:
program
.
Block
(
0
).
AllOps
())
{
if
(
op
->
Type
()
==
"send"
)
{
// append send op if program is distributed trainer main program.
// always use the first device
CreateSendOp
(
&
result
,
*
op
);
}
else
if
(
IsDistTrainOp
(
*
op
,
send_op
))
{
CreateComputationalOps
(
&
result
,
*
op
,
1
);
}
else
if
(
IsScaleLossOp
(
*
op
))
{
if
(
!
skip_scale_loss_
)
{
CreateScaleLossGradOp
(
&
result
);
}
is_forwarding
=
false
;
}
else
{
CreateComputationalOps
(
&
result
,
*
op
);
CreateComputationalOps
(
&
result
,
*
op
,
places_
.
size
()
);
if
(
!
is_forwarding
)
{
// Currently, we assume that once gradient is generated, it can be
// broadcast, and each gradient is only broadcast once. But there are no
...
...
@@ -199,8 +237,9 @@ void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(SSAGraph *result) const {
}
void
MultiDevSSAGraphBuilder
::
CreateComputationalOps
(
SSAGraph
*
result
,
const
OpDesc
&
op
)
const
{
for
(
size_t
scope_idx
=
0
;
scope_idx
<
places_
.
size
();
++
scope_idx
)
{
const
OpDesc
&
op
,
size_t
num_places
)
const
{
for
(
size_t
scope_idx
=
0
;
scope_idx
<
num_places
;
++
scope_idx
)
{
auto
p
=
places_
[
scope_idx
];
auto
s
=
local_scopes_
[
scope_idx
];
result
->
ops_
.
emplace_back
(
new
ComputationOpHandle
(
op
,
s
,
p
));
...
...
paddle/fluid/framework/details/multi_devices_graph_builder.h
浏览文件 @
2c8fe4e5
...
...
@@ -65,7 +65,10 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
void
CreateSendOp
(
SSAGraph
*
result
,
const
OpDesc
&
op
)
const
;
void
CreateComputationalOps
(
SSAGraph
*
result
,
const
OpDesc
&
op
)
const
;
bool
IsDistTrainOp
(
const
OpDesc
&
op
,
OpDesc
*
send_op
)
const
;
void
CreateComputationalOps
(
SSAGraph
*
result
,
const
OpDesc
&
op
,
size_t
num_places
)
const
;
void
CreateScaleLossGradOp
(
SSAGraph
*
result
)
const
;
...
...
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