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
sven357
Paddle
提交
c4f8afa2
P
Paddle
项目概览
sven357
/
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 >>
提交
c4f8afa2
编写于
8月 08, 2018
作者:
Y
Yu Yang
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Fix bug when memopt optimize loss.grad and use ParallelExecutor
上级
7c8b69c7
变更
3
显示空白变更内容
内联
并排
Showing
3 changed file
with
15 addition
and
9 deletion
+15
-9
paddle/fluid/framework/details/multi_devices_graph_builder.cc
...le/fluid/framework/details/multi_devices_graph_builder.cc
+8
-6
paddle/fluid/framework/details/multi_devices_graph_builder.h
paddle/fluid/framework/details/multi_devices_graph_builder.h
+3
-1
paddle/fluid/framework/operator.cc
paddle/fluid/framework/operator.cc
+4
-2
未找到文件。
paddle/fluid/framework/details/multi_devices_graph_builder.cc
浏览文件 @
c4f8afa2
...
...
@@ -275,7 +275,8 @@ std::unique_ptr<ir::Graph> MultiDevSSAGraphBuilder::ApplyImpl(
if
(
strategy_
.
gradient_scale_
!=
BuildStrategy
::
GradientScaleStrategy
::
kCustomized
)
{
// TODO(paddle-dev): Why is there no input for this op_handle?
CreateScaleLossGradOp
(
&
result
);
auto
loss_grad_name
=
node
->
Op
()
->
OutputArgumentNames
()[
0
];
CreateScaleLossGradOp
(
&
result
,
loss_grad_name
);
}
// This assumes the backward generating code will ensure IsScaleLossOp
// is true only for the op that scale the final scalar loss.
...
...
@@ -535,7 +536,8 @@ int MultiDevSSAGraphBuilder::GetVarDeviceID(const ir::Graph &graph,
return
got
==
sharded_var_device
.
end
()
?
-
1
:
got
->
second
;
}
void
MultiDevSSAGraphBuilder
::
CreateScaleLossGradOp
(
ir
::
Graph
*
result
)
const
{
void
MultiDevSSAGraphBuilder
::
CreateScaleLossGradOp
(
ir
::
Graph
*
result
,
const
std
::
string
&
loss_grad_name
)
const
{
for
(
size_t
i
=
0
;
i
<
places_
.
size
();
++
i
)
{
// Insert ScaleCost OpHandle
#ifdef PADDLE_WITH_CUDA
...
...
@@ -558,9 +560,9 @@ void MultiDevSSAGraphBuilder::CreateScaleLossGradOp(ir::Graph *result) const {
// loss->pending_ops_.emplace_back(op_handle);
// op_handle->inputs_.emplace_back(loss);
CreateOpOutput
(
result
,
op_handle
,
result
->
CreateEmptyNode
(
GradVarName
(
loss_var_name_
)
,
ir
::
Node
::
Type
::
kVariable
),
CreateOpOutput
(
result
,
op_handle
,
result
->
CreateEmptyNode
(
loss_grad_name
,
ir
::
Node
::
Type
::
kVariable
),
places_
[
i
],
i
);
}
}
...
...
paddle/fluid/framework/details/multi_devices_graph_builder.h
浏览文件 @
c4f8afa2
...
...
@@ -75,7 +75,9 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
void
CreateComputationalOps
(
ir
::
Graph
*
result
,
ir
::
Node
*
node
,
size_t
num_places
)
const
;
void
CreateScaleLossGradOp
(
ir
::
Graph
*
result
)
const
;
void
CreateScaleLossGradOp
(
ir
::
Graph
*
result
,
const
std
::
string
&
loss_grad_name
)
const
;
VarHandle
*
CreateReduceOp
(
ir
::
Graph
*
result
,
const
std
::
string
&
og
,
int
dst_dev_id
)
const
;
void
CreateComputationalOp
(
ir
::
Graph
*
result
,
ir
::
Node
*
node
,
...
...
paddle/fluid/framework/operator.cc
浏览文件 @
c4f8afa2
...
...
@@ -778,6 +778,7 @@ proto::VarType::Type OperatorWithKernel::IndicateDataType(
const
ExecutionContext
&
ctx
)
const
{
auto
&
scope
=
ctx
.
scope
();
int
data_type
=
-
1
;
std
::
string
last_input_name
;
for
(
auto
&
input
:
this
->
inputs_
)
{
for
(
auto
&
ipt_name
:
input
.
second
)
{
auto
*
var
=
scope
.
FindVar
(
ipt_name
);
...
...
@@ -794,9 +795,10 @@ proto::VarType::Type OperatorWithKernel::IndicateDataType(
int
tmp
=
static_cast
<
int
>
(
ToDataType
(
t
->
type
()));
PADDLE_ENFORCE
(
tmp
==
data_type
||
data_type
==
-
1
,
"DataType of Paddle Op %s must be the same. Get %
d != %d"
,
Type
()
,
data_typ
e
,
tmp
);
"DataType of Paddle Op %s must be the same. Get %
s(%d) != %s(%d)"
,
Type
(),
last_input_name
,
data_type
,
ipt_nam
e
,
tmp
);
data_type
=
tmp
;
last_input_name
=
ipt_name
;
}
}
}
...
...
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