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
a6868c91
P
Paddle
项目概览
PaddlePaddle
/
Paddle
1 年多 前同步成功
通知
2302
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
未验证
提交
a6868c91
编写于
10月 12, 2021
作者:
A
Aurelius84
提交者:
GitHub
10月 12, 2021
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Fix stop_gradient in RunProgramOp (#36339) (#36353)
* Fix stop_gradient in RunProgramOp * fix reference
上级
10eebfa0
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
67 addition
and
7 deletion
+67
-7
paddle/fluid/operators/run_program_op.h
paddle/fluid/operators/run_program_op.h
+19
-7
python/paddle/fluid/tests/unittests/test_run_program_op.py
python/paddle/fluid/tests/unittests/test_run_program_op.py
+48
-0
未找到文件。
paddle/fluid/operators/run_program_op.h
浏览文件 @
a6868c91
...
@@ -142,10 +142,15 @@ static void ShareVarsIntoScope(const std::vector<Variable *> &vars,
...
@@ -142,10 +142,15 @@ static void ShareVarsIntoScope(const std::vector<Variable *> &vars,
static
void
ShareVarsFromScope
(
const
std
::
vector
<
Variable
*>
&
vars
,
static
void
ShareVarsFromScope
(
const
std
::
vector
<
Variable
*>
&
vars
,
const
std
::
vector
<
std
::
string
>
&
var_names
,
const
std
::
vector
<
std
::
string
>
&
var_names
,
const
BlockDesc
&
global_block
,
framework
::
Scope
*
scope
)
{
framework
::
Scope
*
scope
)
{
for
(
size_t
i
=
0
;
i
<
vars
.
size
();
++
i
)
{
for
(
size_t
i
=
0
;
i
<
vars
.
size
();
++
i
)
{
// NOTE: In case of setting out_tmp.stop_gradient = True in model code, all
// parameters before generating out_tmp have no @GRAD, it will raise error
// because we can't findthem in scope. So we skip sharing these vars or
// var@GRAD if they don't appear in global block.
if
(
var_names
[
i
]
==
framework
::
kEmptyVarName
||
if
(
var_names
[
i
]
==
framework
::
kEmptyVarName
||
var_names
[
i
]
==
"Fake_var"
)
{
var_names
[
i
]
==
"Fake_var"
||
!
global_block
.
HasVar
(
var_names
[
i
])
)
{
VLOG
(
2
)
<<
"find variable name is "
<<
var_names
[
i
]
<<
", skip it!"
;
VLOG
(
2
)
<<
"find variable name is "
<<
var_names
[
i
]
<<
", skip it!"
;
continue
;
continue
;
}
}
...
@@ -214,8 +219,10 @@ class RunProgramOpKernel : public framework::OpKernel<T> {
...
@@ -214,8 +219,10 @@ class RunProgramOpKernel : public framework::OpKernel<T> {
details
::
ShareVarsIntoScope
(
input_vars
,
input_var_names
,
&
scope
);
details
::
ShareVarsIntoScope
(
input_vars
,
input_var_names
,
&
scope
);
details
::
ShareVarsIntoScope
(
param_vars
,
param_names
,
&
scope
);
details
::
ShareVarsIntoScope
(
param_vars
,
param_names
,
&
scope
);
auto
*
global_block
=
ctx
.
Attr
<
BlockDesc
*>
(
"global_block"
);
if
(
end_op_index
>
start_op_index
)
{
if
(
end_op_index
>
start_op_index
)
{
auto
*
program
=
ctx
.
Attr
<
BlockDesc
*>
(
"global_block"
)
->
Program
();
auto
*
program
=
global_block
->
Program
();
auto
cache_info
=
framework
::
GetExecutorInfoFromCache
(
auto
cache_info
=
framework
::
GetExecutorInfoFromCache
(
*
program
,
ctx
.
GetPlace
(),
start_op_index
,
end_op_index
,
*
program
,
ctx
.
GetPlace
(),
start_op_index
,
end_op_index
,
/*is_grad=*/
false
,
program_id
,
&
scope
);
/*is_grad=*/
false
,
program_id
,
&
scope
);
...
@@ -240,8 +247,10 @@ class RunProgramOpKernel : public framework::OpKernel<T> {
...
@@ -240,8 +247,10 @@ class RunProgramOpKernel : public framework::OpKernel<T> {
parallel_executor
->
RunWithoutFetch
(
skip_eager_delete_vars
);
parallel_executor
->
RunWithoutFetch
(
skip_eager_delete_vars
);
}
}
// Step 4. Get Output
// Step 4. Get Output
details
::
ShareVarsFromScope
(
output_vars
,
output_var_names
,
&
scope
);
details
::
ShareVarsFromScope
(
output_vars
,
output_var_names
,
*
global_block
,
details
::
ShareVarsFromScope
(
dout_vars
,
dout_var_names
,
&
scope
);
&
scope
);
details
::
ShareVarsFromScope
(
dout_vars
,
dout_var_names
,
*
global_block
,
&
scope
);
// Debug info: scope info when run end
// Debug info: scope info when run end
VLOG
(
3
)
<<
framework
::
GenScopeTreeDebugInfo
(
out_scope_vec
->
front
());
VLOG
(
3
)
<<
framework
::
GenScopeTreeDebugInfo
(
out_scope_vec
->
front
());
...
@@ -307,10 +316,11 @@ class RunProgramGradOpKernel : public framework::OpKernel<T> {
...
@@ -307,10 +316,11 @@ class RunProgramGradOpKernel : public framework::OpKernel<T> {
"least one sub scope."
));
"least one sub scope."
));
auto
&
scope
=
*
(
global_inner_scope
->
kids
().
front
());
auto
&
scope
=
*
(
global_inner_scope
->
kids
().
front
());
auto
*
global_block
=
ctx
.
Attr
<
BlockDesc
*>
(
"global_block"
);
if
(
end_op_index
>
start_op_index
)
{
if
(
end_op_index
>
start_op_index
)
{
// Step 2. prepare executor and scope
// Step 2. prepare executor and scope
auto
*
program
=
ctx
.
Attr
<
BlockDesc
*>
(
"global_block"
)
->
Program
();
auto
*
program
=
global_block
->
Program
();
auto
cache_info
=
framework
::
GetExecutorInfoFromCache
(
auto
cache_info
=
framework
::
GetExecutorInfoFromCache
(
*
program
,
ctx
.
GetPlace
(),
start_op_index
,
end_op_index
,
*
program
,
ctx
.
GetPlace
(),
start_op_index
,
end_op_index
,
/*is_grad*/
true
,
program_id
,
&
scope
);
/*is_grad*/
true
,
program_id
,
&
scope
);
...
@@ -341,8 +351,10 @@ class RunProgramGradOpKernel : public framework::OpKernel<T> {
...
@@ -341,8 +351,10 @@ class RunProgramGradOpKernel : public framework::OpKernel<T> {
}
}
// Step 4. get outputs
// Step 4. get outputs
details
::
ShareVarsFromScope
(
input_grad_vars
,
input_grad_var_names
,
&
scope
);
details
::
ShareVarsFromScope
(
input_grad_vars
,
input_grad_var_names
,
details
::
ShareVarsFromScope
(
param_grad_vars
,
param_grad_names
,
&
scope
);
*
global_block
,
&
scope
);
details
::
ShareVarsFromScope
(
param_grad_vars
,
param_grad_names
,
*
global_block
,
&
scope
);
// Step5. drop current scope
// Step5. drop current scope
global_inner_scope
->
DeleteScope
(
&
scope
);
global_inner_scope
->
DeleteScope
(
&
scope
);
...
...
python/paddle/fluid/tests/unittests/test_run_program_op.py
浏览文件 @
a6868c91
...
@@ -343,5 +343,53 @@ class TestRunProgramOpWithEmbedding(RunProgramOpTest):
...
@@ -343,5 +343,53 @@ class TestRunProgramOpWithEmbedding(RunProgramOpTest):
return
fwd_op_num
return
fwd_op_num
class
Net
(
paddle
.
nn
.
Layer
):
def
__init__
(
self
):
super
(
Net
,
self
).
__init__
()
self
.
fc1
=
paddle
.
nn
.
Linear
(
10
,
10
)
self
.
fc2
=
paddle
.
nn
.
Linear
(
10
,
1
)
def
forward
(
self
,
x
):
out
=
self
.
fc1
(
x
)
out
.
stop_gradient
=
True
out
=
self
.
fc2
(
out
)
return
out
class
TestParametersWithStopGradient
(
unittest
.
TestCase
):
def
setUp
(
self
):
self
.
seed
=
2021
self
.
iter
=
5
def
train
(
self
,
to_static
):
# prepare env
paddle
.
seed
(
self
.
seed
)
net
=
Net
()
if
to_static
:
net
=
paddle
.
jit
.
to_static
(
net
)
sgd
=
paddle
.
optimizer
.
SGD
(
0.01
,
parameters
=
net
.
parameters
())
for
i
in
range
(
self
.
iter
):
x
=
paddle
.
rand
([
4
,
10
])
out
=
net
(
x
)
loss
=
paddle
.
mean
(
out
)
loss
.
backward
()
sgd
.
minimize
(
loss
)
net
.
clear_gradients
()
return
loss
def
test_stop_gradient
(
self
):
paddle
.
disable_static
()
dy_loss
=
self
.
train
(
to_static
=
False
)
st_loss
=
self
.
train
(
to_static
=
True
)
self
.
assertEqual
(
dy_loss
[
0
],
st_loss
[
0
])
paddle
.
enable_static
()
if
__name__
==
"__main__"
:
if
__name__
==
"__main__"
:
unittest
.
main
()
unittest
.
main
()
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