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
866fcb0c
P
PaddleDetection
项目概览
PaddlePaddle
/
PaddleDetection
大约 1 年 前同步成功
通知
695
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
未验证
提交
866fcb0c
编写于
7月 17, 2018
作者:
W
Wu Yi
提交者:
GitHub
7月 17, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #12171 from typhoonzero/fix_pserver_with_condition_block
fix pserver with condition block
上级
d24fd2c6
32d81909
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
56 addition
and
9 deletion
+56
-9
paddle/fluid/operators/listen_and_serv_op.cc
paddle/fluid/operators/listen_and_serv_op.cc
+9
-5
python/paddle/fluid/tests/unittests/test_dist_transpiler.py
python/paddle/fluid/tests/unittests/test_dist_transpiler.py
+44
-4
python/paddle/fluid/transpiler/distribute_transpiler.py
python/paddle/fluid/transpiler/distribute_transpiler.py
+3
-0
未找到文件。
paddle/fluid/operators/listen_and_serv_op.cc
浏览文件 @
866fcb0c
...
...
@@ -61,6 +61,8 @@ static void ParallelExecuteBlocks(
framework
::
Async
([
&
executor
,
&
prepared
,
&
program
,
&
scope
,
idx
]()
{
int
run_block
=
idx
;
// thread local
try
{
VLOG
(
3
)
<<
"running server block: "
<<
run_block
<<
"pointer: "
<<
prepared
[
run_block
].
get
();
executor
->
RunPreparedContext
(
prepared
[
run_block
].
get
(),
scope
);
}
catch
(
const
std
::
exception
&
e
)
{
LOG
(
ERROR
)
<<
"run sub program error "
<<
e
.
what
();
...
...
@@ -107,12 +109,14 @@ void ListenAndServOp::RunSyncLoop(
PADDLE_ENFORCE_GE
(
num_blocks
,
2
,
"server program should have at least 2 blocks"
);
std
::
vector
<
int
>
optimize_blocks_idx
;
for
(
auto
blk
:
optimize_blocks
)
{
optimize_blocks_idx
.
push_back
(
blk
->
ID
());
// Prepare all the server block
std
::
vector
<
int
>
optimize_blocks_list
;
for
(
size_t
i
=
1
;
i
<
program
->
Size
();
++
i
)
{
optimize_blocks_list
.
push_back
(
i
);
}
auto
optimize_prepared
=
executor
->
Prepare
(
*
program
,
optimize_blocks_idx
);
// Insert placeholder for block0 which holds current op itself.
auto
optimize_prepared
=
executor
->
Prepare
(
*
program
,
optimize_blocks_list
);
// Insert placeholder for block0 which holds current op itself,
// NOTE the first block in `optimize_prepared` should never be ran.
optimize_prepared
.
insert
(
optimize_prepared
.
begin
(),
std
::
shared_ptr
<
framework
::
ExecutorPrepareContext
>
(
nullptr
));
...
...
python/paddle/fluid/tests/unittests/test_dist_transpiler.py
浏览文件 @
866fcb0c
...
...
@@ -304,10 +304,50 @@ class TestL2Decay(TranspilerTest):
# TODO(typhoonzero): test clipping and L2Decay ops are removed from trainer
# FIXME(typhoonzero): need to add test for async case:
# see https://github.com/PaddlePaddle/Paddle/issues/11691
class
TestAsyncSGD
(
TranspilerTest
):
pass
class
TestL2DecayWithPiecewise
(
TranspilerTest
):
def
net_conf
(
self
):
x
=
fluid
.
layers
.
data
(
name
=
'x'
,
shape
=
[
1000
],
dtype
=
'float32'
)
y_predict
=
fluid
.
layers
.
fc
(
input
=
x
,
size
=
1000
,
act
=
None
,
param_attr
=
fluid
.
ParamAttr
(
name
=
'fc_w'
),
bias_attr
=
fluid
.
ParamAttr
(
name
=
'fc_b'
))
y
=
fluid
.
layers
.
data
(
name
=
'y'
,
shape
=
[
1
],
dtype
=
'float32'
)
cost
=
fluid
.
layers
.
square_error_cost
(
input
=
y_predict
,
label
=
y
)
avg_cost
=
fluid
.
layers
.
mean
(
cost
)
base_lr
=
1.0
bd
=
[
1
,
10
,
20
,
30
]
lr
=
[
base_lr
*
(
0.1
**
i
)
for
i
in
range
(
len
(
bd
)
+
1
)]
sgd_optimizer
=
fluid
.
optimizer
.
Momentum
(
learning_rate
=
fluid
.
layers
.
piecewise_decay
(
boundaries
=
bd
,
values
=
lr
),
momentum
=
0.9
,
regularization
=
fluid
.
regularizer
.
L2Decay
(
1e-4
))
sgd_optimizer
.
minimize
(
avg_cost
)
return
def
test_transpiler
(
self
):
pserver
,
startup
=
self
.
get_pserver
(
self
.
pserver1_ep
)
trainer
=
self
.
get_trainer
()
self
.
assertEqual
(
len
(
pserver
.
blocks
),
9
)
self
.
assertEqual
([
op
.
type
for
op
in
pserver
.
blocks
[
1
].
ops
],
[
"increment"
,
"cast"
,
"fill_constant"
,
"fill_constant"
,
"less_than"
,
"logical_not"
,
"conditional_block"
,
"fill_constant"
,
"fill_constant"
,
"less_than"
,
"logical_not"
,
"logical_and"
,
"logical_and"
,
"conditional_block"
,
"fill_constant"
,
"fill_constant"
,
"less_than"
,
"logical_not"
,
"logical_and"
,
"logical_and"
,
"conditional_block"
,
"fill_constant"
,
"fill_constant"
,
"less_than"
,
"logical_not"
,
"logical_and"
,
"logical_and"
,
"conditional_block"
,
"fill_constant"
,
"conditional_block"
])
self
.
assertEqual
(
[
op
.
type
for
op
in
pserver
.
blocks
[
7
].
ops
],
[
"sum"
,
"scale"
,
"scale"
,
"elementwise_add"
,
"momentum"
])
self
.
assertEqual
(
[
op
.
type
for
op
in
pserver
.
blocks
[
8
].
ops
],
[
"sum"
,
"scale"
,
"scale"
,
"elementwise_add"
,
"momentum"
])
if
__name__
==
"__main__"
:
...
...
python/paddle/fluid/transpiler/distribute_transpiler.py
浏览文件 @
866fcb0c
...
...
@@ -461,6 +461,8 @@ class DistributeTranspiler(object):
per_opt_block
=
pserver_program
.
create_block
(
pre_block_idx
)
optimize_blocks
.
append
(
per_opt_block
)
# append grad merging ops before clip and weight decay
# cases may like:
# L2Decay op -> clip op -> optimize
for
_
,
op
in
enumerate
(
self
.
optimize_ops
):
# find the origin @GRAD var before clipping
grad_varname_for_block
=
__op_have_grad_input__
(
op
)
...
...
@@ -468,6 +470,7 @@ class DistributeTranspiler(object):
merged_var
=
self
.
_append_pserver_grad_merge_ops
(
per_opt_block
,
grad_varname_for_block
,
endpoint
,
grad_to_block_id
,
self
.
origin_program
)
break
# append optimize op once then append other ops.
for
_
,
op
in
enumerate
(
self
.
optimize_ops
):
# optimizer is connected to itself
if
ufind
.
is_connected
(
op
,
opt_op
)
and
op
not
in
global_ops
:
...
...
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