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
wk1234123
Paddle
提交
18be7256
P
Paddle
项目概览
wk1234123
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
4
Star
0
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
0
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
0
Issue
0
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 >>
未验证
提交
18be7256
编写于
10月 26, 2018
作者:
X
Xin Pan
提交者:
GitHub
10月 26, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #14019 from panyx0718/fix6
fix multi device dependency issue
上级
b717a32f
4cd44c00
变更
8
隐藏空白更改
内联
并排
Showing
8 changed file
with
40 addition
and
15 deletion
+40
-15
paddle/fluid/framework/details/multi_devices_graph_pass.cc
paddle/fluid/framework/details/multi_devices_graph_pass.cc
+3
-3
paddle/fluid/framework/op_proto_maker.cc
paddle/fluid/framework/op_proto_maker.cc
+2
-0
paddle/fluid/framework/op_proto_maker.h
paddle/fluid/framework/op_proto_maker.h
+3
-0
python/paddle/fluid/clip.py
python/paddle/fluid/clip.py
+4
-2
python/paddle/fluid/framework.py
python/paddle/fluid/framework.py
+12
-3
python/paddle/fluid/optimizer.py
python/paddle/fluid/optimizer.py
+9
-4
python/paddle/fluid/regularizer.py
python/paddle/fluid/regularizer.py
+2
-1
python/paddle/fluid/transpiler/distribute_transpiler.py
python/paddle/fluid/transpiler/distribute_transpiler.py
+5
-2
未找到文件。
paddle/fluid/framework/details/multi_devices_graph_pass.cc
浏览文件 @
18be7256
...
...
@@ -252,9 +252,9 @@ std::vector<ir::Node *> SortOpsAndDelayOptimizeOp(const ir::Graph &graph) {
std
::
vector
<
ir
::
Node
*>
sorted_ret
;
for
(
size_t
i
=
0
;
i
<
ret
.
size
();
++
i
)
{
if
(
i
<
last_backward
)
{
if
(
boost
::
get
<
int
>
(
ret
[
i
]
->
Op
()
->
GetAttr
(
OpProtoAndCheckerMaker
::
OpRoleAttrName
()))
==
static_cast
<
int
>
(
OpRole
::
kOptimize
))
{
if
(
static_cast
<
bool
>
(
boost
::
get
<
int
>
(
ret
[
i
]
->
Op
()
->
GetAttr
(
OpProtoAndCheckerMaker
::
OpRoleAttrName
()))
&
static_cast
<
int
>
(
OpRole
::
kOptimize
)
))
{
optimize_ops
.
push_back
(
ret
[
i
]);
}
else
{
sorted_ret
.
push_back
(
ret
[
i
]);
...
...
paddle/fluid/framework/op_proto_maker.cc
浏览文件 @
18be7256
...
...
@@ -71,6 +71,8 @@ void OpProtoAndCheckerMaker::operator()(proto::OpProto* proto,
static_cast
<
int
>
(
OpRole
::
kLoss
)
|
static_cast
<
int
>
(
OpRole
::
kForward
),
static_cast
<
int
>
(
OpRole
::
kLoss
)
|
static_cast
<
int
>
(
OpRole
::
kBackward
),
static_cast
<
int
>
(
OpRole
::
kOptimize
)
|
static_cast
<
int
>
(
OpRole
::
kLRSched
),
static_cast
<
int
>
(
OpRole
::
kNotSpecified
)})
.
SetDefault
(
static_cast
<
int
>
(
OpRole
::
kNotSpecified
));
AddAttr
<
std
::
vector
<
std
::
string
>>
(
OpRoleVarAttrName
(),
...
...
paddle/fluid/framework/op_proto_maker.h
浏览文件 @
18be7256
...
...
@@ -20,6 +20,9 @@ limitations under the License. */
namespace
paddle
{
namespace
framework
{
//////////////////////////
// Don't add more roles to make this too complicated!
//////////////////////////
enum
class
OpRole
{
kForward
=
0x0000
,
kBackward
=
0x0001
,
...
...
python/paddle/fluid/clip.py
浏览文件 @
18be7256
...
...
@@ -333,7 +333,8 @@ def append_gradient_clip_ops(param_grads):
for
p
,
g
in
param_grads
:
if
g
is
None
:
continue
with
p
.
block
.
program
.
_optimized_guard
([
p
,
g
]):
with
p
.
block
.
program
.
_optimized_guard
(
[
p
,
g
]),
framework
.
name_scope
(
'append_clip'
):
clip_attr
=
getattr
(
p
,
'gradient_clip_attr'
,
NullGradientClipAttr
())
if
clip_attr
is
None
:
clip_attr
=
NullGradientClipAttr
()
...
...
@@ -348,7 +349,8 @@ def append_gradient_clip_ops(param_grads):
for
p
,
g
in
param_grads
:
if
g
is
None
:
continue
with
p
.
block
.
program
.
_optimized_guard
([
p
,
g
]):
with
p
.
block
.
program
.
_optimized_guard
(
[
p
,
g
]),
framework
.
name_scope
(
'append_graident_clip'
):
res
.
append
(
clip_attr
.
_create_operators
(
param
=
p
,
grad
=
g
))
return
res
...
...
python/paddle/fluid/framework.py
浏览文件 @
18be7256
...
...
@@ -1496,6 +1496,9 @@ class Program(object):
>>> with program._optimized_guard([p,g]):
>>> p = p - 0.001 * g
"""
tmp_role
=
self
.
_current_role
tmp_var
=
self
.
_op_role_var
OpRole
=
core
.
op_proto_and_checker_maker
.
OpRole
self
.
_current_role
=
OpRole
.
Optimize
self
.
_op_role_var
=
[
...
...
@@ -1503,11 +1506,11 @@ class Program(object):
for
var
in
param_and_grads
]
yield
self
.
_op_role_var
=
[]
self
.
_current_role
=
OpRole
.
Forward
self
.
_op_role_var
=
tmp_var
self
.
_current_role
=
tmp_role
@
contextlib
.
contextmanager
def
_lr_schedule_guard
(
self
):
def
_lr_schedule_guard
(
self
,
is_with_opt
=
False
):
"""
A with guard to set :code:`LRSched` :code:`OpRole` and
:code:`OpRoleVar` automatically. The :code:`OpRoleVar` is
...
...
@@ -1515,6 +1518,10 @@ class Program(object):
Notes: This is a very low level API. Users should not use it directly.
Args:
is_with_opt: Only set to true if these ops a in the middle
of a bunch of optimize ops so that it can be treated
correctly. For example, sgd->lr_op->sgd->lr_op->sgd.
Examples:
...
...
@@ -1528,6 +1535,8 @@ class Program(object):
OpRole
=
core
.
op_proto_and_checker_maker
.
OpRole
self
.
_current_role
=
OpRole
.
LRSched
if
is_with_opt
:
self
.
_current_role
=
int
(
OpRole
.
LRSched
)
|
int
(
OpRole
.
Optimize
)
# TODO(typhoonzero): how to set target learning rate var
self
.
_op_role_var
=
[]
yield
...
...
python/paddle/fluid/optimizer.py
浏览文件 @
18be7256
...
...
@@ -111,7 +111,9 @@ class Optimizer(object):
if
param_lr
==
1.0
:
return
self
.
_global_learning_rate
()
else
:
with
default_main_program
().
_lr_schedule_guard
():
with
default_main_program
().
_lr_schedule_guard
(
is_with_opt
=
True
),
framework
.
name_scope
(
'scale_with_param_lr'
):
return
self
.
_global_learning_rate
()
*
param_lr
def
_create_accumulators
(
self
,
block
,
parameters
):
...
...
@@ -602,7 +604,8 @@ class AdamOptimizer(Optimizer):
for
param
,
grad
in
param_and_grads
:
if
grad
is
None
:
continue
with
param
.
block
.
program
.
_optimized_guard
([
param
,
grad
]):
with
param
.
block
.
program
.
_optimized_guard
(
[
param
,
grad
]),
name_scope
(
"optimizer"
):
beta1_pow_acc
=
self
.
_get_accumulator
(
self
.
_beta1_pow_acc_str
,
param
)
beta2_pow_acc
=
self
.
_get_accumulator
(
self
.
_beta2_pow_acc_str
,
...
...
@@ -740,7 +743,8 @@ class AdamaxOptimizer(Optimizer):
for
param
,
grad
in
parameters_and_grads
:
if
grad
is
None
:
continue
with
param
.
block
.
program
.
_optimized_guard
([
param
,
grad
]):
with
param
.
block
.
program
.
_optimized_guard
(
[
param
,
grad
]),
name_scope
(
'adamx'
):
beta1_pow_acc
=
self
.
_get_accumulator
(
self
.
_beta1_pow_acc_str
,
param
)
main_block
.
append_op
(
...
...
@@ -1279,7 +1283,8 @@ class ModelAverage(Optimizer):
for
param
,
grad
in
self
.
params_grads
:
if
grad
is
None
:
continue
with
param
.
block
.
program
.
_optimized_guard
([
param
,
grad
]):
with
param
.
block
.
program
.
_optimized_guard
(
[
param
,
grad
]),
name_scope
(
'move_average'
):
self
.
_append_average_accumulate_op
(
param
)
self
.
apply_program
=
Program
()
...
...
python/paddle/fluid/regularizer.py
浏览文件 @
18be7256
...
...
@@ -47,7 +47,8 @@ def append_regularization_ops(parameters_and_grads, regularization=None):
if
grad
is
None
:
params_and_grads
.
append
((
param
,
grad
))
continue
with
param
.
block
.
program
.
_optimized_guard
([
param
,
grad
]):
with
param
.
block
.
program
.
_optimized_guard
(
[
param
,
grad
]),
framework
.
name_scope
(
'regularization'
):
regularization_term
=
None
if
param
.
regularizer
is
not
None
:
# Add variable for regularization term in grad block
...
...
python/paddle/fluid/transpiler/distribute_transpiler.py
浏览文件 @
18be7256
...
...
@@ -49,6 +49,7 @@ LOOKUP_TABLE_GRAD_TYPE = "lookup_table_grad"
OP_ROLE_VAR_ATTR_NAME
=
core
.
op_proto_and_checker_maker
.
kOpRoleVarAttrName
()
RPC_OP_ROLE_ATTR_NAME
=
op_role_attr_name
=
core
.
op_proto_and_checker_maker
.
kOpRoleAttrName
(
)
OPT_OP_ROLE_ATTR_VALUE
=
core
.
op_proto_and_checker_maker
.
OpRole
.
Optimize
RPC_OP_ROLE_ATTR_VALUE
=
core
.
op_proto_and_checker_maker
.
OpRole
.
RPC
DIST_OP_ROLE_ATTR_VALUE
=
core
.
op_proto_and_checker_maker
.
OpRole
.
Dist
LR_SCHED_OP_ROLE_ATTR_VALUE
=
core
.
op_proto_and_checker_maker
.
OpRole
.
LRSched
...
...
@@ -1717,8 +1718,10 @@ to transpile() call.")
lr_ops
=
[]
block
=
self
.
origin_program
.
global_block
()
for
op
in
block
.
ops
:
if
int
(
op
.
attr
(
RPC_OP_ROLE_ATTR_NAME
))
==
int
(
LR_SCHED_OP_ROLE_ATTR_VALUE
):
role_id
=
int
(
op
.
attr
(
RPC_OP_ROLE_ATTR_NAME
))
if
role_id
==
int
(
LR_SCHED_OP_ROLE_ATTR_VALUE
)
or
\
role_id
==
int
(
LR_SCHED_OP_ROLE_ATTR_VALUE
)
|
\
int
(
OPT_OP_ROLE_ATTR_VALUE
):
lr_ops
.
append
(
op
)
log
(
"append lr op: "
,
op
.
type
)
return
lr_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