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
wzpei9
Paddle
提交
6398c15c
P
Paddle
项目概览
wzpei9
/
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
6398c15c
编写于
7月 09, 2017
作者:
D
dzhwinter
提交者:
GitHub
7月 09, 2017
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #2718 from dzhwinter/lr_state
"lr state serialization"
上级
bc368525
45adbfc4
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
68 addition
and
39 deletion
+68
-39
paddle/optimizer/adadelta_optimizer.cc
paddle/optimizer/adadelta_optimizer.cc
+5
-3
paddle/optimizer/adagrad_optimizer.cc
paddle/optimizer/adagrad_optimizer.cc
+6
-3
paddle/optimizer/adam_optimizer.cc
paddle/optimizer/adam_optimizer.cc
+6
-3
paddle/optimizer/lr_policy.h
paddle/optimizer/lr_policy.h
+34
-14
paddle/optimizer/sgd_optimizer.cc
paddle/optimizer/sgd_optimizer.cc
+5
-1
proto/OptimizerConfig.proto
proto/OptimizerConfig.proto
+12
-15
未找到文件。
paddle/optimizer/adadelta_optimizer.cc
浏览文件 @
6398c15c
...
...
@@ -27,22 +27,24 @@ void AdadeltaOptimizer::Update(const Tensor* gradient) {
const
char
*
AdadeltaOptimizer
::
SerializeState
(
int
*
state_len
)
{
AdadeltaOptimizerState
state
;
// TODO(zhihong) : add lr_policy serialization
state
.
set_num_sample_passed
(
num_sample_passed_
);
std
::
string
lr_str
=
this
->
lr_policy_
->
SerializeState
(
state_len
);
state
.
mutable_lr_state
()
->
ParseFromString
(
lr_str
);
TensorToProto
(
*
parameter_
,
state
.
mutable_parameter
());
TensorToProto
(
*
accum_gradient_
,
state
.
mutable_accum_gradient
());
TensorToProto
(
*
accum_delta_
,
state
.
mutable_accum_delta
());
TensorToProto
(
*
update_delta_
,
state
.
mutable_update_delta
());
auto
str
=
state
.
SerializeAsString
();
*
state_len
=
str
.
size
();
*
state_len
+
=
str
.
size
();
return
str
.
c_str
();
}
void
AdadeltaOptimizer
::
DeserializeState
(
const
std
::
string
&
str
)
{
AdadeltaOptimizerState
state
;
state
.
ParseFromString
(
str
);
// TODO(zhihong) : add lr_policy DeserializeState
auto
lr_state
=
state
.
lr_state
();
this
->
lr_policy_
->
DeserializeState
(
lr_state
.
SerializeAsString
());
num_sample_passed_
=
state
.
num_sample_passed
();
ProtoToTensor
(
state
.
parameter
(),
parameter_
);
...
...
paddle/optimizer/adagrad_optimizer.cc
浏览文件 @
6398c15c
...
...
@@ -19,20 +19,23 @@ void AdagradOptimizer::Update(const Tensor* gradient) {
}
const
char
*
AdagradOptimizer
::
SerializeState
(
int
*
state_len
)
{
AdagradOptimizerState
state
;
// TODO(zhihong) : add lr_policy serialization
state
.
set_num_sample_passed
(
num_sample_passed_
);
std
::
string
lr_str
=
this
->
lr_policy_
->
SerializeState
(
state_len
);
state
.
mutable_lr_state
()
->
ParseFromString
(
lr_str
);
TensorToProto
(
*
parameter_
,
state
.
mutable_parameter
());
TensorToProto
(
*
accum_gradient_
,
state
.
mutable_accum_gradient
());
auto
str
=
state
.
SerializeAsString
();
*
state_len
=
str
.
size
();
*
state_len
+
=
str
.
size
();
return
str
.
c_str
();
}
void
AdagradOptimizer
::
DeserializeState
(
const
std
::
string
&
str
)
{
AdagradOptimizerState
state
;
state
.
ParseFromString
(
str
);
// TODO(zhihong) : add lr_policy DeserializeState
auto
lr_state
=
state
.
lr_state
();
this
->
lr_policy_
->
DeserializeState
(
lr_state
.
SerializeAsString
());
num_sample_passed_
=
state
.
num_sample_passed
();
ProtoToTensor
(
state
.
parameter
(),
parameter_
);
ProtoToTensor
(
state
.
accum_gradient
(),
accum_gradient_
);
...
...
paddle/optimizer/adam_optimizer.cc
浏览文件 @
6398c15c
...
...
@@ -24,20 +24,23 @@ void AdamOptimizer::Update(const Tensor *gradient) {
const
char
*
AdamOptimizer
::
SerializeState
(
int
*
state_len
)
{
AdamOptimizerState
state
;
// TODO(zhihong) : add lr_policy serialization
std
::
string
lr_str
=
this
->
lr_policy_
->
SerializeState
(
state_len
);
state
.
mutable_lr_state
()
->
ParseFromString
(
lr_str
);
state
.
set_num_sample_passed
(
num_sample_passed_
);
TensorToProto
(
*
parameter_
,
state
.
mutable_parameter
());
TensorToProto
(
*
momentums_
,
state
.
mutable_momentums
());
TensorToProto
(
*
velocitys_
,
state
.
mutable_velocitys
());
auto
str
=
state
.
SerializeAsString
();
*
state_len
=
str
.
size
();
*
state_len
+
=
str
.
size
();
return
str
.
c_str
();
}
void
AdamOptimizer
::
DeserializeState
(
const
std
::
string
&
str
)
{
AdamOptimizerState
state
;
state
.
ParseFromString
(
str
);
// TODO(zhihong) : add lr_policy DeserializeState
auto
lr_state
=
state
.
lr_state
();
this
->
lr_policy_
->
DeserializeState
(
lr_state
.
SerializeAsString
());
num_sample_passed_
=
state
.
num_sample_passed
();
ProtoToTensor
(
state
.
parameter
(),
parameter_
);
...
...
paddle/optimizer/lr_policy.h
浏览文件 @
6398c15c
...
...
@@ -17,36 +17,56 @@ public:
// constant learning rate policy
class
ConstLr
final
:
public
LrPolicy
{
public:
ConstLr
(
double
lr
)
:
learning_rate
(
lr
){};
ConstLr
(
double
lr
)
:
learning_rate
_
(
lr
){};
double
LearningRate
(
const
uint64_t
num_sample_passed
)
{
return
learning_rate
;
return
learning_rate_
;
}
const
char
*
SerializeState
(
int
*
state_len
)
{
LrPolicyState
state
;
state
.
set_learning_rate
(
learning_rate_
);
auto
str
=
state
.
SerializeAsString
();
*
state_len
=
str
.
size
();
return
str
.
c_str
();
}
void
DeserializeState
(
const
std
::
string
&
str
)
{
LrPolicyState
state
;
state
.
ParseFromString
(
str
);
learning_rate_
=
state
.
learning_rate
();
}
const
char
*
SerializeState
(
int
*
state_len
)
{
return
nullptr
;
}
void
DeserializeState
(
const
std
::
string
&
state
)
{}
private:
double
learning_rate
;
double
learning_rate
_
;
};
class
LinearLr
final
:
public
LrPolicy
{
public:
LinearLr
(
double
lr
,
double
lr_decay_a
,
double
lr_decay_b
)
:
learning_rate
(
lr
),
lr_decay_a
(
lr_decay_a
),
lr_decay_b
(
lr_decay_b
)
{}
:
learning_rate
_
(
lr
),
lr_decay_a_
(
lr_decay_a
),
lr_decay_b_
(
lr_decay_b
)
{}
double
LearningRate
(
const
uint64_t
num_sample_passed
)
{
return
std
::
max
(
learning_rate
-
lr_decay_a
*
num_sample_passed
,
lr_decay_b
);
return
std
::
max
(
learning_rate_
-
lr_decay_a_
*
num_sample_passed
,
lr_decay_b_
);
}
const
char
*
SerializeState
(
int
*
state_len
)
{
// TODO(zhihong) : add lr_policy serialization
return
nullptr
;
LrPolicyState
state
;
state
.
set_learning_rate
(
learning_rate_
);
state
.
set_lr_decay_a
(
lr_decay_a_
);
state
.
set_lr_decay_b
(
lr_decay_b_
);
auto
str
=
state
.
SerializeAsString
();
*
state_len
=
str
.
size
();
return
str
.
c_str
();
}
void
DeserializeState
(
const
std
::
string
&
state
)
{
// TODO(zhihong) : add lr_policy serialization
void
DeserializeState
(
const
std
::
string
&
str
)
{
LrPolicyState
state
;
state
.
ParseFromString
(
str
);
learning_rate_
=
state
.
learning_rate
();
lr_decay_a_
=
state
.
lr_decay_a
();
lr_decay_b_
=
state
.
lr_decay_b
();
}
private:
double
learning_rate
;
double
lr_decay_a
;
double
lr_decay_b
;
double
learning_rate
_
;
double
lr_decay_a
_
;
double
lr_decay_b
_
;
};
}
// namespace optimizer
...
...
paddle/optimizer/sgd_optimizer.cc
浏览文件 @
6398c15c
...
...
@@ -30,16 +30,20 @@ void SGDOptimizer::Update(const Tensor *gradient) {
const
char
*
SGDOptimizer
::
SerializeState
(
int
*
state_len
)
{
SGDOptimizerState
state
;
state
.
set_num_sample_passed
(
num_sample_passed_
);
std
::
string
lr_str
=
this
->
lr_policy_
->
SerializeState
(
state_len
);
state
.
mutable_lr_state
()
->
ParseFromString
(
lr_str
);
TensorToProto
(
*
parameter_
,
state
.
mutable_parameter
());
if
(
momentum_
!=
0.0
)
TensorToProto
(
*
momentums_
,
state
.
mutable_momentums
());
auto
str
=
state
.
SerializeAsString
();
*
state_len
=
str
.
size
();
*
state_len
+
=
str
.
size
();
return
str
.
c_str
();
}
void
SGDOptimizer
::
DeserializeState
(
const
std
::
string
&
str
)
{
SGDOptimizerState
state
;
state
.
ParseFromString
(
str
);
auto
lr_state
=
state
.
lr_state
();
this
->
lr_policy_
->
DeserializeState
(
lr_state
.
SerializeAsString
());
num_sample_passed_
=
state
.
num_sample_passed
();
ProtoToTensor
(
state
.
parameter
(),
parameter_
);
if
(
momentum_
!=
0.0
)
ProtoToTensor
(
state
.
parameter
(),
momentums_
);
...
...
proto/OptimizerConfig.proto
浏览文件 @
6398c15c
...
...
@@ -78,11 +78,15 @@ enum DataType {
repeated
bytes
content
=
2
;
}
message
LrPolicyState
{
// learninRate Policy
optional
double
learning_rate
=
1
[
default
=
1.0
];
optional
double
lr_decay_a
=
2
;
optional
double
lr_decay_b
=
3
;
}
message
SGDOptimizerState
{
// learning rate policy
optional
double
learning_rate
=
101
;
optional
double
lr_decay_a
=
102
;
optional
double
lr_decay_b
=
103
;
optional
LrPolicyState
lr_state
=
101
;
optional
double
num_sample_passed
=
104
;
// state
optional
TensorProto
parameter
=
1
;
...
...
@@ -91,9 +95,7 @@ message SGDOptimizerState {
message
AdadeltaOptimizerState
{
// learning rate policy
optional
double
learning_rate
=
101
;
optional
double
lr_decay_a
=
102
;
optional
double
lr_decay_b
=
103
;
optional
LrPolicyState
lr_state
=
101
;
optional
double
num_sample_passed
=
104
;
// state
optional
TensorProto
parameter
=
1
;
...
...
@@ -102,11 +104,9 @@ message AdadeltaOptimizerState {
optional
TensorProto
update_delta
=
4
;
}
message
AdagradOptimizerState
{
// learning rate policy
optional
double
learning_rate
=
101
;
optional
double
lr_decay_a
=
102
;
optional
double
lr_decay_b
=
103
;
optional
LrPolicyState
lr_state
=
101
;
optional
double
num_sample_passed
=
104
;
// state
optional
TensorProto
parameter
=
1
;
...
...
@@ -114,10 +114,7 @@ message AdagradOptimizerState {
}
message
AdamOptimizerState
{
// learning rate policy
optional
double
learning_rate
=
101
;
optional
double
lr_decay_a
=
102
;
optional
double
lr_decay_b
=
103
;
optional
LrPolicyState
lr_state
=
101
;
optional
double
num_sample_passed
=
104
;
// state
optional
TensorProto
parameter
=
1
;
...
...
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