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
机器未来
Paddle
提交
163d2871
P
Paddle
项目概览
机器未来
/
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
提交
163d2871
编写于
10月 02, 2017
作者:
K
Kavya Srinet
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Made learning rate the input
上级
61c03f9d
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
20 addition
and
13 deletion
+20
-13
paddle/operators/rmsprop_op.cc
paddle/operators/rmsprop_op.cc
+11
-5
paddle/operators/rmsprop_op.h
paddle/operators/rmsprop_op.h
+1
-1
python/paddle/v2/framework/tests/test_rmsprop_op.py
python/paddle/v2/framework/tests/test_rmsprop_op.py
+8
-7
未找到文件。
paddle/operators/rmsprop_op.cc
浏览文件 @
163d2871
...
...
@@ -24,11 +24,13 @@ class RmspropOp : public framework::OperatorWithKernel {
protected:
void
InferShape
(
framework
::
InferShapeContextBase
*
ctx
)
const
override
{
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Param"
),
"Input(
p
aram) of RmspropOp should not be null."
);
"Input(
P
aram) of RmspropOp should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Grad"
),
"Input(
g
rad) of RmspropOp should not be null."
);
"Input(
G
rad) of RmspropOp should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Moment"
),
"Input(moment) of RmspropOp should not be null."
);
"Input(Moment) of RmspropOp should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"LearningRate"
),
"Input(LearningRate) of RmspropOp should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasOutput
(
"ParamOut"
),
"Output(param_out) of RmspropOp should not be null."
);
...
...
@@ -43,6 +45,10 @@ class RmspropOp : public framework::OperatorWithKernel {
param_dim
,
ctx
->
GetInputDim
(
"Moment"
),
"Param and moment input of RmspropOp should have the same dimension."
);
auto
lr_dim
=
ctx
->
GetInputDim
(
"LearningRate"
);
PADDLE_ENFORCE_EQ
(
framework
::
product
(
lr_dim
),
1
,
"Learning Rate should be a scalar."
);
ctx
->
SetOutputDim
(
"ParamOut"
,
param_dim
);
ctx
->
SetOutputDim
(
"MomentOut"
,
param_dim
);
}
...
...
@@ -56,11 +62,11 @@ class RmspropOpMaker : public framework::OpProtoAndCheckerMaker {
AddInput
(
"Param"
,
"Input parameter"
);
AddInput
(
"Grad"
,
"Input gradient"
);
AddInput
(
"Moment"
,
"Second moment"
);
AddInput
(
"LearningRate"
,
"Learning Rate"
);
AddOutput
(
"ParamOut"
,
"Output parameter"
);
AddOutput
(
"MomentOut"
,
"Output second moment"
);
AddAttr
<
float
>
(
"learningRate"
,
"Learning rate"
);
AddAttr
<
float
>
(
"epsilon"
,
"Constant for numerical stability"
);
AddAttr
<
float
>
(
"decayRate"
,
"Decay rate for moving average of gradients"
);
AddComment
(
R"DOC(
...
...
@@ -68,7 +74,7 @@ class RmspropOpMaker : public framework::OpProtoAndCheckerMaker {
RMSprop
MomentOut = decayRate * Moment + (1 - decayRate) * Grad * Grad
ParamOut = Param -
l
earningRate * Grad / (sqrt(MomentOut) + epsilon)
ParamOut = Param -
L
earningRate * Grad / (sqrt(MomentOut) + epsilon)
The original slide(Slide 29 of
http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf)
...
...
paddle/operators/rmsprop_op.h
浏览文件 @
163d2871
...
...
@@ -34,13 +34,13 @@ class RmspropOpKernel : public framework::OpKernel<T> {
param_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
moment_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
float
lr
=
ctx
.
Attr
<
float
>
(
"learningRate"
);
float
epsilon
=
ctx
.
Attr
<
float
>
(
"epsilon"
);
float
decay
=
ctx
.
Attr
<
float
>
(
"decayRate"
);
auto
p
=
EigenVector
<
T
>::
Flatten
(
*
ctx
.
Input
<
Tensor
>
(
"Param"
));
auto
g
=
EigenVector
<
T
>::
Flatten
(
*
ctx
.
Input
<
Tensor
>
(
"Grad"
));
auto
m
=
EigenVector
<
T
>::
Flatten
(
*
ctx
.
Input
<
Tensor
>
(
"Moment"
));
float
lr
=
ctx
.
Input
<
Tensor
>
(
"LearningRate"
)
->
data
<
float
>
()[
0
];
auto
p_out
=
EigenVector
<
T
>::
Flatten
(
*
param_out
);
auto
m_out
=
EigenVector
<
T
>::
Flatten
(
*
moment_out
);
auto
place
=
ctx
.
GetEigenDevice
<
Place
>
();
...
...
python/paddle/v2/framework/tests/test_rmsprop_op.py
浏览文件 @
163d2871
...
...
@@ -10,19 +10,20 @@ class TestRmspropOp(OpTest):
param
=
np
.
random
.
random
((
123
,
321
)).
astype
(
"float32"
)
grad
=
np
.
random
.
random
((
123
,
321
)).
astype
(
"float32"
)
moment
=
np
.
zeros
((
123
,
321
)).
astype
(
"float32"
)
learning_rate
=
np
.
array
([
0.01
]).
astype
(
"float32"
)
learning_rate
=
0.01
epsilon
=
1e-6
decay_rate
=
0.9
self
.
inputs
=
{
'Param'
:
param
,
'Grad'
:
grad
,
'Moment'
:
moment
}
self
.
attrs
=
{
'learningRate'
:
learning_rate
,
'epsilon'
:
epsilon
,
'decayRate'
:
decay_rate
self
.
inputs
=
{
'Param'
:
param
,
'Grad'
:
grad
,
'Moment'
:
moment
,
'LearningRate'
:
learning_rate
}
self
.
attrs
=
{
'epsilon'
:
epsilon
,
'decayRate'
:
decay_rate
}
moment_out
=
decay_rate
*
moment
+
(
1
-
decay_rate
)
*
grad
*
grad
param_out
=
param
-
learning_rate
*
grad
/
(
np
.
sqrt
(
moment_out
)
+
epsilon
)
...
...
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