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
迷途归一
Paddle
提交
00e8791f
P
Paddle
项目概览
迷途归一
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
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 >>
提交
00e8791f
编写于
10月 17, 2018
作者:
D
dzhwinter
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
fix compile in cpu error. test=develop
上级
d239cf2e
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
37 addition
and
38 deletion
+37
-38
paddle/fluid/operators/momentum_op.cc
paddle/fluid/operators/momentum_op.cc
+9
-6
paddle/fluid/operators/momentum_op.h
paddle/fluid/operators/momentum_op.h
+12
-10
python/paddle/fluid/tests/unittests/test_momentum_op.py
python/paddle/fluid/tests/unittests/test_momentum_op.py
+16
-22
未找到文件。
paddle/fluid/operators/momentum_op.cc
浏览文件 @
00e8791f
...
...
@@ -45,12 +45,15 @@ class MomentumOp : public framework::OperatorWithKernel {
"Output(VelocityOut) of Momentum should not be null."
);
auto
param_dim
=
ctx
->
GetInputDim
(
"Param"
);
PADDLE_ENFORCE_EQ
(
param_dim
,
ctx
->
GetInputDim
(
"Grad"
),
"Param and Grad input of MomentumOp should have the same dimension."
);
PADDLE_ENFORCE_EQ
(
param_dim
,
ctx
->
GetInputDim
(
"Velocity"
),
"Param and Velocity of MomentumOp should have the same dimension."
);
if
(
ctx
->
GetInputsVarType
(
"Grad"
)[
0
]
==
framework
::
proto
::
VarType
::
LOD_TENSOR
)
{
PADDLE_ENFORCE_EQ
(
param_dim
,
ctx
->
GetInputDim
(
"Grad"
),
"Param and Grad input of MomentumOp should have the same dimension."
);
PADDLE_ENFORCE_EQ
(
param_dim
,
ctx
->
GetInputDim
(
"Velocity"
),
"Param and Velocity of MomentumOp should have the same dimension."
);
}
PADDLE_ENFORCE_EQ
(
framework
::
product
(
ctx
->
GetInputDim
(
"LearningRate"
)),
1
,
"Learning_rate should be a scalar"
);
...
...
paddle/fluid/operators/momentum_op.h
浏览文件 @
00e8791f
...
...
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <string>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/algorithm.h"
...
...
@@ -303,28 +304,30 @@ class MomentumOpKernel : public framework::OpKernel<T> {
auto
*
merged_grad
=
const_cast
<
framework
::
Scope
&>
(
ctx
.
scope
())
.
Var
()
->
GetMutable
<
framework
::
SelectedRows
>
();
math
::
scatter
::
MergeAdd
<
DeviceContext
,
T
>
merge_func
;
merge_func
(
ctx
.
template
device_context
<
DeviceContext
>(),
*
grad
,
merged_grad
);
platform
::
ForRange
<
DeviceContext
>
for_range
(
static_cast
<
const
DeviceContext
&>
(
ctx
.
device_context
()),
param
->
numel
());
const
int64_t
*
rows
=
nullptr
;
#ifdef PADDLE_WITH_CUDA
if
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()))
{
rows
=
merged_grad
->
rows
().
CUDAData
(
ctx
.
GetPlace
());
}
else
{
#endif
rows
=
merged_grad
->
rows
().
data
();
#ifdef PADDLE_WITH_CUDA
}
#endif
int64_t
row_numel
=
merged_grad
->
value
().
numel
()
/
merged_grad
->
rows
().
size
();
platform
::
ForRange
<
DeviceContext
>
for_range
(
static_cast
<
const
DeviceContext
&>
(
ctx
.
device_context
()),
param
->
numel
());
if
(
use_nesterov
)
{
SparseMomentumFunctor
<
T
,
UseNesterov
>
functor
(
param
->
data
<
T
>
(),
merged_grad
->
value
().
data
<
T
>
(),
velocity
->
data
<
T
>
(),
learning_rate
->
data
<
T
>
(),
mu
,
rows
,
velocity
->
data
<
T
>
(),
learning_rate
->
data
<
T
>
(),
mu
,
rows
,
row_numel
,
static_cast
<
int64_t
>
(
merged_grad
->
rows
().
size
()),
static_cast
<
int64_t
>
(
merged_grad
->
height
()),
param_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
()),
velocity_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
()));
for_range
(
functor
);
...
...
@@ -332,9 +335,8 @@ class MomentumOpKernel : public framework::OpKernel<T> {
}
else
{
SparseMomentumFunctor
<
T
,
NoNesterov
>
functor
(
param
->
data
<
T
>
(),
merged_grad
->
value
().
data
<
T
>
(),
velocity
->
data
<
T
>
(),
learning_rate
->
data
<
T
>
(),
mu
,
rows
,
velocity
->
data
<
T
>
(),
learning_rate
->
data
<
T
>
(),
mu
,
rows
,
row_numel
,
static_cast
<
int64_t
>
(
merged_grad
->
rows
().
size
()),
static_cast
<
int64_t
>
(
merged_grad
->
height
()),
param_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
()),
velocity_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
()));
for_range
(
functor
);
...
...
python/paddle/fluid/tests/unittests/test_momentum_op.py
浏览文件 @
00e8791f
...
...
@@ -121,22 +121,13 @@ class TestSparseMomentumOp(unittest.TestCase):
grad_tensor
=
grad_selected_rows
.
get_tensor
()
grad_tensor
.
set
(
grad_np_array
,
place
)
velocity_selected_rows
=
scope
.
var
(
'Velocity'
).
get_selected_rows
()
velocity_selected_rows
.
set_height
(
height
)
velocity_selected_rows
.
set_rows
(
rows
)
velocity_np_array
=
np
.
ones
((
len
(
rows
),
row_numel
)).
astype
(
"float32"
)
velocity_np_array
[
0
,
0
]
=
2.0
velocity_np_array
[
2
,
8
]
=
2.0
velocity_tensor
=
velocity_selected_rows
.
get_tensor
()
velocity_tensor
.
set
(
velocity_np_array
,
place
)
velocity_out_selected_rows
=
scope
.
var
(
'VelocityOut'
).
get_selected_rows
(
)
velocity_out_selected_rows
.
set_height
(
height
)
velocity_out_selected_rows
.
set_rows
(
rows
)
velocity_out_np_array
=
np
.
full
((
len
(
rows
),
row_numel
),
velocity
=
scope
.
var
(
'Velocity'
).
get_tensor
()
velocity_np_array
=
np
.
ones
((
height
,
row_numel
)).
astype
(
"float32"
)
velocity
.
set
(
velocity_np_array
,
place
)
velocity_out
=
scope
.
var
(
'VelocityOut'
).
get_tensor
()
velocity_out_np_array
=
np
.
full
((
height
,
row_numel
),
0.0
).
astype
(
"float32"
)
velocity_out_tensor
=
velocity_out_selected_rows
.
get_tensor
()
velocity_out_tensor
.
set
(
velocity_out_np_array
,
place
)
velocity_out
.
set
(
velocity_out_np_array
,
place
)
# create and initialize LeraningRate Variable
lr
=
scope
.
var
(
'LearningRate'
).
get_tensor
()
...
...
@@ -158,19 +149,22 @@ class TestSparseMomentumOp(unittest.TestCase):
# get and compare result
param_out_np_array
=
np
.
array
(
param_out
)
velocity_out_np_array
=
np
.
array
(
velocity_out
_tensor
)
velocity_out_np_array
=
np
.
array
(
velocity_out
)
# TODO(dzh): add a more suitable general numpy interface
# for sparse update.
_velocity_out
=
mu
*
velocity_np_array
+
grad_np_array
_param
=
param_array
[
rows
]
_grad_np_array
=
np
.
full
((
height
,
row_numel
),
0.0
).
astype
(
"float32"
)
for
i
in
range
(
len
(
rows
)):
_grad_np_array
[
rows
[
i
]]
=
grad_np_array
[
i
]
_velocity_out
=
mu
*
velocity_np_array
+
_grad_np_array
_param
=
param_array
if
use_nesterov
:
_param_out
=
_param
-
grad_np_array
*
lr_array
-
\
_velocity_out
*
mu
*
lr_array
_param_out
=
_param
-
(
_grad_np_array
+
_velocity_out
*
mu
)
*
lr_array
else
:
_param_out
=
_param
-
lr
*
_velocity_out
self
.
assertTrue
((
_param_out
==
param_out_np_array
[
rows
]).
all
())
_param_out
=
_param
-
lr_array
*
_velocity_out
self
.
assertTrue
((
_velocity_out
==
velocity_out_np_array
).
all
())
self
.
assertTrue
((
_param_out
==
param_out_np_array
).
all
())
def
init_kernel
(
self
):
pass
...
...
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