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
f0d108f5
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
f0d108f5
编写于
3月 15, 2019
作者:
S
sneaxiy
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
fix const_cast
test=develop
上级
1a4a90a8
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
25 addition
and
61 deletion
+25
-61
paddle/fluid/operators/optimizers/adam_op.h
paddle/fluid/operators/optimizers/adam_op.h
+15
-34
paddle/fluid/operators/optimizers/momentum_op.h
paddle/fluid/operators/optimizers/momentum_op.h
+6
-13
paddle/fluid/operators/optimizers/rmsprop_op.h
paddle/fluid/operators/optimizers/rmsprop_op.h
+4
-14
未找到文件。
paddle/fluid/operators/optimizers/adam_op.h
浏览文件 @
f0d108f5
...
...
@@ -15,6 +15,7 @@ limitations under the License. */
#pragma once
#include <math.h> // for sqrt in CPU and CUDA
#include <Eigen/Dense>
#include <unordered_map>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/threadpool.h"
...
...
@@ -311,17 +312,17 @@ struct SparseAdamFunctor<T, CPUAdam> {
T
beta1_pow
=
*
beta1_pow_
;
T
beta2_pow
=
*
beta2_pow_
;
lr
*=
sqrt
(
1
-
beta2_pow
)
/
(
1
-
beta1_pow
);
size_t
row_count
=
numel
/
row_numel_
;
int64_t
row_count
=
static_cast
<
int64_t
>
(
numel
/
row_numel_
)
;
for
(
size_t
i
=
0U
,
j
=
0U
;
i
!=
row_count
;
++
i
)
{
for
(
int64_t
i
=
0
,
j
=
0
;
i
!=
row_count
;
++
i
)
{
if
(
i
==
*
(
rows_
+
j
))
{
for
(
size_t
k
=
0U
;
k
!=
row_numel_
;
++
k
)
{
for
(
int64_t
k
=
0
;
k
!=
row_numel_
;
++
k
)
{
T
g
=
grad_
[
j
*
row_numel_
+
k
];
adam_update
(
i
*
row_numel_
+
k
,
g
);
}
++
j
;
}
else
{
for
(
size_t
k
=
0U
;
k
!=
row_numel_
;
++
k
)
{
for
(
int64_t
k
=
0
;
k
!=
row_numel_
;
++
k
)
{
T
mom1
=
moment1_
[
i
*
row_numel_
+
k
];
T
mom2
=
moment2_
[
i
*
row_numel_
+
k
];
T
p
=
param_
[
i
*
row_numel_
+
k
];
...
...
@@ -427,43 +428,23 @@ class AdamOpKernel : public framework::OpKernel<T> {
}
}
framework
::
SelectedRows
cpu
_grad_merge
;
framework
::
SelectedRows
tmp
_grad_merge
;
const
framework
::
SelectedRows
*
grad_merge_ptr
;
if
(
is_strict_sorted
)
{
grad_merge_ptr
=
&
grad
;
}
else
{
// merge duplicated rows if any.
// The rows of grad_merge have been sorted inside MergeAdd functor
framework
::
SelectedRows
*
grad_merge_var
;
scatter
::
MergeAdd
<
DeviceContext
,
T
>
merge_func
;
if
(
platform
::
is_cpu_place
(
ctx
.
GetPlace
()))
{
grad_merge_var
=
&
cpu_grad_merge
;
}
else
{
// FIXME(qiao): GPU also need to fix this
grad_merge_var
=
const_cast
<
framework
::
Scope
&>
(
ctx
.
scope
())
.
Var
()
->
GetMutable
<
framework
::
SelectedRows
>
();
}
merge_func
(
ctx
.
template
device_context
<
DeviceContext
>(),
grad
,
grad_merge_var
,
true
);
grad_merge_ptr
=
grad_merge_var
;
&
tmp_grad_merge
,
true
);
grad_merge_ptr
=
&
tmp_grad_merge
;
}
auto
&
grad_merge
=
*
grad_merge_ptr
;
auto
&
grad_tensor
=
grad_merge
.
value
();
const
T
*
grad_data
=
grad_tensor
.
template
data
<
T
>();
const
int64_t
*
rows
=
nullptr
;
// When compiled without CUDA, the CUDAData() interface should not be
// provided.
#if defined(PADDLE_WITH_CUDA)
if
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()))
{
rows
=
grad_merge
.
rows
().
CUDAData
(
ctx
.
GetPlace
());
}
else
{
#endif
rows
=
grad_merge
.
rows
().
data
();
#if defined(PADDLE_WITH_CUDA)
}
#endif
const
int64_t
*
rows
=
grad_merge
.
rows
().
Data
(
ctx
.
GetPlace
());
auto
row_numel
=
grad_tensor
.
numel
()
/
grad_merge
.
rows
().
size
();
if
(
platform
::
is_cpu_place
(
ctx
.
GetPlace
()))
{
...
...
@@ -488,7 +469,7 @@ class AdamOpKernel : public framework::OpKernel<T> {
}
}
#ifndef _WIN32
else
if
(
FLAGS_inner_op_parallelism
>
1
&&
else
if
(
FLAGS_inner_op_parallelism
>
1
&&
// NOLINT
min_row_size_to_use_multithread
>
0
&&
param
.
dims
()[
0
]
>
min_row_size_to_use_multithread
)
{
VLOG
(
3
)
<<
"use multi thread, inner_op_parallelism="
...
...
@@ -516,11 +497,11 @@ class AdamOpKernel : public framework::OpKernel<T> {
for
(
int
i
=
0
;
i
<
FLAGS_inner_op_parallelism
;
++
i
)
{
int64_t
start
=
i
*
line_in_each_thread
;
int64_t
end
=
(
i
+
1
)
*
line_in_each_thread
;
if
(
start
>=
param_row_count
)
{
if
(
start
>=
static_cast
<
int64_t
>
(
param_row_count
)
)
{
break
;
}
if
(
end
>
param_row_count
)
{
end
=
param_row_count
;
if
(
end
>
static_cast
<
int64_t
>
(
param_row_count
)
)
{
end
=
static_cast
<
int64_t
>
(
param_row_count
)
;
}
fs
.
push_back
(
framework
::
Async
([
&
functor
,
&
row_id_to_grad_row_offset
,
...
...
@@ -545,8 +526,8 @@ class AdamOpKernel : public framework::OpKernel<T> {
}
for
(
size_t
i
=
0
;
i
<
fs
.
size
();
++
i
)
fs
[
i
].
wait
();
}
#endif // !_WIN32
else
{
#endif
// !_WIN32
else
{
// NOLINT
functor
(
param
.
numel
());
}
}
else
if
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()))
{
...
...
paddle/fluid/operators/optimizers/momentum_op.h
浏览文件 @
f0d108f5
...
...
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#pragma once
#include <memory>
#include <string>
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
...
...
@@ -69,6 +70,7 @@ class MomentumOp : public framework::OperatorWithKernel {
ctx
->
SetOutputDim
(
"ParamOut"
,
param_dim
);
ctx
->
SetOutputDim
(
"VelocityOut"
,
param_dim
);
}
framework
::
OpKernelType
GetExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
input_data_type
=
framework
::
GetDataTypeOfVar
(
ctx
.
InputVar
(
"Param"
));
...
...
@@ -351,23 +353,14 @@ class MomentumOpKernel : public framework::OpKernel<T> {
VLOG
(
3
)
<<
"Grad SelectedRows contains no data!"
;
return
;
}
auto
*
merged_grad
=
const_cast
<
framework
::
Scope
&>
(
ctx
.
scope
())
.
Var
()
->
GetMutable
<
framework
::
SelectedRows
>
()
;
framework
::
SelectedRows
tmp_merged_grad
;
framework
::
SelectedRows
*
merged_grad
=
&
tmp_merged_grad
;
math
::
scatter
::
MergeAdd
<
DeviceContext
,
T
>
merge_func
;
merge_func
(
ctx
.
template
device_context
<
DeviceContext
>(),
*
grad
,
merged_grad
);
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
const
int64_t
*
rows
=
merged_grad
->
rows
().
Data
(
ctx
.
GetPlace
());
int64_t
row_numel
=
merged_grad
->
value
().
numel
()
/
merged_grad
->
rows
().
size
();
platform
::
ForRange
<
DeviceContext
>
for_range
(
...
...
paddle/fluid/operators/optimizers/rmsprop_op.h
浏览文件 @
f0d108f5
...
...
@@ -216,24 +216,14 @@ class RmspropOpKernel : public framework::OpKernel<T> {
}
}
else
if
(
grad_var
->
IsType
<
framework
::
SelectedRows
>
())
{
auto
&
grad
=
grad_var
->
Get
<
framework
::
SelectedRows
>
();
auto
*
merged_grad
=
const_cast
<
framework
::
Scope
&>
(
ctx
.
scope
())
.
Var
()
->
GetMutable
<
framework
::
SelectedRows
>
();
framework
::
SelectedRows
tmp_merged_grad
;
framework
::
SelectedRows
*
merged_grad
=
&
tmp_merged_grad
;
math
::
scatter
::
MergeAdd
<
DeviceContext
,
T
>
merge_func
;
merge_func
(
dev_ctx
,
grad
,
merged_grad
);
platform
::
ForRange
<
DeviceContext
>
for_range
(
dev_ctx
,
limit
);
const
int64_t
*
rows
;
#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
const
int64_t
*
rows
=
merged_grad
->
rows
().
Data
(
ctx
.
GetPlace
());
auto
&
merged_tensor
=
merged_grad
->
value
();
int64_t
row_count
=
merged_grad
->
rows
().
size
();
int64_t
row_numel
=
merged_tensor
.
numel
()
/
row_count
;
...
...
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