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
马赛克医院
Paddle
提交
29382db6
P
Paddle
项目概览
马赛克医院
/
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
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
29382db6
编写于
10月 18, 2018
作者:
D
dzhwinter
提交者:
GitHub
10月 18, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #13874 from dzhwinter/fix/momentum
add sparse update momentum. test=develop
上级
6a54c3de
00e8791f
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
448 addition
and
105 deletion
+448
-105
paddle/fluid/operators/momentum_op.cc
paddle/fluid/operators/momentum_op.cc
+40
-13
paddle/fluid/operators/momentum_op.cu
paddle/fluid/operators/momentum_op.cu
+3
-72
paddle/fluid/operators/momentum_op.h
paddle/fluid/operators/momentum_op.h
+311
-20
python/paddle/fluid/tests/unittests/test_momentum_op.py
python/paddle/fluid/tests/unittests/test_momentum_op.py
+94
-0
未找到文件。
paddle/fluid/operators/momentum_op.cc
浏览文件 @
29382db6
...
...
@@ -24,7 +24,7 @@ class MomentumOp : public framework::OperatorWithKernel {
using
framework
::
OperatorWithKernel
::
OperatorWithKernel
;
protected:
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
void
InferShape
(
framework
::
InferShapeContext
*
ctx
)
const
override
{
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Param"
),
"Input(param) of Momentum should not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Grad"
),
...
...
@@ -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"
);
...
...
@@ -58,13 +61,34 @@ class MomentumOp : public framework::OperatorWithKernel {
ctx
->
SetOutputDim
(
"VelocityOut"
,
param_dim
);
}
framework
::
OpKernelType
GetExpectedKernelType
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
input_data_type
=
framework
::
ToDataType
(
ctx
.
Input
<
Tensor
>
(
"Param"
)
->
type
());
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
input_data_type
=
framework
::
GetDataTypeOfVar
(
ctx
.
InputVar
(
"Param"
));
return
framework
::
OpKernelType
(
input_data_type
,
ctx
.
GetPlace
());
}
};
class
MomentumOpInferVarType
:
public
framework
::
VarTypeInference
{
public:
void
operator
()(
const
framework
::
OpDesc
&
op_desc
,
framework
::
BlockDesc
*
block
)
const
override
{
auto
input_var
=
op_desc
.
Input
(
"Param"
)[
0
];
for
(
auto
&
out_var
:
op_desc
.
Output
(
"ParamOut"
))
{
if
(
block
->
FindRecursiveOrCreateVar
(
input_var
).
GetType
()
==
framework
::
proto
::
VarType
::
SELECTED_ROWS
)
{
block
->
FindRecursiveOrCreateVar
(
out_var
).
SetType
(
framework
::
proto
::
VarType
::
SELECTED_ROWS
);
}
else
if
(
block
->
FindRecursiveOrCreateVar
(
input_var
).
GetType
()
==
framework
::
proto
::
VarType
::
LOD_TENSOR
)
{
block
->
FindRecursiveOrCreateVar
(
out_var
).
SetType
(
framework
::
proto
::
VarType
::
LOD_TENSOR
);
}
else
{
PADDLE_THROW
(
"Only support LodTensor and SelectedRows, Unexpected Input Type."
);
}
}
}
};
class
MomentumOpMaker
:
public
framework
::
OpProtoAndCheckerMaker
{
public:
void
Make
()
override
{
...
...
@@ -115,6 +139,9 @@ $$
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_WITHOUT_GRADIENT
(
momentum
,
ops
::
MomentumOp
,
ops
::
MomentumOpMaker
);
REGISTER_OP_CPU_KERNEL
(
momentum
,
ops
::
MomentumOpKernel
<
float
>
,
ops
::
MomentumOpKernel
<
double
>
);
REGISTER_OPERATOR
(
momentum
,
ops
::
MomentumOp
,
ops
::
MomentumOpMaker
,
paddle
::
framework
::
EmptyGradOpMaker
,
ops
::
MomentumOpInferVarType
);
REGISTER_OP_CPU_KERNEL
(
momentum
,
ops
::
MomentumOpKernel
<
paddle
::
platform
::
CPUDeviceContext
,
float
>
,
ops
::
MomentumOpKernel
<
paddle
::
platform
::
CPUDeviceContext
,
double
>
);
paddle/fluid/operators/momentum_op.cu
浏览文件 @
29382db6
...
...
@@ -15,76 +15,7 @@ limitations under the License. */
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/momentum_op.h"
namespace
paddle
{
namespace
operators
{
template
<
typename
T
>
__global__
void
MomentumKernel
(
const
T
*
p
,
const
T
*
g
,
const
T
*
v
,
const
T
*
learning_rate
,
const
T
mu
,
const
int64_t
num
,
bool
use_nesterov
,
T
*
p_out
,
T
*
v_out
)
{
T
lr
=
learning_rate
[
0
];
if
(
use_nesterov
)
{
for
(
int
i
=
blockIdx
.
x
*
blockDim
.
x
+
threadIdx
.
x
;
i
<
num
;
i
+=
blockDim
.
x
*
gridDim
.
x
)
{
T
g_val
=
g
[
i
];
T
v_new
=
v
[
i
]
*
mu
+
g_val
;
v_out
[
i
]
=
v_new
;
p_out
[
i
]
=
p
[
i
]
-
(
g_val
+
v_new
*
mu
)
*
lr
;
}
}
else
{
for
(
int
i
=
blockIdx
.
x
*
blockDim
.
x
+
threadIdx
.
x
;
i
<
num
;
i
+=
blockDim
.
x
*
gridDim
.
x
)
{
T
v_new
=
v
[
i
]
*
mu
+
g
[
i
];
v_out
[
i
]
=
v_new
;
p_out
[
i
]
=
p
[
i
]
-
lr
*
v_new
;
}
}
}
template
<
typename
T
>
class
MomentumOpCUDAKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
const
auto
*
param_var
=
ctx
.
InputVar
(
"Param"
);
PADDLE_ENFORCE
(
param_var
->
IsType
<
framework
::
LoDTensor
>
(),
"The Var(%s)'s type should be LoDTensor, "
"but the received is %s"
,
ctx
.
Inputs
(
"Param"
).
front
(),
param_var
->
Type
().
name
());
const
auto
*
grad_var
=
ctx
.
InputVar
(
"Grad"
);
PADDLE_ENFORCE
(
grad_var
->
IsType
<
framework
::
LoDTensor
>
(),
"The Var(%s)'s type should be LoDTensor, "
"but the received is %s"
,
ctx
.
Inputs
(
"Grad"
).
front
(),
grad_var
->
Type
().
name
());
auto
param_out
=
ctx
.
Output
<
framework
::
Tensor
>
(
"ParamOut"
);
auto
velocity_out
=
ctx
.
Output
<
framework
::
Tensor
>
(
"VelocityOut"
);
auto
param
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Param"
);
auto
velocity
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Velocity"
);
auto
grad
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Grad"
);
auto
learning_rate
=
ctx
.
Input
<
framework
::
Tensor
>
(
"LearningRate"
);
T
*
p_out
=
param_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
T
*
v_out
=
velocity_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
T
mu
=
static_cast
<
T
>
(
ctx
.
Attr
<
float
>
(
"mu"
));
bool
use_nesterov
=
ctx
.
Attr
<
bool
>
(
"use_nesterov"
);
auto
*
p
=
param
->
data
<
T
>
();
auto
*
v
=
velocity
->
data
<
T
>
();
auto
*
g
=
grad
->
data
<
T
>
();
auto
*
lr
=
learning_rate
->
data
<
T
>
();
int
block
=
512
;
int
grid
=
(
param
->
numel
()
+
block
-
1
)
/
block
;
MomentumKernel
<
T
><<<
grid
,
block
,
0
,
ctx
.
cuda_device_context
().
stream
()
>>>
(
p
,
g
,
v
,
lr
,
mu
,
param
->
numel
(),
use_nesterov
,
p_out
,
v_out
);
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_CUDA_KERNEL
(
momentum
,
ops
::
MomentumOpCUDAKernel
<
float
>
,
ops
::
MomentumOpCUDAKernel
<
double
>
);
REGISTER_OP_CUDA_KERNEL
(
momentum
,
ops
::
MomentumOpKernel
<
paddle
::
platform
::
CUDADeviceContext
,
float
>
,
ops
::
MomentumOpKernel
<
paddle
::
platform
::
CUDADeviceContext
,
double
>
);
paddle/fluid/operators/momentum_op.h
浏览文件 @
29382db6
...
...
@@ -13,35 +13,48 @@ 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"
#include "paddle/fluid/operators/math/selected_rows_functor.h"
#include "paddle/fluid/platform/for_range.h"
namespace
paddle
{
namespace
operators
{
template
<
typename
T
>
class
MomentumOpKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
const
auto
*
param_var
=
ctx
.
InputVar
(
"Param"
);
PADDLE_ENFORCE
(
param_var
->
IsType
<
framework
::
LoDTensor
>
(),
"The Var(%s)'s type should be LoDTensor, "
"but the received is %s"
,
ctx
.
Inputs
(
"Param"
).
front
(),
param_var
->
Type
().
name
());
auto
param_out
=
ctx
.
Output
<
framework
::
Tensor
>
(
"ParamOut"
);
auto
velocity_out
=
ctx
.
Output
<
framework
::
Tensor
>
(
"VelocityOut"
);
auto
param
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Param"
);
auto
velocity
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Velocity"
);
auto
grad
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Grad"
);
auto
learning_rate
=
ctx
.
Input
<
framework
::
Tensor
>
(
"LearningRate"
);
using
framework
::
Tensor
;
using
framework
::
SelectedRows
;
struct
NoNesterov
;
struct
UseNesterov
;
param_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
velocity_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
template
<
typename
T
>
class
CPUDenseMomentumFunctor
{
private:
const
Tensor
*
param
;
const
Tensor
*
grad
;
const
Tensor
*
velocity
;
const
Tensor
*
learning_rate
;
const
T
mu
;
const
T
use_nesterov
;
Tensor
*
param_out
;
Tensor
*
velocity_out
;
T
mu
=
static_cast
<
T
>
(
ctx
.
Attr
<
float
>
(
"mu"
));
bool
use_nesterov
=
ctx
.
Attr
<
bool
>
(
"use_nesterov"
);
public:
CPUDenseMomentumFunctor
(
const
Tensor
*
param
,
const
Tensor
*
grad
,
const
Tensor
*
velocity
,
const
Tensor
*
learning_rate
,
const
T
mu
,
const
bool
use_nesterov
,
Tensor
*
param_out
,
Tensor
*
velocity_out
)
:
param
(
param
),
grad
(
grad
),
velocity
(
velocity
),
learning_rate
(
learning_rate
),
mu
(
mu
),
use_nesterov
(
use_nesterov
),
param_out
(
param_out
),
velocity_out
(
velocity_out
)
{}
inline
void
operator
()()
{
auto
p_out
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
param_out
);
auto
v_out
=
framework
::
EigenVector
<
T
>::
Flatten
(
*
velocity_out
);
...
...
@@ -59,5 +72,283 @@ class MomentumOpKernel : public framework::OpKernel<T> {
}
};
template
<
typename
T
,
typename
UpdateMethod
>
class
DenseMomentumFunctor
;
// NOTE(dzh) for performance.
// avoid if/else in inside kernel, implement GPU UseNesterov/NoNesterov as two
// functor.
template
<
typename
T
>
class
DenseMomentumFunctor
<
T
,
UseNesterov
>
{
private:
const
T
*
p_
;
const
T
*
g_
;
const
T
*
v_
;
const
T
*
lr_
;
const
T
mu_
;
const
int64_t
num_
;
T
*
p_out_
;
T
*
v_out_
;
public:
DenseMomentumFunctor
(
const
T
*
p
,
const
T
*
g
,
const
T
*
v
,
const
T
*
learning_rate
,
const
T
mu
,
const
int64_t
num
,
T
*
p_out
,
T
*
v_out
)
:
p_
(
p
),
g_
(
g
),
v_
(
v
),
lr_
(
learning_rate
),
mu_
(
mu
),
num_
(
num
),
p_out_
(
p_out
),
v_out_
(
v_out
)
{}
inline
HOSTDEVICE
void
operator
()(
size_t
i
)
const
{
// put memory access in register
const
T
p
=
p_
[
i
];
const
T
g
=
g_
[
i
];
const
T
lr
=
lr_
[
0
];
const
T
v
=
v_
[
i
];
T
v_out
=
v
*
mu_
+
g
;
T
p_out
=
p
-
(
g
+
v_out
*
mu_
)
*
lr
;
// write reigster to memory
v_out_
[
i
]
=
v_out
;
p_out_
[
i
]
=
p_out
;
}
};
template
<
typename
T
>
class
DenseMomentumFunctor
<
T
,
NoNesterov
>
{
private:
const
T
*
p_
;
const
T
*
g_
;
const
T
*
v_
;
const
T
*
lr_
;
const
T
mu_
;
const
int64_t
num_
;
T
*
p_out_
;
T
*
v_out_
;
public:
DenseMomentumFunctor
(
const
T
*
p
,
const
T
*
g
,
const
T
*
v
,
const
T
*
learning_rate
,
const
T
mu
,
const
int64_t
num
,
T
*
p_out
,
T
*
v_out
)
:
p_
(
p
),
g_
(
g
),
v_
(
v
),
lr_
(
learning_rate
),
mu_
(
mu
),
num_
(
num
),
p_out_
(
p_out
),
v_out_
(
v_out
)
{}
inline
HOSTDEVICE
void
operator
()(
size_t
i
)
const
{
// put memory access in register
const
T
p
=
p_
[
i
];
const
T
g
=
g_
[
i
];
const
T
lr
=
lr_
[
0
];
const
T
v
=
v_
[
i
];
T
v_out
=
v
*
mu_
+
g
;
T
p_out
=
p
-
lr
*
v_out
;
// write reigster to memory
v_out_
[
i
]
=
v_out
;
p_out_
[
i
]
=
p_out
;
}
};
template
<
typename
T
,
typename
UpdateMethod
>
class
SparseMomentumFunctor
;
template
<
typename
T
>
class
SparseMomentumFunctor
<
T
,
UseNesterov
>
{
private:
const
T
*
p_
;
const
T
*
g_
;
const
T
*
v_
;
const
T
*
lr_
;
const
T
mu_
;
const
int64_t
*
rows_
;
const
int64_t
row_numel_
;
const
int64_t
row_height_
;
T
*
p_out_
;
T
*
v_out_
;
public:
SparseMomentumFunctor
(
const
T
*
p
,
const
T
*
g
,
const
T
*
v
,
const
T
*
lr
,
const
T
mu
,
const
int64_t
*
rows
,
int64_t
row_numel
,
int64_t
row_height
,
T
*
p_out
,
T
*
v_out
)
:
p_
(
p
),
g_
(
g
),
v_
(
v
),
lr_
(
lr
),
mu_
(
mu
),
rows_
(
rows
),
row_numel_
(
row_numel
),
row_height_
(
row_height
),
p_out_
(
p_out
),
v_out_
(
v_out
)
{}
inline
HOSTDEVICE
void
operator
()(
size_t
i
)
{
auto
row_idx
=
math
::
BinarySearch
<
int64_t
>
(
rows_
,
row_height_
,
i
/
row_numel_
);
T
g
=
row_idx
>=
0
?
g_
[
row_idx
*
row_numel_
+
i
%
row_numel_
]
:
0
;
// put memory access in register
const
T
p
=
p_
[
i
];
const
T
lr
=
lr_
[
0
];
const
T
v
=
v_
[
i
];
T
v_out
=
v
*
mu_
+
g
;
T
p_out
=
p
-
(
g
+
v_out
*
mu_
)
*
lr
;
// write reigster to memory
v_out_
[
i
]
=
v_out
;
p_out_
[
i
]
=
p_out
;
}
};
template
<
typename
T
>
class
SparseMomentumFunctor
<
T
,
NoNesterov
>
{
private:
const
T
*
p_
;
const
T
*
g_
;
const
T
*
v_
;
const
T
*
lr_
;
const
T
mu_
;
const
int64_t
*
rows_
;
const
int64_t
row_numel_
;
const
int64_t
row_height_
;
T
*
p_out_
;
T
*
v_out_
;
public:
SparseMomentumFunctor
(
const
T
*
p
,
const
T
*
g
,
const
T
*
v
,
const
T
*
lr
,
const
T
mu
,
const
int64_t
*
rows
,
int64_t
row_numel
,
int64_t
row_height
,
T
*
p_out
,
T
*
v_out
)
:
p_
(
p
),
g_
(
g
),
v_
(
v
),
lr_
(
lr
),
mu_
(
mu
),
rows_
(
rows
),
row_numel_
(
row_numel
),
row_height_
(
row_height
),
p_out_
(
p_out
),
v_out_
(
v_out
)
{}
inline
HOSTDEVICE
void
operator
()(
size_t
i
)
{
auto
row_idx
=
math
::
BinarySearch
<
int64_t
>
(
rows_
,
row_height_
,
i
/
row_numel_
);
T
g
=
row_idx
>=
0
?
g_
[
row_idx
*
row_numel_
+
i
%
row_numel_
]
:
0
;
// put memory access in register
const
T
p
=
p_
[
i
];
const
T
lr
=
lr_
[
0
];
const
T
v
=
v_
[
i
];
T
v_out
=
v
*
mu_
+
g
;
T
p_out
=
p
-
v_out
*
lr
;
// write reigster to memory
v_out_
[
i
]
=
v_out
;
p_out_
[
i
]
=
p_out
;
}
};
template
<
typename
DeviceContext
,
typename
T
>
class
MomentumOpKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
T
mu
=
static_cast
<
T
>
(
ctx
.
Attr
<
float
>
(
"mu"
));
bool
use_nesterov
=
ctx
.
Attr
<
bool
>
(
"use_nesterov"
);
auto
learning_rate
=
ctx
.
Input
<
framework
::
Tensor
>
(
"LearningRate"
);
auto
param
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Param"
);
auto
param_out
=
ctx
.
Output
<
framework
::
Tensor
>
(
"ParamOut"
);
auto
*
velocity
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Velocity"
);
auto
velocity_out
=
ctx
.
Output
<
framework
::
Tensor
>
(
"VelocityOut"
);
param_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
velocity_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
auto
*
grad_var
=
ctx
.
InputVar
(
"Grad"
);
if
(
grad_var
->
IsType
<
framework
::
LoDTensor
>
())
{
auto
grad
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Grad"
);
if
(
platform
::
is_cpu_place
(
ctx
.
GetPlace
()))
{
CPUDenseMomentumFunctor
<
T
>
functor
(
param
,
grad
,
velocity
,
learning_rate
,
mu
,
use_nesterov
,
param_out
,
velocity_out
);
functor
();
}
else
if
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()))
{
platform
::
ForRange
<
DeviceContext
>
for_range
(
static_cast
<
const
DeviceContext
&>
(
ctx
.
device_context
()),
param
->
numel
());
if
(
use_nesterov
)
{
DenseMomentumFunctor
<
T
,
UseNesterov
>
functor
(
param
->
data
<
T
>
(),
grad
->
data
<
T
>
(),
velocity
->
data
<
T
>
(),
learning_rate
->
data
<
T
>
(),
mu
,
param
->
numel
(),
param_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
()),
velocity_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
()));
for_range
(
functor
);
}
else
{
DenseMomentumFunctor
<
T
,
NoNesterov
>
functor
(
param
->
data
<
T
>
(),
grad
->
data
<
T
>
(),
velocity
->
data
<
T
>
(),
learning_rate
->
data
<
T
>
(),
mu
,
param
->
numel
(),
param_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
()),
velocity_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
()));
for_range
(
functor
);
}
}
}
else
if
(
grad_var
->
IsType
<
framework
::
SelectedRows
>
())
{
// sparse update embedding with selectedrows
auto
grad
=
ctx
.
Input
<
framework
::
SelectedRows
>
(
"Grad"
);
// sparse update maybe empty.
if
(
grad
->
rows
().
size
()
==
0
)
{
VLOG
(
3
)
<<
"Grad SelectedRows contains no data!"
;
return
;
}
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
);
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
,
row_numel
,
static_cast
<
int64_t
>
(
merged_grad
->
rows
().
size
()),
param_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
()),
velocity_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
()));
for_range
(
functor
);
}
else
{
SparseMomentumFunctor
<
T
,
NoNesterov
>
functor
(
param
->
data
<
T
>
(),
merged_grad
->
value
().
data
<
T
>
(),
velocity
->
data
<
T
>
(),
learning_rate
->
data
<
T
>
(),
mu
,
rows
,
row_numel
,
static_cast
<
int64_t
>
(
merged_grad
->
rows
().
size
()),
param_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
()),
velocity_out
->
mutable_data
<
T
>
(
ctx
.
GetPlace
()));
for_range
(
functor
);
}
}
else
{
PADDLE_THROW
(
string
::
Sprintf
(
"MomentumOp only supports LoDTensor or SelectedRows "
"gradient, but the received Variable Type is %s"
,
grad_var
->
Type
().
name
()));
}
}
};
}
// namespace operators
}
// namespace paddle
python/paddle/fluid/tests/unittests/test_momentum_op.py
浏览文件 @
29382db6
...
...
@@ -16,6 +16,8 @@ from __future__ import print_function
import
unittest
import
numpy
as
np
import
paddle.fluid.core
as
core
from
paddle.fluid.op
import
Operator
from
op_test
import
OpTest
...
...
@@ -88,5 +90,97 @@ class TestMomentumOp2(OpTest):
self
.
check_output
()
class
TestSparseMomentumOp
(
unittest
.
TestCase
):
def
setUp
(
self
):
self
.
use_nesterov
=
False
def
check_with_place
(
self
,
place
):
self
.
init_kernel
()
scope
=
core
.
Scope
()
# create and initialize Grad Variable
height
=
10
rows
=
[
0
,
4
,
7
]
row_numel
=
12
mu
=
1.0
use_nesterov
=
self
.
use_nesterov
# create and initialize Param Variable
param
=
scope
.
var
(
'Param'
).
get_tensor
()
param_array
=
np
.
full
((
height
,
row_numel
),
5.0
).
astype
(
"float32"
)
param
.
set
(
param_array
,
place
)
param_out
=
scope
.
var
(
"ParamOut"
).
get_tensor
()
param_out_array
=
np
.
full
((
height
,
row_numel
),
0.0
).
astype
(
"float32"
)
param_out
.
set
(
param_out_array
,
place
)
grad_selected_rows
=
scope
.
var
(
'Grad'
).
get_selected_rows
()
grad_selected_rows
.
set_height
(
height
)
grad_selected_rows
.
set_rows
(
rows
)
grad_np_array
=
np
.
ones
((
len
(
rows
),
row_numel
)).
astype
(
"float32"
)
grad_np_array
[
0
,
0
]
=
2.0
grad_np_array
[
2
,
8
]
=
4.0
grad_tensor
=
grad_selected_rows
.
get_tensor
()
grad_tensor
.
set
(
grad_np_array
,
place
)
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
.
set
(
velocity_out_np_array
,
place
)
# create and initialize LeraningRate Variable
lr
=
scope
.
var
(
'LearningRate'
).
get_tensor
()
lr_array
=
np
.
full
((
1
),
2.0
).
astype
(
"float32"
)
lr
.
set
(
lr_array
,
place
)
# create and run operator
op
=
Operator
(
"momentum"
,
Param
=
'Param'
,
Grad
=
'Grad'
,
Velocity
=
'Velocity'
,
ParamOut
=
'ParamOut'
,
VelocityOut
=
'VelocityOut'
,
LearningRate
=
'LearningRate'
,
mu
=
mu
,
use_nesterov
=
use_nesterov
)
op
.
run
(
scope
,
place
)
# get and compare result
param_out_np_array
=
np
.
array
(
param_out
)
velocity_out_np_array
=
np
.
array
(
velocity_out
)
# TODO(dzh): add a more suitable general numpy interface
# for sparse update.
_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
+
_velocity_out
*
mu
)
*
lr_array
else
:
_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
def
test_sparse_momentum
(
self
):
places
=
[
core
.
CPUPlace
()]
if
core
.
is_compiled_with_cuda
():
places
.
append
(
core
.
CUDAPlace
(
0
))
for
place
in
places
:
self
.
check_with_place
(
place
)
class
TestSparseMomentumOp2
(
TestSparseMomentumOp
):
def
init_kernel
(
self
):
self
.
use_nesterov
=
True
if
__name__
==
"__main__"
:
unittest
.
main
()
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