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
PaddlePaddle
Paddle
提交
82630408
P
Paddle
项目概览
PaddlePaddle
/
Paddle
12 个月 前同步成功
通知
2292
Star
20931
Fork
5422
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1423
列表
看板
标记
里程碑
合并请求
543
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
1,423
Issue
1,423
列表
看板
标记
里程碑
合并请求
543
合并请求
543
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
82630408
编写于
12月 22, 2020
作者:
W
whs
提交者:
GitHub
12月 22, 2020
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Support double backward rsqrt (#29589)
上级
b76f5a84
变更
4
隐藏空白更改
内联
并排
Showing
4 changed file
with
176 addition
and
1 deletion
+176
-1
paddle/fluid/operators/activation_op.cc
paddle/fluid/operators/activation_op.cc
+48
-0
paddle/fluid/operators/activation_op.cu
paddle/fluid/operators/activation_op.cu
+14
-0
paddle/fluid/operators/activation_op.h
paddle/fluid/operators/activation_op.h
+90
-1
python/paddle/fluid/tests/unittests/test_activation_nn_grad.py
...n/paddle/fluid/tests/unittests/test_activation_nn_grad.py
+24
-0
未找到文件。
paddle/fluid/operators/activation_op.cc
100755 → 100644
浏览文件 @
82630408
...
...
@@ -896,6 +896,25 @@ class SqrtDoubleGradMaker : public ::paddle::framework::SingleGradOpMaker<T> {
}
};
// rsqrt Grad: dx = -0.5 * dy * y * y * y
// rsqrt GradGrad: ddy = -0.5 * ddx * y * y * y, dy = (3/y) * ddx
template
<
typename
T
>
class
RsqrtDoubleGradMaker
:
public
::
paddle
::
framework
::
SingleGradOpMaker
<
T
>
{
public:
using
::
paddle
::
framework
::
SingleGradOpMaker
<
T
>::
SingleGradOpMaker
;
protected:
void
Apply
(
GradOpPtr
<
T
>
op
)
const
override
{
op
->
SetType
(
"rsqrt_grad_grad"
);
op
->
SetInput
(
"Out"
,
this
->
Input
(
"Out"
));
op
->
SetInput
(
"DX"
,
this
->
Output
(
framework
::
GradVarName
(
"X"
)));
op
->
SetInput
(
"DDX"
,
this
->
OutputGrad
(
framework
::
GradVarName
(
"X"
)));
op
->
SetAttrMap
(
this
->
Attrs
());
op
->
SetOutput
(
"DOut"
,
this
->
InputGrad
(
"Out"
));
op
->
SetOutput
(
"DDOut"
,
this
->
InputGrad
(
framework
::
GradVarName
(
"Out"
)));
}
};
// square Grad: dx=2x*dy
// square GradGrad: ddy=2x*ddx, dx=2dy*ddx
template
<
typename
T
>
...
...
@@ -1167,6 +1186,35 @@ REGISTER_OP_CPU_KERNEL(
ops
::
SqrtGradGradFunctor
<
plat
::
float16
>>
);
/* ========================================================================== */
/* =========================== rsqrt register =============================
*/
REGISTER_OPERATOR
(
rsqrt
,
ops
::
ActivationOp
,
ops
::
RsqrtOpMaker
,
ops
::
ActivationOpInferVarType
,
ops
::
ActivationGradOpMaker
<
ops
::
RsqrtGradFunctor
<
float
>::
FwdDeps
(),
paddle
::
framework
::
OpDesc
>
,
ops
::
ActivationGradOpMaker
<
ops
::
RsqrtGradFunctor
<
float
>::
FwdDeps
(),
paddle
::
imperative
::
OpBase
>
,
ops
::
ActFwdInplaceInferer
);
REGISTER_OPERATOR
(
rsqrt_grad
,
ops
::
ActivationOpGrad
,
ops
::
ActivationGradOpInplaceInferer
,
ops
::
RsqrtDoubleGradMaker
<
paddle
::
framework
::
OpDesc
>
,
ops
::
RsqrtDoubleGradMaker
<
paddle
::
imperative
::
OpBase
>
);
REGISTER_OPERATOR
(
rsqrt_grad_grad
,
ops
::
ActivationOpDoubleGrad
<
ops
::
RsqrtGradGradFunctor
<
float
>::
FwdDeps
()
>
,
ops
::
ActivationDoubleGradOpInplaceInferer
);
REGISTER_ACTIVATION_CPU_KERNEL
(
rsqrt
,
Rsqrt
,
RsqrtFunctor
,
RsqrtGradFunctor
);
REGISTER_OP_CPU_KERNEL
(
rsqrt_grad_grad
,
ops
::
RsqrtDoubleGradKernel
<
plat
::
CPUDeviceContext
,
ops
::
RsqrtGradGradFunctor
<
float
>>
,
ops
::
RsqrtDoubleGradKernel
<
plat
::
CPUDeviceContext
,
ops
::
RsqrtGradGradFunctor
<
double
>>
,
ops
::
RsqrtDoubleGradKernel
<
plat
::
CPUDeviceContext
,
ops
::
RsqrtGradGradFunctor
<
plat
::
float16
>>
);
/* ========================================================================== */
/* ========================== square register ============================ */
REGISTER_OPERATOR
(
square
,
ops
::
ActivationOp
,
ops
::
SquareOpMaker
,
...
...
paddle/fluid/operators/activation_op.cu
浏览文件 @
82630408
...
...
@@ -85,6 +85,20 @@ REGISTER_OP_CUDA_KERNEL(
ops
::
SqrtGradGradFunctor
<
plat
::
float16
>>
);
/* ========================================================================== */
/* =========================== rsqrt register =============================
*/
REGISTER_ACTIVATION_CUDA_KERNEL
(
rsqrt
,
Rsqrt
,
RsqrtFunctor
,
RsqrtGradFunctor
);
REGISTER_OP_CUDA_KERNEL
(
rsqrt_grad_grad
,
ops
::
RsqrtDoubleGradKernel
<
paddle
::
platform
::
CUDADeviceContext
,
ops
::
RsqrtGradGradFunctor
<
float
>>
,
ops
::
RsqrtDoubleGradKernel
<
paddle
::
platform
::
CUDADeviceContext
,
ops
::
RsqrtGradGradFunctor
<
double
>>
,
ops
::
RsqrtDoubleGradKernel
<
paddle
::
platform
::
CUDADeviceContext
,
ops
::
RsqrtGradGradFunctor
<
plat
::
float16
>>
);
/* ========================================================================== */
/* =========================== square register ============================ */
REGISTER_OP_CUDA_KERNEL
(
square
,
...
...
paddle/fluid/operators/activation_op.h
100755 → 100644
浏览文件 @
82630408
...
...
@@ -1643,6 +1643,35 @@ struct SqrtGradGradFunctor : public BaseActivationFunctor<T> {
static
constexpr
ActBwdOpFwdDeps
FwdDeps
()
{
return
kDepOut
;
}
};
template
<
typename
T
>
struct
RsqrtGradGradFunctor
:
public
BaseActivationFunctor
<
T
>
{
template
<
typename
Device
>
void
operator
()(
const
Device
&
dev
,
const
framework
::
Tensor
*
Out
,
const
framework
::
Tensor
*
ddX
,
framework
::
Tensor
*
ddOut
,
framework
::
Tensor
*
dOut
,
const
framework
::
Tensor
*
dX
)
const
{
auto
*
d
=
dev
.
eigen_device
();
auto
ddx
=
framework
::
EigenVector
<
T
>::
Flatten
(
GET_DATA_SAFELY
(
ddX
,
"Input"
,
"DDX"
,
"RsqrtGradGrad"
));
auto
out
=
framework
::
EigenVector
<
T
>::
Flatten
(
GET_DATA_SAFELY
(
Out
,
"Output"
,
"Out"
,
"RsqrtGradGrad"
));
// rsqrt GradGrad: ddy = -0.5 * ddx * y * y * y, dy = (3/y) * dx * ddx
if
(
dOut
)
{
auto
dx
=
framework
::
EigenVector
<
T
>::
Flatten
(
GET_DATA_SAFELY
(
dX
,
"Output"
,
"DX"
,
"RsqrtGradGrad"
));
auto
dout
=
framework
::
EigenVector
<
T
>::
Flatten
(
GET_DATA_SAFELY
(
dOut
,
"Output"
,
"DOut"
,
"RsqrtGradGrad"
));
dout
.
device
(
*
d
)
=
(
static_cast
<
T
>
(
3.0
)
/
out
)
*
dx
*
ddx
;
}
if
(
ddOut
)
{
auto
ddout
=
framework
::
EigenVector
<
T
>::
Flatten
(
GET_DATA_SAFELY
(
ddOut
,
"Output"
,
"DDOut"
,
"RsqrtGradGrad"
));
ddout
.
device
(
*
d
)
=
ddx
*
static_cast
<
T
>
(
-
0.5
)
*
out
*
out
*
out
;
}
}
static
constexpr
ActBwdOpFwdDeps
FwdDeps
()
{
return
kDepOut
;
}
};
template
<
typename
T
>
struct
SquareGradGradFunctor
:
public
BaseActivationFunctor
<
T
>
{
template
<
typename
Device
>
...
...
@@ -1828,6 +1857,67 @@ class SqrtDoubleGradKernel
}
};
// rsqrt Grad: dx = -0.5 * dy * y * y * y
// rsqrt GradGrad: ddy = -0.5 * ddx * y * y * y, dy = (3 / y) * dx * ddx
template
<
typename
DeviceContext
,
typename
Functor
>
class
RsqrtDoubleGradKernel
:
public
framework
::
OpKernel
<
typename
Functor
::
ELEMENT_TYPE
>
{
public:
using
T
=
typename
Functor
::
ELEMENT_TYPE
;
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
const
framework
::
Tensor
*
Out
,
*
dX
,
*
ddX
;
Out
=
dX
=
ddX
=
nullptr
;
framework
::
Tensor
*
ddOut
,
*
dOut
;
ddOut
=
dOut
=
nullptr
;
// extract ddx(input), ddout(output)
auto
ddx_var
=
ctx
.
InputVar
(
"DDX"
);
auto
ddo_var
=
ctx
.
OutputVar
(
"DDOut"
);
PADDLE_ENFORCE_NOT_NULL
(
ddx_var
,
platform
::
errors
::
NotFound
(
"Cannot get input Variable DDX, variable name = %s"
,
ctx
.
InputName
(
"DDX"
)));
ddX
=
ctx
.
Input
<
framework
::
Tensor
>
(
"DDX"
);
if
(
ddo_var
)
{
ddOut
=
ctx
.
Output
<
framework
::
Tensor
>
(
"DDOut"
);
}
PADDLE_ENFORCE_NOT_NULL
(
ddX
,
platform
::
errors
::
NotFound
(
"Cannot get input Variable DDX, variable name = %s"
,
ctx
.
InputName
(
"DDX"
)));
// extract out(input), dout(output)
auto
out_var
=
ctx
.
InputVar
(
"Out"
);
PADDLE_ENFORCE_NOT_NULL
(
out_var
,
platform
::
errors
::
NotFound
(
"Cannot get input Variable Out, variable name = %s"
,
ctx
.
InputName
(
"Out"
)));
auto
dout_var
=
ctx
.
OutputVar
(
"DOut"
);
Out
=
ctx
.
Input
<
framework
::
Tensor
>
(
"Out"
);
if
(
dout_var
)
{
dOut
=
ctx
.
Output
<
framework
::
Tensor
>
(
"DOut"
);
}
// extract dx(input)
auto
dx_var
=
ctx
.
InputVar
(
"DX"
);
PADDLE_ENFORCE_NOT_NULL
(
dx_var
,
platform
::
errors
::
NotFound
(
"Cannot get input Variable DX, variable name = %s"
,
ctx
.
InputName
(
"DX"
)));
if
(
dx_var
)
{
dX
=
ctx
.
Input
<
framework
::
Tensor
>
(
"DX"
);
}
if
(
dOut
)
dOut
->
mutable_data
<
T
>
(
Out
->
dims
(),
ctx
.
GetPlace
());
if
(
ddOut
)
ddOut
->
mutable_data
<
T
>
(
Out
->
dims
(),
ctx
.
GetPlace
());
auto
&
place
=
ctx
.
template
device_context
<
DeviceContext
>();
Functor
functor
;
functor
(
place
,
Out
,
ddX
,
ddOut
,
dOut
,
dX
);
}
};
template
<
typename
DeviceContext
,
typename
Functor
>
class
PowKernel
:
public
framework
::
OpKernel
<
typename
Functor
::
ELEMENT_TYPE
>
{
public:
...
...
@@ -1971,7 +2061,6 @@ struct LogGradGradFunctor : public BaseActivationFunctor<T> {
__macro(tanh, Tanh, TanhFunctor, TanhGradFunctor); \
__macro(atan, Atan, AtanFunctor, AtanGradFunctor); \
__macro(softshrink, SoftShrink, SoftShrinkFunctor, SoftShrinkGradFunctor); \
__macro(rsqrt, Rsqrt, RsqrtFunctor, RsqrtGradFunctor); \
__macro(ceil, Ceil, CeilFunctor, ZeroGradFunctor); \
__macro(floor, Floor, FloorFunctor, ZeroGradFunctor); \
__macro(cos, Cos, CosFunctor, CosGradFunctor); \
...
...
python/paddle/fluid/tests/unittests/test_activation_nn_grad.py
浏览文件 @
82630408
...
...
@@ -125,6 +125,30 @@ class TestSqrtDoubleGradCheck(unittest.TestCase):
self
.
func
(
p
)
class
TestRsqrtDoubleGradCheck
(
unittest
.
TestCase
):
@
prog_scope
()
def
func
(
self
,
place
):
shape
=
[
2
,
3
,
7
,
9
]
eps
=
0.0001
dtype
=
np
.
float64
x
=
layers
.
data
(
'x'
,
shape
,
False
,
dtype
)
x
.
persistable
=
True
y
=
layers
.
rsqrt
(
x
)
x_arr
=
np
.
random
.
uniform
(
0.1
,
1
,
shape
).
astype
(
dtype
)
gradient_checker
.
double_grad_check
(
[
x
],
y
,
x_init
=
x_arr
,
place
=
place
,
eps
=
eps
)
def
test_grad
(
self
):
places
=
[
fluid
.
CPUPlace
()]
if
core
.
is_compiled_with_cuda
():
places
=
[
fluid
.
CUDAPlace
(
0
)]
for
p
in
places
:
self
.
func
(
p
)
class
TestSquareDoubleGradCheck
(
unittest
.
TestCase
):
@
prog_scope
()
def
func
(
self
,
place
):
...
...
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