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
d6b6a337
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 >>
提交
d6b6a337
编写于
7月 19, 2019
作者:
A
Adam
提交者:
Tao Luo
7月 19, 2019
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Add LeakyRelu MKLDNN support (#18656)
test=develop
上级
0b9acb49
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
68 addition
and
14 deletion
+68
-14
paddle/fluid/operators/activation_op.cc
paddle/fluid/operators/activation_op.cc
+9
-0
paddle/fluid/operators/activation_op.h
paddle/fluid/operators/activation_op.h
+1
-1
paddle/fluid/operators/mkldnn/activation_mkldnn_op.cc
paddle/fluid/operators/mkldnn/activation_mkldnn_op.cc
+17
-12
python/paddle/fluid/tests/unittests/mkldnn/test_activation_mkldnn_op.py
...fluid/tests/unittests/mkldnn/test_activation_mkldnn_op.py
+22
-1
python/paddle/fluid/tests/unittests/test_activation_op.py
python/paddle/fluid/tests/unittests/test_activation_op.py
+19
-0
未找到文件。
paddle/fluid/operators/activation_op.cc
浏览文件 @
d6b6a337
...
...
@@ -363,6 +363,13 @@ class LeakyReluOpMaker : public framework::OpProtoAndCheckerMaker {
AddInput
(
"X"
,
"Input of LeakyRelu operator"
);
AddOutput
(
"Out"
,
"Output of LeakyRelu operator"
);
AddAttr
<
float
>
(
"alpha"
,
"The small negative slope"
).
SetDefault
(
0.02
f
);
AddAttr
<
bool
>
(
"use_mkldnn"
,
"(bool, default false) Only used in mkldnn kernel"
)
.
SetDefault
(
false
);
AddAttr
<
bool
>
(
"is_test"
,
"(bool, default false) Set to true for inference only, false "
"for training. Some layers may run faster when this is true."
)
.
SetDefault
(
false
);
AddComment
(
R"DOC(
LeakyRelu Activation Operator.
...
...
@@ -695,6 +702,8 @@ class LeakyReluDoubleGradMaker
op
->
SetType
(
"leaky_relu_grad_grad"
);
// input1: X
op
->
SetInput
(
"X"
,
Input
(
"X"
));
// input2: Out
op
->
SetInput
(
"Out"
,
Input
(
"Out"
));
// X@GRAD@GRAD: ddx
op
->
SetInput
(
"DDX"
,
OutputGrad
(
framework
::
GradVarName
(
"X"
)));
op
->
SetAttrMap
(
Attrs
());
...
...
paddle/fluid/operators/activation_op.h
浏览文件 @
d6b6a337
...
...
@@ -1001,7 +1001,7 @@ struct LeakyReluGradFunctor : public BaseActivationFunctor<T> {
dx
.
device
(
d
)
=
dout
*
(
temp1
+
temp2
).
template
cast
<
T
>();
}
static
constexpr
ActBwdOpFwdDeps
FwdDeps
()
{
return
kDepX
;
}
static
constexpr
ActBwdOpFwdDeps
FwdDeps
()
{
return
kDepX
Out
;
}
};
template
<
typename
T
>
...
...
paddle/fluid/operators/mkldnn/activation_mkldnn_op.cc
浏览文件 @
d6b6a337
...
...
@@ -77,8 +77,7 @@ class MKLDNNActivationGradKernel
template
<
typename
T
>
void
eltwise_forward
(
const
framework
::
ExecutionContext
&
ctx
,
mkldnn
::
algorithm
algorithm
,
const
T
alpha
=
0
,
const
T
beta
=
0
)
{
mkldnn
::
algorithm
algorithm
)
{
PADDLE_ENFORCE
(
paddle
::
platform
::
is_cpu_place
(
ctx
.
GetPlace
()),
"It must use CPUPlace."
);
auto
&
dev_ctx
=
ctx
.
template
device_context
<
MKLDNNDeviceContext
>();
...
...
@@ -90,6 +89,9 @@ void eltwise_forward(const framework::ExecutionContext &ctx,
const
T
*
x_data
=
x
->
data
<
T
>
();
T
*
y_data
=
y
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
const
T
alpha
=
ctx
.
op
().
HasAttr
(
"alpha"
)
?
ctx
.
Attr
<
T
>
(
"alpha"
)
:
0
;
const
T
beta
=
ctx
.
op
().
HasAttr
(
"beta"
)
?
ctx
.
Attr
<
T
>
(
"beta"
)
:
0
;
PADDLE_ENFORCE
(
x
->
dims
().
size
()
==
2
||
x
->
dims
().
size
()
==
3
||
x
->
dims
().
size
()
==
4
,
"Input dim must be with 2, 3 or 4"
);
...
...
@@ -101,10 +103,9 @@ void eltwise_forward(const framework::ExecutionContext &ctx,
bool
is_test
=
ctx
.
Attr
<
bool
>
(
"is_test"
);
// TODO(jczaja): When adding leaky-relu , swish , elu make sure to extend key
// with alpha, beta
std
::
string
key
=
platform
::
MKLDNNHandler
::
GetHash
(
src_tz
,
std
::
to_string
(
algorithm
)
+
ctx
.
op
().
Output
(
"Out"
));
src_tz
,
std
::
to_string
(
algorithm
)
+
std
::
to_string
(
alpha
)
+
std
::
to_string
(
beta
)
+
ctx
.
op
().
Output
(
"Out"
));
// TODO(jczaja): Make it Thread safe
// save input data and layout to be referred in backward path
...
...
@@ -153,8 +154,7 @@ void eltwise_forward(const framework::ExecutionContext &ctx,
template
<
typename
T
>
void
eltwise_grad
(
const
framework
::
ExecutionContext
&
ctx
,
mkldnn
::
algorithm
algorithm
,
const
T
alpha
=
0
,
const
T
beta
=
0
)
{
mkldnn
::
algorithm
algorithm
)
{
auto
&
dev_ctx
=
ctx
.
template
device_context
<
MKLDNNDeviceContext
>();
const
auto
&
mkldnn_engine
=
dev_ctx
.
GetEngine
();
...
...
@@ -164,6 +164,9 @@ void eltwise_grad(const framework::ExecutionContext &ctx,
const
T
*
diff_y_data
=
diff_y
->
data
<
T
>
();
T
*
diff_x_data
=
diff_x
->
mutable_data
<
T
>
(
ctx
.
GetPlace
());
const
T
alpha
=
ctx
.
op
().
HasAttr
(
"alpha"
)
?
ctx
.
Attr
<
T
>
(
"alpha"
)
:
0
;
const
T
beta
=
ctx
.
op
().
HasAttr
(
"beta"
)
?
ctx
.
Attr
<
T
>
(
"beta"
)
:
0
;
std
::
vector
<
int
>
diff_dst_tz
=
framework
::
vectorize2int
(
diff_y
->
dims
());
auto
diff_y_format
=
...
...
@@ -173,7 +176,8 @@ void eltwise_grad(const framework::ExecutionContext &ctx,
diff_dst_tz
,
platform
::
MKLDNNGetDataType
<
T
>
(),
diff_y_format
);
std
::
string
key
=
platform
::
MKLDNNHandler
::
GetHash
(
diff_dst_tz
,
std
::
to_string
(
algorithm
)
+
ctx
.
op
().
Input
(
"Out"
));
diff_dst_tz
,
std
::
to_string
(
algorithm
)
+
std
::
to_string
(
alpha
)
+
std
::
to_string
(
beta
)
+
ctx
.
op
().
Input
(
"Out"
));
const
std
::
string
key_src_data
=
key
+
"@eltwise_fwd_src_data"
;
const
std
::
string
key_src_layout
=
key
+
"@eltwise_fwd_src_layout"
;
...
...
@@ -273,10 +277,11 @@ namespace ops = paddle::operators;
act_type##_grad, MKLDNN, ::paddle::platform::CPUPlace, \
ops::MKLDNNActivationGradKernel<ops::grad_functor<float>>);
#define FOR_EACH_MKLDNN_KERNEL_FUNCTOR(__macro) \
__macro(relu, ReluMKLDNNFunctor, ReluMKLDNNGradFunctor); \
__macro(tanh, TanhMKLDNNFunctor, TanhMKLDNNGradFunctor); \
__macro(sqrt, SqrtMKLDNNFunctor, SqrtMKLDNNGradFunctor); \
#define FOR_EACH_MKLDNN_KERNEL_FUNCTOR(__macro) \
__macro(relu, ReluMKLDNNFunctor, ReluMKLDNNGradFunctor); \
__macro(leaky_relu, ReluMKLDNNFunctor, ReluMKLDNNGradFunctor); \
__macro(tanh, TanhMKLDNNFunctor, TanhMKLDNNGradFunctor); \
__macro(sqrt, SqrtMKLDNNFunctor, SqrtMKLDNNGradFunctor); \
__macro(abs, AbsMKLDNNFunctor, AbsMKLDNNGradFunctor);
FOR_EACH_MKLDNN_KERNEL_FUNCTOR
(
REGISTER_ACTIVATION_MKLDNN_KERNEL
);
python/paddle/fluid/tests/unittests/mkldnn/test_activation_mkldnn_op.py
浏览文件 @
d6b6a337
...
...
@@ -18,7 +18,7 @@ import unittest
import
numpy
as
np
import
paddle.fluid.core
as
core
from
paddle.fluid.tests.unittests.op_test
import
OpTest
from
paddle.fluid.tests.unittests.test_activation_op
import
TestRelu
,
TestTanh
,
TestSqrt
,
TestAbs
from
paddle.fluid.tests.unittests.test_activation_op
import
TestRelu
,
TestTanh
,
TestSqrt
,
TestAbs
,
TestLeakyRelu
from
mkldnn_op_test
import
check_if_mkldnn_primitives_exist_in_bwd
...
...
@@ -29,6 +29,13 @@ class TestMKLDNNReluDim2(TestRelu):
self
.
attrs
=
{
"use_mkldnn"
:
True
}
class
TestMKLDNNLeakyReluDim2
(
TestLeakyRelu
):
def
setUp
(
self
):
super
(
TestMKLDNNLeakyReluDim2
,
self
).
setUp
()
self
.
attrs
=
{
"use_mkldnn"
:
True
}
class
TestMKLDNNTanhDim2
(
TestTanh
):
def
setUp
(
self
):
super
(
TestMKLDNNTanhDim2
,
self
).
setUp
()
...
...
@@ -63,6 +70,20 @@ class TestMKLDNNReluDim4(TestRelu):
self
.
attrs
=
{
"use_mkldnn"
:
True
}
class
TestMKLDNNLeakyReluDim4
(
TestLeakyRelu
):
def
setUp
(
self
):
super
(
TestMKLDNNLeakyReluDim4
,
self
).
setUp
()
x
=
np
.
random
.
uniform
(
-
1
,
1
,
[
2
,
4
,
3
,
5
]).
astype
(
"float32"
)
# The same reason with TestAbs
x
[
np
.
abs
(
x
)
<
0.005
]
=
0.02
out
=
np
.
maximum
(
x
,
0.02
*
x
)
self
.
inputs
=
{
'X'
:
OpTest
.
np_dtype_to_fluid_dtype
(
x
)}
self
.
outputs
=
{
'Out'
:
out
}
self
.
attrs
=
{
"use_mkldnn"
:
True
}
class
TestMKLDNNTanhDim4
(
TestTanh
):
def
setUp
(
self
):
super
(
TestMKLDNNTanhDim4
,
self
).
setUp
()
...
...
python/paddle/fluid/tests/unittests/test_activation_op.py
浏览文件 @
d6b6a337
...
...
@@ -367,6 +367,25 @@ class TestRelu(TestActivation):
self
.
check_grad
([
'X'
],
'Out'
,
max_relative_error
=
0.007
)
class
TestLeakyRelu
(
TestActivation
):
def
setUp
(
self
):
self
.
op_type
=
"leaky_relu"
self
.
init_dtype
()
x
=
np
.
random
.
uniform
(
-
1
,
1
,
[
11
,
17
]).
astype
(
self
.
dtype
)
# The same reason with TestAbs
x
[
np
.
abs
(
x
)
<
0.005
]
=
0.02
out
=
np
.
maximum
(
x
,
0.02
*
x
)
self
.
inputs
=
{
'X'
:
OpTest
.
np_dtype_to_fluid_dtype
(
x
)}
self
.
outputs
=
{
'Out'
:
out
}
def
test_check_grad
(
self
):
if
self
.
dtype
==
np
.
float16
:
return
self
.
check_grad
([
'X'
],
'Out'
,
max_relative_error
=
0.007
)
class
TestGelu
(
TestActivation
):
def
setUp
(
self
):
self
.
op_type
=
"gelu"
...
...
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