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
wmsofts
Paddle
提交
3e636ec9
P
Paddle
项目概览
wmsofts
/
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
前往新版Gitcode,体验更适合开发者的 AI 搜索 >>
未验证
提交
3e636ec9
编写于
3月 15, 2023
作者:
Z
zhangyuqin1998
提交者:
GitHub
3月 15, 2023
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Delete hardswish_raw op (#51634)
* Delete hardswish_raw op * fix ut
上级
bcec0dce
变更
14
隐藏空白更改
内联
并排
Showing
14 changed file
with
45 addition
and
87 deletion
+45
-87
paddle/phi/api/yaml/legacy_backward.yaml
paddle/phi/api/yaml/legacy_backward.yaml
+1
-1
paddle/phi/kernels/activation_grad_kernel.h
paddle/phi/kernels/activation_grad_kernel.h
+0
-3
paddle/phi/kernels/activation_kernel.cc
paddle/phi/kernels/activation_kernel.cc
+0
-25
paddle/phi/kernels/activation_kernel.h
paddle/phi/kernels/activation_kernel.h
+0
-8
paddle/phi/kernels/cpu/activation_grad_kernel.cc
paddle/phi/kernels/cpu/activation_grad_kernel.cc
+3
-3
paddle/phi/kernels/cpu/activation_kernel.cc
paddle/phi/kernels/cpu/activation_kernel.cc
+7
-7
paddle/phi/kernels/gpu/activation_grad_kernel.cu
paddle/phi/kernels/gpu/activation_grad_kernel.cu
+3
-3
paddle/phi/kernels/gpu/activation_kernel.cu
paddle/phi/kernels/gpu/activation_kernel.cu
+7
-7
paddle/phi/kernels/onednn/activation_grad_kernel.cc
paddle/phi/kernels/onednn/activation_grad_kernel.cc
+0
-3
paddle/phi/kernels/onednn/activation_kernel.cc
paddle/phi/kernels/onednn/activation_kernel.cc
+5
-8
paddle/phi/kernels/xpu/activation_grad_kernel.cc
paddle/phi/kernels/xpu/activation_grad_kernel.cc
+3
-3
paddle/phi/kernels/xpu/activation_kernel.cc
paddle/phi/kernels/xpu/activation_kernel.cc
+7
-7
paddle/phi/ops/compat/activation_sig.cc
paddle/phi/ops/compat/activation_sig.cc
+6
-6
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_hard_swish.py
...sts/unittests/ir/inference/test_trt_convert_hard_swish.py
+3
-3
未找到文件。
paddle/phi/api/yaml/legacy_backward.yaml
浏览文件 @
3e636ec9
...
...
@@ -532,7 +532,7 @@
-
backward_op
:
hardswish_grad
forward
:
hardswish (Tensor x) -> Tensor(out)
args
:
(Tensor x, Tensor out_grad
, float threshold = 6.0, float scale = 6.0, float offset = 3.0
)
args
:
(Tensor x, Tensor out_grad)
output
:
Tensor(x_grad)
infer_meta
:
func
:
UnchangedInferMeta
...
...
paddle/phi/kernels/activation_grad_kernel.h
浏览文件 @
3e636ec9
...
...
@@ -225,9 +225,6 @@ template <typename T, typename Context>
void
HardSwishGradKernel
(
const
Context
&
dev_ctx
,
const
DenseTensor
&
x
,
const
DenseTensor
&
dout
,
float
threshold
,
float
scale
,
float
offset
,
DenseTensor
*
dx
);
template
<
typename
T
,
typename
Context
>
...
...
paddle/phi/kernels/activation_kernel.cc
浏览文件 @
3e636ec9
...
...
@@ -19,13 +19,6 @@
namespace
phi
{
template
<
typename
T
,
typename
Context
>
void
HardSwishKernel
(
const
Context
&
dev_ctx
,
const
DenseTensor
&
x
,
DenseTensor
*
out
)
{
HardSwishRawKernel
<
T
,
Context
>
(
dev_ctx
,
x
,
6
,
6
,
3
,
out
);
}
template
<
typename
T
,
typename
Context
>
void
Relu6Kernel
(
const
Context
&
dev_ctx
,
const
DenseTensor
&
x
,
...
...
@@ -44,21 +37,10 @@ void SwishKernel(const Context& dev_ctx,
using
complex64
=
::
phi
::
dtype
::
complex
<
float
>
;
using
complex128
=
::
phi
::
dtype
::
complex
<
double
>
;
PD_REGISTER_KERNEL
(
hardswish
,
CPU
,
ALL_LAYOUT
,
phi
::
HardSwishKernel
,
float
,
double
)
{}
PD_REGISTER_KERNEL
(
relu6
,
CPU
,
ALL_LAYOUT
,
phi
::
Relu6Kernel
,
float
,
double
)
{}
PD_REGISTER_KERNEL
(
swish
,
CPU
,
ALL_LAYOUT
,
phi
::
SwishKernel
,
float
,
double
)
{}
#if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
PD_REGISTER_KERNEL
(
hardswish
,
GPU
,
ALL_LAYOUT
,
phi
::
HardSwishKernel
,
float
,
double
,
phi
::
dtype
::
float16
,
phi
::
dtype
::
bfloat16
)
{}
PD_REGISTER_KERNEL
(
relu6
,
GPU
,
ALL_LAYOUT
,
...
...
@@ -80,18 +62,11 @@ PD_REGISTER_KERNEL(swish,
#endif
#if defined PADDLE_WITH_XPU
PD_REGISTER_KERNEL
(
hardswish
,
XPU
,
ALL_LAYOUT
,
phi
::
HardSwishKernel
,
float
)
{}
PD_REGISTER_KERNEL
(
relu6
,
XPU
,
ALL_LAYOUT
,
phi
::
Relu6Kernel
,
float
)
{}
PD_REGISTER_KERNEL
(
swish
,
XPU
,
ALL_LAYOUT
,
phi
::
SwishKernel
,
float
)
{}
#endif
#ifdef PADDLE_WITH_MKLDNN
PD_REGISTER_KERNEL
(
hardswish
,
OneDNN
,
ONEDNN
,
phi
::
HardSwishKernel
,
float
,
phi
::
dtype
::
bfloat16
)
{}
PD_REGISTER_KERNEL
(
relu6
,
OneDNN
,
ONEDNN
,
phi
::
Relu6Kernel
,
float
,
phi
::
dtype
::
bfloat16
)
{}
PD_REGISTER_KERNEL
(
...
...
paddle/phi/kernels/activation_kernel.h
浏览文件 @
3e636ec9
...
...
@@ -90,14 +90,6 @@ DECLARE_ACTIVATION_KERNEL_WITH_TWO_ATTRS(STanh, scale_a, scale_b)
DECLARE_ACTIVATION_KERNEL_WITH_TWO_ATTRS
(
Softplus
,
beta
,
threshold
)
DECLARE_ACTIVATION_KERNEL_WITH_TWO_ATTRS
(
HardSigmoid
,
slope
,
offset
)
template
<
typename
T
,
typename
Context
>
void
HardSwishRawKernel
(
const
Context
&
dev_ctx
,
const
DenseTensor
&
x
,
float
threshold
,
float
scale
,
float
offset
,
DenseTensor
*
out
);
template
<
typename
T
,
typename
Context
>
void
HardSwishKernel
(
const
Context
&
dev_ctx
,
const
DenseTensor
&
x
,
...
...
paddle/phi/kernels/cpu/activation_grad_kernel.cc
浏览文件 @
3e636ec9
...
...
@@ -226,11 +226,11 @@ template <typename T, typename Context>
void
HardSwishGradKernel
(
const
Context
&
dev_ctx
,
const
DenseTensor
&
x
,
const
DenseTensor
&
dout
,
float
threshold
,
float
scale
,
float
offset
,
DenseTensor
*
dx
)
{
funcs
::
HardSwishGradFunctor
<
T
>
functor
;
float
threshold
=
6
;
float
scale
=
6
;
float
offset
=
3
;
auto
attrs
=
functor
.
GetAttrs
();
*
(
attrs
[
0
].
second
)
=
threshold
;
*
(
attrs
[
1
].
second
)
=
scale
;
...
...
paddle/phi/kernels/cpu/activation_kernel.cc
浏览文件 @
3e636ec9
...
...
@@ -113,13 +113,13 @@ DEFINE_CPU_ACT_KERNEL_WITH_TWO_ATTRS(HardSigmoid,
offset
)
template
<
typename
T
,
typename
Context
>
void
HardSwishRawKernel
(
const
Context
&
dev_ctx
,
const
DenseTensor
&
x
,
float
threshold
,
float
scale
,
float
offset
,
DenseTensor
*
out
)
{
void
HardSwishKernel
(
const
Context
&
dev_ctx
,
const
DenseTensor
&
x
,
DenseTensor
*
out
)
{
funcs
::
HardSwishFunctor
<
T
>
functor
;
float
threshold
=
6
;
float
scale
=
6
;
float
offset
=
3
;
auto
attrs
=
functor
.
GetAttrs
();
*
(
attrs
[
0
].
second
)
=
threshold
;
*
(
attrs
[
1
].
second
)
=
scale
;
...
...
@@ -183,7 +183,7 @@ PD_REGISTER_ACTIVATION_KERNEL(log2, Log2Kernel)
PD_REGISTER_ACTIVATION_KERNEL
(
log10
,
Log10Kernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
log1p
,
Log1pKernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
swish_raw
,
SwishRawKernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
hardswish
_raw
,
HardSwishRaw
Kernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
hardswish
,
HardSwish
Kernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
round
,
RoundKernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
floor
,
FloorKernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
ceil
,
CeilKernel
)
...
...
paddle/phi/kernels/gpu/activation_grad_kernel.cu
浏览文件 @
3e636ec9
...
...
@@ -274,11 +274,11 @@ template <typename T, typename Context>
void
HardSwishGradKernel
(
const
Context
&
dev_ctx
,
const
DenseTensor
&
x
,
const
DenseTensor
&
dout
,
float
threshold
,
float
scale
,
float
offset
,
DenseTensor
*
dx
)
{
funcs
::
CudaHardSwishGradFunctor
<
T
>
functor
;
float
threshold
=
6
;
float
scale
=
6
;
float
offset
=
3
;
auto
attrs
=
functor
.
GetAttrs
();
*
(
attrs
[
0
].
second
)
=
threshold
;
*
(
attrs
[
1
].
second
)
=
scale
;
...
...
paddle/phi/kernels/gpu/activation_kernel.cu
浏览文件 @
3e636ec9
...
...
@@ -138,13 +138,13 @@ DEFINE_GPU_ACT_KERNEL_WITH_TWO_ATTRS(HardSigmoid,
DEFINE_GPU_ACT_KERNEL_WITH_TWO_ATTRS
(
Selu
,
CudaSeluFunctor
,
scale
,
alpha
)
template
<
typename
T
,
typename
Context
>
void
HardSwishRawKernel
(
const
Context
&
dev_ctx
,
const
DenseTensor
&
x
,
float
threshold
,
float
scale
,
float
offset
,
DenseTensor
*
out
)
{
void
HardSwishKernel
(
const
Context
&
dev_ctx
,
const
DenseTensor
&
x
,
DenseTensor
*
out
)
{
funcs
::
CudaHardSwishFunctor
<
T
>
functor
;
float
threshold
=
6
;
float
scale
=
6
;
float
offset
=
3
;
auto
attrs
=
functor
.
GetAttrs
();
*
(
attrs
[
0
].
second
)
=
threshold
;
*
(
attrs
[
1
].
second
)
=
scale
;
...
...
@@ -257,7 +257,7 @@ PD_REGISTER_ACTIVATION_KERNEL(log, LogKernel)
PD_REGISTER_ACTIVATION_KERNEL
(
log2
,
Log2Kernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
log10
,
Log10Kernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
log1p
,
Log1pKernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
hardswish
_raw
,
HardSwishRaw
Kernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
hardswish
,
HardSwish
Kernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
swish_raw
,
SwishRawKernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
round
,
RoundKernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
floor
,
FloorKernel
)
...
...
paddle/phi/kernels/onednn/activation_grad_kernel.cc
浏览文件 @
3e636ec9
...
...
@@ -238,9 +238,6 @@ template <typename T, typename Context>
void
HardSwishGradKernel
(
const
Context
&
dev_ctx
,
const
DenseTensor
&
x
,
const
DenseTensor
&
dout
,
float
threshold
,
float
scale
,
float
offset
,
DenseTensor
*
dx
)
{
HardSwishOneDNNGradFunctor
<
T
>
functor
;
functor
(
dev_ctx
,
x
,
dout
,
0
,
0
,
dx
);
...
...
paddle/phi/kernels/onednn/activation_kernel.cc
浏览文件 @
3e636ec9
...
...
@@ -157,14 +157,11 @@ DEFINE_ONEDNN_ACT_KERNEL_WITH_ONE_ATTRS(Mish, MishOneDNNFunctor, threshold)
DEFINE_ONEDNN_ACT_KERNEL_WITH_ONE_ATTRS
(
SwishRaw
,
SwishOneDNNFunctor
,
beta
)
template
<
typename
T
,
typename
Context
>
void
HardSwishRawKernel
(
const
Context
&
dev_ctx
,
const
DenseTensor
&
x
,
float
threshold
,
float
scale
,
float
offset
,
DenseTensor
*
out
)
{
void
HardSwishKernel
(
const
Context
&
dev_ctx
,
const
DenseTensor
&
x
,
DenseTensor
*
out
)
{
HardSwishOneDNNFunctor
<
T
>
functor
;
functor
(
dev_ctx
,
x
,
threshold
,
0
,
out
);
functor
(
dev_ctx
,
x
,
6
,
0
,
out
);
}
template
<
typename
T
,
typename
Context
>
...
...
@@ -202,7 +199,7 @@ PD_REGISTER_ACTIVATION_KERNEL(abs, AbsKernel)
PD_REGISTER_ACTIVATION_KERNEL
(
elu
,
EluKernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
exp
,
ExpKernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
gelu
,
GeluKernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
hardswish
_raw
,
HardSwishRaw
Kernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
hardswish
,
HardSwish
Kernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
leaky_relu
,
LeakyReluKernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
mish
,
MishKernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
relu
,
ReluKernel
)
...
...
paddle/phi/kernels/xpu/activation_grad_kernel.cc
浏览文件 @
3e636ec9
...
...
@@ -603,11 +603,11 @@ template <typename T, typename Context>
void
HardSwishGradKernel
(
const
Context
&
dev_ctx
,
const
DenseTensor
&
x
,
const
DenseTensor
&
dout
,
float
threshold
,
float
scale
,
float
offset
,
DenseTensor
*
dx
)
{
XPUHardSwishGradFunctor
<
T
>
functor
;
float
threshold
=
6
;
float
scale
=
6
;
float
offset
=
3
;
auto
attrs
=
functor
.
GetAttrs
();
*
(
attrs
[
0
].
second
)
=
threshold
;
*
(
attrs
[
1
].
second
)
=
scale
;
...
...
paddle/phi/kernels/xpu/activation_kernel.cc
浏览文件 @
3e636ec9
...
...
@@ -513,13 +513,13 @@ DEFINE_XPU_ACTIVATION_KERNEL_WITH_TWO_ATTRS(HardSigmoid,
offset
)
template
<
typename
T
,
typename
Context
>
void
HardSwishRawKernel
(
const
Context
&
dev_ctx
,
const
DenseTensor
&
x
,
float
threshold
,
float
scale
,
float
offset
,
DenseTensor
*
out
)
{
void
HardSwishKernel
(
const
Context
&
dev_ctx
,
const
DenseTensor
&
x
,
DenseTensor
*
out
)
{
XPUHardSwishFunctor
<
T
>
functor
;
float
threshold
=
6
;
float
scale
=
6
;
float
offset
=
3
;
auto
attrs
=
functor
.
GetAttrs
();
*
(
attrs
[
0
].
second
)
=
threshold
;
*
(
attrs
[
1
].
second
)
=
scale
;
...
...
@@ -551,7 +551,7 @@ PD_REGISTER_ACTIVATION_KERNEL(exp, ExpKernel) // no grad
PD_REGISTER_ACTIVATION_KERNEL
(
floor
,
FloorKernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
leaky_relu
,
LeakyReluKernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
hard_sigmoid
,
HardSigmoidKernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
hardswish
_raw
,
HardSwishRaw
Kernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
hardswish
,
HardSwish
Kernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
mish
,
MishKernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
pow
,
PowKernel
)
PD_REGISTER_ACTIVATION_KERNEL
(
reciprocal
,
ReciprocalKernel
)
...
...
paddle/phi/ops/compat/activation_sig.cc
浏览文件 @
3e636ec9
...
...
@@ -41,10 +41,6 @@ namespace phi {
DEFINE_ACT_GRAD_DEPX_OP_ARGMAP
(
HardTanh
,
"hardtanh"
,
"t_min"
comma
"t_max"
);
DEFINE_ACT_GRAD_DEPX_OP_ARGMAP
(
Mish
,
"mish"
,
"threshold"
);
DEFINE_ACT_GRAD_DEPX_OP_ARGMAP
(
HardSwish
,
"hardswish"
,
"threshold"
comma
"scale"
comma
"offset"
);
// NOLINT
DEFINE_ACT_GRAD_DEPX_OP_ARGMAP
(
Swish
,
"swish"
,
"beta"
);
// NOLINT
DEFINE_ACT_GRAD_DEPX_OP_ARGMAP
(
STanh
,
...
...
@@ -53,9 +49,13 @@ DEFINE_ACT_GRAD_DEPX_OP_ARGMAP(STanh,
DEFINE_ACT_GRAD_DEPOUT_OP_ARGMAP
(
Relu6
,
"relu6"
,
"threshold"
);
// NOLINT
KernelSignature
HardSwishGradOpArgumentMapping
(
const
ArgumentMappingContext
&
ctx
)
{
return
KernelSignature
(
"hardswish_grad"
,
{
"X"
,
"Out@GRAD"
},
{},
{
"X@GRAD"
});
}
KernelSignature
HardSwishOpArgumentMapping
(
const
ArgumentMappingContext
&
ctx
)
{
return
KernelSignature
(
"hardswish_raw"
,
{
"X"
},
{
"threshold"
,
"scale"
,
"offset"
},
{
"Out"
});
return
KernelSignature
(
"hardswish"
,
{
"X"
},
{},
{
"Out"
});
}
KernelSignature
SwishOpArgumentMapping
(
const
ArgumentMappingContext
&
ctx
)
{
...
...
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_hard_swish.py
浏览文件 @
3e636ec9
...
...
@@ -40,9 +40,9 @@ class TrtConvertHardSwishTest(TrtLayerAutoScanTest):
def
generate_input1
(
attrs
:
List
[
Dict
[
str
,
Any
]]):
return
np
.
ones
([
1
,
3
,
32
,
32
]).
astype
(
np
.
float32
)
for
threshold
in
[
6.0
,
7.0
,
100.0
,
0.0
,
-
1.0
]:
for
scale
in
[
5.0
,
7.0
,
-
1.0
,
0.0
,
100
.0
]:
for
offset
in
[
3.0
,
5.0
,
-
1.0
,
0.0
,
100.0
]:
for
threshold
in
[
6.0
]:
for
scale
in
[
6
.0
]:
for
offset
in
[
3.0
]:
dics
=
[
{
"threshold"
:
threshold
,
...
...
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