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
cfa69133
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
体验新版 GitCode,发现更多精彩内容 >>
未验证
提交
cfa69133
编写于
8月 12, 2021
作者:
W
wuhuachaocoding
提交者:
GitHub
8月 12, 2021
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[NPU] Support npu kernel for smooth_l1_loss op (#34674)
上级
bc543e35
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
350 addition
and
0 deletion
+350
-0
paddle/fluid/operators/smooth_l1_loss_op_npu.cc
paddle/fluid/operators/smooth_l1_loss_op_npu.cc
+203
-0
python/paddle/fluid/tests/unittests/npu/test_smooth_l1_loss_op_npu.py
...e/fluid/tests/unittests/npu/test_smooth_l1_loss_op_npu.py
+147
-0
未找到文件。
paddle/fluid/operators/smooth_l1_loss_op_npu.cc
0 → 100644
浏览文件 @
cfa69133
/* Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/smooth_l1_loss_op.h"
#include "paddle/fluid/framework/tensor_util.h"
#include "paddle/fluid/operators/npu_op_runner.h"
namespace
paddle
{
namespace
operators
{
template
<
typename
DeviceContext
,
typename
T
>
class
SmoothL1LossNPUKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
auto
*
in_x
=
context
.
Input
<
Tensor
>
(
"X"
);
auto
*
in_y
=
context
.
Input
<
Tensor
>
(
"Y"
);
auto
*
inside_weight
=
context
.
Input
<
Tensor
>
(
"InsideWeight"
);
auto
*
outside_weight
=
context
.
Input
<
Tensor
>
(
"OutsideWeight"
);
auto
*
out_diff
=
context
.
Output
<
Tensor
>
(
"Diff"
);
auto
*
out_loss
=
context
.
Output
<
Tensor
>
(
"Out"
);
out_diff
->
mutable_data
<
T
>
(
context
.
GetPlace
());
out_loss
->
mutable_data
<
T
>
(
context
.
GetPlace
());
auto
sigma
=
context
.
Attr
<
float
>
(
"sigma"
);
T
sigma2
=
1.0
/
(
sigma
*
sigma
);
bool
has_weight
=
(
inside_weight
!=
nullptr
)
&&
(
outside_weight
!=
nullptr
);
// out_diff = in_x - in_y
auto
stream
=
context
.
template
device_context
<
paddle
::
platform
::
NPUDeviceContext
>()
.
stream
();
const
auto
&
runner1
=
NpuOpRunner
(
"Sub"
,
{
*
in_x
,
*
in_y
},
{
*
out_diff
},
{});
runner1
.
Run
(
stream
);
Tensor
no_reduce_loss
(
in_x
->
type
());
no_reduce_loss
.
Resize
(
in_x
->
dims
());
no_reduce_loss
.
mutable_data
<
T
>
(
context
.
GetPlace
());
// multiply inside weight before get the loss
if
(
has_weight
)
{
Tensor
tmp_diff
(
out_diff
->
type
());
tmp_diff
.
Resize
(
out_diff
->
dims
());
tmp_diff
.
mutable_data
<
T
>
(
context
.
GetPlace
());
const
auto
&
runner2
=
NpuOpRunner
(
"Mul"
,
{
*
out_diff
,
*
inside_weight
},
{
tmp_diff
},
{});
runner2
.
Run
(
stream
);
framework
::
TensorCopy
(
tmp_diff
,
context
.
GetPlace
(),
context
.
template
device_context
<
paddle
::
platform
::
NPUDeviceContext
>(),
out_diff
);
Tensor
tmp_x
(
in_x
->
type
());
tmp_x
.
Resize
(
in_x
->
dims
());
tmp_x
.
mutable_data
<
T
>
(
context
.
GetPlace
());
Tensor
tmp_y
(
in_y
->
type
());
tmp_y
.
Resize
(
in_y
->
dims
());
tmp_y
.
mutable_data
<
T
>
(
context
.
GetPlace
());
// mul input and inside_weight
const
auto
&
runner_x
=
NpuOpRunner
(
"Mul"
,
{
*
in_x
,
*
inside_weight
},
{
tmp_x
},
{});
runner_x
.
Run
(
stream
);
const
auto
&
runner_y
=
NpuOpRunner
(
"Mul"
,
{
*
in_y
,
*
inside_weight
},
{
tmp_y
},
{});
runner_y
.
Run
(
stream
);
const
auto
&
runner3
=
NpuOpRunner
(
"SmoothL1Loss"
,
{
tmp_x
,
tmp_y
},
{
no_reduce_loss
},
{{
"sigma"
,
sigma2
}});
runner3
.
Run
(
stream
);
}
else
{
const
auto
&
runner3
=
NpuOpRunner
(
"SmoothL1Loss"
,
{
*
in_x
,
*
in_y
},
{
no_reduce_loss
},
{{
"sigma"
,
sigma2
}});
runner3
.
Run
(
stream
);
}
// multiply outside weight and loss
// reduceSum because the output'shape must be [B,1]
if
(
has_weight
)
{
Tensor
tmp_loss
(
no_reduce_loss
.
type
());
tmp_loss
.
Resize
(
no_reduce_loss
.
dims
());
tmp_loss
.
mutable_data
<
T
>
(
context
.
GetPlace
());
const
auto
&
runner4
=
NpuOpRunner
(
"Mul"
,
{
no_reduce_loss
,
*
outside_weight
},
{
tmp_loss
},
{});
runner4
.
Run
(
stream
);
const
auto
&
runner5
=
NpuOpRunner
(
"ReduceSumD"
,
{
tmp_loss
},
{
*
out_loss
},
{{
"axes"
,
std
::
vector
<
int
>
{
1
}},
{
"keep_dims"
,
true
}});
runner5
.
Run
(
stream
);
}
else
{
const
auto
&
runner5
=
NpuOpRunner
(
"ReduceSumD"
,
{
no_reduce_loss
},
{
*
out_loss
},
{{
"axes"
,
std
::
vector
<
int
>
{
1
}},
{
"keep_dims"
,
true
}});
runner5
.
Run
(
stream
);
}
}
};
template
<
typename
DeviceContext
,
typename
T
>
class
SmoothL1LossGradNPUKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
auto
*
inside_weight
=
context
.
Input
<
Tensor
>
(
"InsideWeight"
);
auto
*
outside_weight
=
context
.
Input
<
Tensor
>
(
"OutsideWeight"
);
auto
*
diff
=
context
.
Input
<
Tensor
>
(
"Diff"
);
auto
*
og
=
context
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"Out"
));
auto
*
outx_grad
=
context
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"X"
));
auto
*
outy_grad
=
context
.
Output
<
Tensor
>
(
framework
::
GradVarName
(
"Y"
));
auto
sigma
=
context
.
Attr
<
T
>
(
"sigma"
);
T
sigma2
=
1.0
/
(
sigma
*
sigma
);
bool
has_weight
=
(
inside_weight
!=
nullptr
)
&&
(
outside_weight
!=
nullptr
);
auto
stream
=
context
.
template
device_context
<
paddle
::
platform
::
NPUDeviceContext
>()
.
stream
();
// diff == in_x - in_y == diff - 0
Tensor
tmp_zero
(
diff
->
type
());
tmp_zero
.
Resize
(
diff
->
dims
());
tmp_zero
.
mutable_data
<
T
>
(
context
.
GetPlace
());
const
auto
&
runner_zero
=
NpuOpRunner
(
"ZerosLike"
,
{
*
diff
},
{
tmp_zero
},
{});
runner_zero
.
Run
(
stream
);
Tensor
grad
(
diff
->
type
());
grad
.
Resize
(
diff
->
dims
());
grad
.
mutable_data
<
T
>
(
context
.
GetPlace
());
// broadcast og(output_grad) to adapt to the npu interface
const
auto
&
runner_broad
=
NpuOpRunner
(
"BroadcastToD"
,
{
*
og
},
{
grad
},
{{
"shape"
,
framework
::
vectorize
(
diff
->
dims
())}});
runner_broad
.
Run
(
stream
);
Tensor
gradient
(
diff
->
type
());
gradient
.
Resize
(
diff
->
dims
());
gradient
.
mutable_data
<
T
>
(
context
.
GetPlace
());
// diff == diff - 0 == in_x - in_y
const
auto
&
runner_grad
=
NpuOpRunner
(
"SmoothL1LossGrad"
,
{
*
diff
,
tmp_zero
,
grad
},
{
gradient
},
{{
"sigma"
,
sigma2
}});
runner_grad
.
Run
(
stream
);
// mul weight and gradient
if
(
has_weight
)
{
Tensor
weight
(
inside_weight
->
type
());
weight
.
Resize
(
inside_weight
->
dims
());
weight
.
mutable_data
<
T
>
(
context
.
GetPlace
());
const
auto
&
runner_weight
=
NpuOpRunner
(
"Mul"
,
{
*
inside_weight
,
*
outside_weight
},
{
weight
},
{});
runner_weight
.
Run
(
stream
);
Tensor
tmp_grad
(
gradient
.
type
());
tmp_grad
.
Resize
(
gradient
.
dims
());
tmp_grad
.
mutable_data
<
T
>
(
context
.
GetPlace
());
const
auto
&
runner_weight_grad
=
NpuOpRunner
(
"Mul"
,
{
gradient
,
weight
},
{
tmp_grad
},
{});
runner_weight_grad
.
Run
(
stream
);
framework
::
TensorCopy
(
tmp_grad
,
context
.
GetPlace
(),
context
.
template
device_context
<
paddle
::
platform
::
NPUDeviceContext
>(),
&
gradient
);
}
// outx_grad = gradient
if
(
outx_grad
)
{
outx_grad
->
mutable_data
<
T
>
(
context
.
GetPlace
());
framework
::
TensorCopy
(
gradient
,
context
.
GetPlace
(),
context
.
template
device_context
<
paddle
::
platform
::
NPUDeviceContext
>(),
outx_grad
);
}
// outy_grad = - gradient
if
(
outy_grad
)
{
outy_grad
->
mutable_data
<
T
>
(
context
.
GetPlace
());
Tensor
coeff
(
framework
::
proto
::
VarType
::
FP32
);
coeff
.
mutable_data
<
float
>
({
1
},
context
.
GetPlace
());
FillNpuTensorWithConstant
<
float
>
(
&
coeff
,
-
1
);
const
auto
&
runner_y_grad
=
NpuOpRunner
(
"Mul"
,
{
coeff
,
gradient
},
{
*
outy_grad
},
{});
runner_y_grad
.
Run
(
stream
);
}
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OP_NPU_KERNEL
(
smooth_l1_loss
,
ops
::
SmoothL1LossNPUKernel
<
paddle
::
platform
::
NPUDeviceContext
,
float
>
);
REGISTER_OP_NPU_KERNEL
(
smooth_l1_loss_grad
,
ops
::
SmoothL1LossGradNPUKernel
<
paddle
::
platform
::
NPUDeviceContext
,
float
>
);
python/paddle/fluid/tests/unittests/npu/test_smooth_l1_loss_op_npu.py
0 → 100644
浏览文件 @
cfa69133
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
from
__future__
import
print_function
import
unittest
import
numpy
as
np
import
sys
sys
.
path
.
append
(
".."
)
from
op_test
import
OpTest
import
paddle
import
paddle.fluid
as
fluid
paddle
.
enable_static
()
def
smooth_l1_loss_forward
(
val
,
sigma2
):
abs_val
=
abs
(
val
)
if
abs_val
<
1.0
/
sigma2
:
return
0.5
*
val
*
val
*
sigma2
else
:
return
abs_val
-
0.5
/
sigma2
class
TestSmoothL1LossOp1
(
OpTest
):
def
setUp
(
self
):
self
.
set_npu
()
self
.
place
=
paddle
.
NPUPlace
(
0
)
self
.
op_type
=
"smooth_l1_loss"
dims
=
(
5
,
20
)
self
.
inputs
=
{
'X'
:
np
.
random
.
random
(
dims
).
astype
(
"float32"
),
'Y'
:
np
.
random
.
random
(
dims
).
astype
(
"float32"
)
}
sigma
=
3.0
self
.
attrs
=
{
'sigma'
:
sigma
}
sigma2
=
sigma
*
sigma
diff
=
self
.
inputs
[
'X'
]
-
self
.
inputs
[
'Y'
]
loss
=
np
.
vectorize
(
smooth_l1_loss_forward
)(
diff
,
sigma2
).
sum
(
1
)
loss
=
loss
.
reshape
((
dims
[
0
],
1
))
self
.
outputs
=
{
'Diff'
:
diff
.
astype
(
'float32'
),
'Out'
:
loss
.
astype
(
'float32'
)
}
def
set_npu
(
self
):
self
.
__class__
.
use_npu
=
True
def
test_check_output
(
self
):
self
.
check_output_with_place
(
self
.
place
)
def
test_check_grad_normal
(
self
):
self
.
check_grad_with_place
(
self
.
place
,
[
'X'
,
'Y'
],
'Out'
,
max_relative_error
=
0.02
)
def
test_check_grad_ingore_x
(
self
):
self
.
check_grad_with_place
(
self
.
place
,
[
'Y'
],
'Out'
,
max_relative_error
=
0.03
,
no_grad_set
=
set
(
"X"
))
def
test_check_grad_ingore_y
(
self
):
self
.
check_grad_with_place
(
self
.
place
,
[
'X'
],
'Out'
,
max_relative_error
=
0.03
,
no_grad_set
=
set
(
'Y'
))
class
TestSmoothL1LossOp2
(
OpTest
):
def
setUp
(
self
):
self
.
set_npu
()
self
.
place
=
paddle
.
NPUPlace
(
0
)
self
.
op_type
=
"smooth_l1_loss"
dims
=
(
5
,
20
)
self
.
inputs
=
{
'X'
:
np
.
random
.
random
(
dims
).
astype
(
"float32"
),
'Y'
:
np
.
random
.
random
(
dims
).
astype
(
"float32"
),
'InsideWeight'
:
np
.
random
.
random
(
dims
).
astype
(
"float32"
),
'OutsideWeight'
:
np
.
random
.
random
(
dims
).
astype
(
"float32"
)
}
sigma
=
3.0
self
.
attrs
=
{
'sigma'
:
sigma
}
sigma2
=
sigma
*
sigma
diff
=
self
.
inputs
[
'X'
]
-
self
.
inputs
[
'Y'
]
diff
=
diff
*
self
.
inputs
[
'InsideWeight'
]
loss
=
np
.
vectorize
(
smooth_l1_loss_forward
)(
diff
,
sigma2
)
loss
=
loss
*
self
.
inputs
[
'OutsideWeight'
]
loss
=
loss
.
sum
(
1
).
reshape
((
dims
[
0
],
1
))
self
.
outputs
=
{
'Diff'
:
diff
.
astype
(
'float32'
),
'Out'
:
loss
.
astype
(
'float32'
)
}
def
set_npu
(
self
):
self
.
__class__
.
use_npu
=
True
def
test_check_output
(
self
):
self
.
check_output_with_place
(
self
.
place
)
def
test_check_grad_normal
(
self
):
self
.
check_grad_with_place
(
self
.
place
,
[
'X'
,
'Y'
],
'Out'
,
max_relative_error
=
0.03
)
def
test_check_grad_ingore_x
(
self
):
self
.
check_grad_with_place
(
self
.
place
,
[
'Y'
],
'Out'
,
max_relative_error
=
0.03
,
no_grad_set
=
set
([
'X'
,
'InsideWeight'
,
'OutsideWeight'
]))
def
test_check_grad_ingore_y
(
self
):
self
.
check_grad_with_place
(
self
.
place
,
[
'X'
],
'Out'
,
max_relative_error
=
0.03
,
no_grad_set
=
set
([
'Y'
,
'InsideWeight'
,
'OutsideWeight'
]))
class
TestSmoothL1LossOpError
(
unittest
.
TestCase
):
def
test_errors
(
self
):
with
fluid
.
program_guard
(
fluid
.
Program
(),
fluid
.
Program
()):
# The input type of accuracy_op must be Variable.
x1
=
fluid
.
create_lod_tensor
(
np
.
array
([[
-
1
]]),
[[
1
]],
fluid
.
NPUPlace
(
0
))
y1
=
fluid
.
create_lod_tensor
(
np
.
array
([[
-
1
]]),
[[
1
]],
fluid
.
NPUPlace
(
0
))
self
.
assertRaises
(
TypeError
,
fluid
.
layers
.
smooth_l1
,
x1
,
y1
)
# The input dtype of accuracy_op must be float32 or float64.
x2
=
fluid
.
layers
.
data
(
name
=
'x2'
,
shape
=
[
4
],
dtype
=
"int32"
)
y2
=
fluid
.
layers
.
data
(
name
=
'x2'
,
shape
=
[
4
],
dtype
=
"int32"
)
self
.
assertRaises
(
TypeError
,
fluid
.
layers
.
smooth_l1
,
x2
,
y2
)
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