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
极客李华
Paddle
提交
f99ea99e
P
Paddle
项目概览
极客李华
/
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
f99ea99e
编写于
10月 11, 2018
作者:
Q
Qiyang Min
提交者:
GitHub
10月 11, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #13720 from velconia/fix_grad_clip
Merge selected_rows for clip_by_norm op
上级
afdc7309
1456b8ec
变更
3
隐藏空白更改
内联
并排
Showing
3 changed file
with
96 addition
and
4 deletion
+96
-4
paddle/fluid/operators/CMakeLists.txt
paddle/fluid/operators/CMakeLists.txt
+2
-1
paddle/fluid/operators/clip_by_norm_op.h
paddle/fluid/operators/clip_by_norm_op.h
+37
-3
python/paddle/fluid/tests/unittests/test_clip_by_norm_op.py
python/paddle/fluid/tests/unittests/test_clip_by_norm_op.py
+57
-0
未找到文件。
paddle/fluid/operators/CMakeLists.txt
浏览文件 @
f99ea99e
...
...
@@ -230,7 +230,7 @@ if(WITH_DISTRIBUTE)
op_library
(
${
dist_op
}
DEPS
${
DISTRIBUTE_DEPS
}
)
set_source_files_properties
(
${
dist_op
}
.cc PROPERTIES COMPILE_FLAGS
${
DISTRIBUTE_COMPILE_FLAGS
}
)
endforeach
()
#set_source_files_properties(send_recv_op_test.cc PROPERTIES COMPILE_FLAGS ${DISTRIBUTE_COMPILE_FLAGS})
#cc_test(test_send_recv SRCS send_recv_op_test.cc DEPS prefetch_op send_op
# listen_and_serv_op sum_op executor SERIAL)
...
...
@@ -268,6 +268,7 @@ if (WITH_GPU AND TENSORRT_FOUND)
else
()
set
(
DEPS_OPS
${
DEPS_OPS
}
tensorrt_engine_op
)
endif
()
op_library
(
clip_by_norm_op DEPS selected_rows_functor selected_rows
)
op_library
(
sum_op DEPS selected_rows_functor
)
op_library
(
sgd_op DEPS selected_rows_functor
)
op_library
(
print_op DEPS lod_tensor
)
...
...
paddle/fluid/operators/clip_by_norm_op.h
浏览文件 @
f99ea99e
...
...
@@ -16,12 +16,15 @@ limitations under the License. */
#include "paddle/fluid/framework/eigen.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/framework/selected_rows.h"
#include "paddle/fluid/operators/math/selected_rows_functor.h"
#include "paddle/fluid/platform/transform.h"
namespace
paddle
{
namespace
operators
{
using
Tensor
=
framework
::
Tensor
;
using
SelectedRows
=
framework
::
SelectedRows
;
template
<
typename
T
,
int
MajorType
=
Eigen
::
RowMajor
,
typename
IndexType
=
Eigen
::
DenseIndex
>
using
EigenVector
=
framework
::
EigenVector
<
T
,
MajorType
,
IndexType
>
;
...
...
@@ -31,9 +34,40 @@ class ClipByNormKernel : public framework::OpKernel<T> {
public:
void
Compute
(
const
framework
::
ExecutionContext
&
context
)
const
override
{
auto
max_norm
=
context
.
Attr
<
T
>
(
"max_norm"
);
auto
*
input
=
context
.
Input
<
Tensor
>
(
"X"
);
auto
*
output
=
context
.
Output
<
Tensor
>
(
"Out"
);
output
->
mutable_data
<
T
>
(
context
.
GetPlace
());
auto
in_var
=
context
.
InputVar
(
"X"
);
Tensor
*
output
=
nullptr
;
const
Tensor
*
input
=
nullptr
;
if
(
in_var
->
IsType
<
framework
::
LoDTensor
>
())
{
input
=
context
.
Input
<
Tensor
>
(
"X"
);
output
=
context
.
Output
<
Tensor
>
(
"Out"
);
output
->
mutable_data
<
T
>
(
context
.
GetPlace
());
}
else
if
(
in_var
->
IsType
<
SelectedRows
>
())
{
auto
*
x
=
context
.
Input
<
SelectedRows
>
(
"X"
);
// merge ids in selected rows first
math
::
scatter
::
MergeAdd
<
DeviceContext
,
T
>
merge_func
;
SelectedRows
*
merged_input
=
const_cast
<
framework
::
Scope
&>
(
context
.
scope
())
.
Var
()
->
GetMutable
<
SelectedRows
>
();
merge_func
(
context
.
template
device_context
<
DeviceContext
>(),
*
x
,
merged_input
);
input
=
&
(
merged_input
->
value
());
SelectedRows
*
output_selected_rows
=
context
.
Output
<
SelectedRows
>
(
"Out"
);
output_selected_rows
->
set_rows
(
merged_input
->
rows
());
output_selected_rows
->
set_height
(
merged_input
->
height
());
output
=
output_selected_rows
->
mutable_value
();
output
->
Resize
(
merged_input
->
value
().
dims
());
output
->
mutable_data
<
T
>
(
context
.
GetPlace
());
}
else
{
PADDLE_THROW
(
"Unexpected branch, input variable type is %s"
,
in_var
->
Type
().
name
());
}
PADDLE_ENFORCE_NOT_NULL
(
input
);
auto
x
=
EigenVector
<
T
>::
Flatten
(
*
input
);
auto
out
=
EigenVector
<
T
>::
Flatten
(
*
output
);
...
...
python/paddle/fluid/tests/unittests/test_clip_by_norm_op.py
浏览文件 @
f99ea99e
...
...
@@ -18,6 +18,9 @@ import unittest
import
numpy
as
np
from
op_test
import
OpTest
import
paddle.fluid
as
fluid
import
paddle.fluid.core
as
core
class
TestClipByNormOp
(
OpTest
):
def
setUp
(
self
):
...
...
@@ -62,5 +65,59 @@ class TestCase3(TestClipByNormOp):
self
.
max_norm
=
1.0
class
TestClipByNormOpWithSelectedRows
(
OpTest
):
def
check_with_place
(
self
,
place
):
self
.
config_test_case
()
scope
=
core
.
Scope
()
# set input
x_selected_rows
=
scope
.
var
(
'X'
).
get_selected_rows
()
x_selected_rows
.
set_rows
(
self
.
grad_rows
)
x_tensor
=
x_selected_rows
.
get_tensor
()
x_np
=
np
.
random
.
random
(
self
.
grad_shape
).
astype
(
"float32"
)
x_np
[
np
.
abs
(
x_np
)
<
self
.
max_relative_error
]
=
0.5
x_tensor
.
set
(
x_np
,
place
)
# set output
out_selected_rows
=
scope
.
var
(
'Out'
).
get_selected_rows
()
# run clip_by_norm_op
clip_by_norm_op
=
fluid
.
op
.
Operator
(
"clip_by_norm"
,
max_norm
=
self
.
max_norm
,
X
=
'X'
,
Out
=
'Out'
)
clip_by_norm_op
.
run
(
scope
,
place
)
# check output
self
.
assertEqual
(
out_selected_rows
.
rows
(),
self
.
grad_clipped_rows
)
out_tensor
=
out_selected_rows
.
get_tensor
()
y_np
=
np
.
zeros
(
self
.
grad_clipped_shape
)
y_np
[
0
]
=
np
.
sum
(
x_np
[
0
:
2
])
y_np
[
1
]
=
x_np
[
2
]
y_np
[
2
]
=
x_np
[
3
]
norm
=
np
.
sqrt
(
np
.
sum
(
np
.
square
(
y_np
)))
if
norm
>
self
.
max_norm
:
output
=
self
.
max_norm
*
y_np
/
norm
else
:
output
=
y_np
self
.
assertTrue
(
np
.
allclose
(
np
.
array
(
out_tensor
),
output
,
atol
=
1e-5
,
equal_nan
=
False
))
def
test_clip_by_norm_with_selected_ros
(
self
):
places
=
[
core
.
CPUPlace
()]
if
core
.
is_compiled_with_cuda
():
places
.
append
(
core
.
CUDAPlace
(
0
))
for
place
in
places
:
self
.
check_with_place
(
place
)
def
config_test_case
(
self
):
self
.
max_norm
=
1.0
self
.
max_relative_error
=
0.006
self
.
grad_shape
=
(
4
,
1
)
self
.
grad_clipped_shape
=
(
3
,
1
)
self
.
grad_rows
=
[
0
,
0
,
1
,
2
]
self
.
grad_clipped_rows
=
[
0
,
1
,
2
]
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