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
PaddleDetection
提交
6af0593c
P
PaddleDetection
项目概览
PaddlePaddle
/
PaddleDetection
大约 1 年 前同步成功
通知
694
Star
11112
Fork
2696
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
184
列表
看板
标记
里程碑
合并请求
40
Wiki
0
Wiki
分析
仓库
DevOps
项目成员
Pages
P
PaddleDetection
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
184
Issue
184
列表
看板
标记
里程碑
合并请求
40
合并请求
40
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
6af0593c
编写于
5月 15, 2018
作者:
S
Siddharth Goyal
提交者:
GitHub
5月 15, 2018
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
Add FP16 option to load_combine op (#10601)
上级
5f6fd26f
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
113 addition
and
13 deletion
+113
-13
paddle/fluid/operators/load_combine_op.cc
paddle/fluid/operators/load_combine_op.cc
+26
-10
paddle/fluid/operators/save_load_combine_op_test.cc
paddle/fluid/operators/save_load_combine_op_test.cc
+87
-3
未找到文件。
paddle/fluid/operators/load_combine_op.cc
浏览文件 @
6af0593c
...
...
@@ -12,7 +12,7 @@ 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 <fstream>
#include "paddle/fluid/framework/data_type_transform.h"
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/platform/device_context.h"
...
...
@@ -31,6 +31,7 @@ class LoadCombineOp : public framework::OperatorBase {
void
RunImpl
(
const
framework
::
Scope
&
scope
,
const
platform
::
Place
&
place
)
const
override
{
auto
filename
=
Attr
<
std
::
string
>
(
"file_path"
);
auto
load_as_fp16
=
Attr
<
bool
>
(
"load_as_fp16"
);
std
::
ifstream
fin
(
filename
);
PADDLE_ENFORCE
(
static_cast
<
bool
>
(
fin
),
...
...
@@ -59,17 +60,25 @@ class LoadCombineOp : public framework::OperatorBase {
// Get data from fin to tensor
DeserializeFromStream
(
fin
,
tensor
,
dev_ctx
);
if
(
platform
::
is_gpu_place
(
place
))
{
// copy CPU to GPU
framework
::
LoDTensor
cpu_tensor
;
cpu_tensor
.
ShareDataWith
(
*
tensor
);
cpu_tensor
.
set_lod
(
tensor
->
lod
());
// reset tensor
auto
in_dtype
=
framework
::
ToDataType
(
tensor
->
type
());
auto
out_dtype
=
load_as_fp16
?
framework
::
proto
::
VarType
::
FP16
:
in_dtype
;
if
(
in_dtype
!=
out_dtype
)
{
// convert to float16 tensor
auto
in_kernel_type
=
framework
::
OpKernelType
(
in_dtype
,
place
);
auto
out_kernel_type
=
framework
::
OpKernelType
(
out_dtype
,
place
);
framework
::
LoDTensor
fp16_tensor
;
// copy LoD info to the new tensor
fp16_tensor
.
set_lod
(
tensor
->
lod
());
framework
::
TransDataType
(
in_kernel_type
,
out_kernel_type
,
*
tensor
,
&
fp16_tensor
);
// reset output tensor
out_var
->
Clear
();
tensor
=
out_var
->
GetMutable
<
framework
::
LoDTensor
>
();
tensor
->
set_lod
(
cpu
_tensor
.
lod
());
TensorCopy
(
cpu_tensor
,
place
,
dev_ctx
,
tensor
);
tensor
->
set_lod
(
fp16
_tensor
.
lod
());
tensor
->
ShareDataWith
(
fp16_
tensor
);
}
}
}
...
...
@@ -82,6 +91,13 @@ class LoadCombineOpProtoMaker : public framework::OpProtoAndCheckerMaker {
"Out"
,
"(vector) The output LoDTensors that will be read from the input file."
)
.
AsDuplicable
();
AddAttr
<
bool
>
(
"load_as_fp16"
,
"(boolean, default false)"
"If true, the tensor will be first loaded and then "
"converted to float16 data type. Otherwise, the tensor will be "
"directly loaded without data type conversion."
)
.
SetDefault
(
false
);
AddAttr
<
std
::
string
>
(
"file_path"
,
"(string) "
"LoDTensors will be loaded from
\"
file_path
\"
."
)
...
...
paddle/fluid/operators/save_load_combine_op_test.cc
浏览文件 @
6af0593c
...
...
@@ -139,8 +139,9 @@ TEST(SaveLoadCombineOp, CPU) {
CheckValues
<
int
,
int
>
(
expect4
,
actual4
,
expect_lod4
,
actual_lod4
,
numel4
);
}
// FP16 version of SaveLoadCombineOp Test
TEST
(
SaveLoadCombineFP16Op
,
CPU
)
{
// FP16 version of SaveLoadCombineOp Test, only altering the saving aspect
// to save as FP16.
TEST
(
SaveCombineFP16Op
,
CPU
)
{
paddle
::
framework
::
Scope
scope
;
paddle
::
platform
::
CPUPlace
place
;
...
...
@@ -169,7 +170,7 @@ TEST(SaveLoadCombineFP16Op, CPU) {
20
,
50
,
lod4
,
"test_var4"
,
place
,
&
scope
,
&
expect_lod4
);
// Set attributes
std
::
string
filename
=
"check_tensor_fp16.ls"
;
std
::
string
filename
=
"check_tensor_fp16
_save
.ls"
;
paddle
::
framework
::
AttributeMap
attrs
;
attrs
.
insert
({
"file_path"
,
std
::
string
(
filename
)});
attrs
.
insert
({
"save_as_fp16"
,
true
});
...
...
@@ -216,6 +217,89 @@ TEST(SaveLoadCombineFP16Op, CPU) {
actual_lod4
,
numel4
);
}
// FP16 version of SaveLoadCombineOp Test, only altering the loading aspect
// to load tensors with FP16 precision.
TEST
(
LoadCombineFP16Op
,
CPU
)
{
paddle
::
framework
::
Scope
scope
;
paddle
::
platform
::
CPUPlace
place
;
std
::
vector
<
int
>
lod1
=
{
0
,
1
,
2
,
3
,
10
};
int
numel1
=
100
;
paddle
::
framework
::
LoD
expect_lod1
;
float
*
expect1
=
CreateForSaveCombineOp
<
float
,
paddle
::
platform
::
float16
>
(
10
,
10
,
lod1
,
"test_var1"
,
place
,
&
scope
,
&
expect_lod1
);
std
::
vector
<
int
>
lod2
=
{
0
,
2
,
5
,
10
};
int
numel2
=
200
;
paddle
::
framework
::
LoD
expect_lod2
;
float
*
expect2
=
CreateForSaveCombineOp
<
float
,
paddle
::
platform
::
float16
>
(
10
,
20
,
lod2
,
"test_var2"
,
place
,
&
scope
,
&
expect_lod2
);
std
::
vector
<
int
>
lod3
=
{
0
,
20
};
int
numel3
=
4000
;
paddle
::
framework
::
LoD
expect_lod3
;
float
*
expect3
=
CreateForSaveCombineOp
<
float
,
paddle
::
platform
::
float16
>
(
20
,
200
,
lod3
,
"test_var3"
,
place
,
&
scope
,
&
expect_lod3
);
std
::
vector
<
int
>
lod4
=
{
0
,
1
,
20
};
int
numel4
=
1000
;
paddle
::
framework
::
LoD
expect_lod4
;
float
*
expect4
=
CreateForSaveCombineOp
<
float
,
paddle
::
platform
::
float16
>
(
20
,
50
,
lod4
,
"test_var4"
,
place
,
&
scope
,
&
expect_lod4
);
// Set attributes
std
::
string
filename
=
"check_tensor_fp16_load.ls"
;
paddle
::
framework
::
AttributeMap
attrs
;
attrs
.
insert
({
"file_path"
,
std
::
string
(
filename
)});
// Run the save_combine_op
auto
save_combine_op
=
paddle
::
framework
::
OpRegistry
::
CreateOp
(
"save_combine"
,
{{
"X"
,
{
"test_var1"
,
"test_var2"
,
"test_var3"
,
"test_var4"
}}},
{},
attrs
);
save_combine_op
->
Run
(
scope
,
place
);
// Set up output vars
auto
load_var1
=
scope
.
Var
(
"out_var1"
);
auto
load_var2
=
scope
.
Var
(
"out_var2"
);
auto
load_var3
=
scope
.
Var
(
"out_var3"
);
auto
load_var4
=
scope
.
Var
(
"out_var4"
);
attrs
.
insert
({
"load_as_fp16"
,
true
});
// Run the load_combine_op
auto
load_combine_op
=
paddle
::
framework
::
OpRegistry
::
CreateOp
(
"load_combine"
,
{},
{{
"Out"
,
{
"out_var1"
,
"out_var2"
,
"out_var3"
,
"out_var4"
}}},
attrs
);
load_combine_op
->
Run
(
scope
,
place
);
auto
*
target1
=
load_var1
->
GetMutable
<
paddle
::
framework
::
LoDTensor
>
();
auto
*
target2
=
load_var2
->
GetMutable
<
paddle
::
framework
::
LoDTensor
>
();
auto
*
target3
=
load_var3
->
GetMutable
<
paddle
::
framework
::
LoDTensor
>
();
auto
*
target4
=
load_var4
->
GetMutable
<
paddle
::
framework
::
LoDTensor
>
();
paddle
::
framework
::
LoD
actual_lod1
,
actual_lod2
,
actual_lod3
,
actual_lod4
;
paddle
::
platform
::
float16
*
actual1
=
GetValuesAfterLoadCombineOp
<
paddle
::
platform
::
float16
>
(
target1
,
scope
,
&
actual_lod1
);
paddle
::
platform
::
float16
*
actual2
=
GetValuesAfterLoadCombineOp
<
paddle
::
platform
::
float16
>
(
target2
,
scope
,
&
actual_lod2
);
paddle
::
platform
::
float16
*
actual3
=
GetValuesAfterLoadCombineOp
<
paddle
::
platform
::
float16
>
(
target3
,
scope
,
&
actual_lod3
);
paddle
::
platform
::
float16
*
actual4
=
GetValuesAfterLoadCombineOp
<
paddle
::
platform
::
float16
>
(
target4
,
scope
,
&
actual_lod4
);
CheckValues
<
float
,
paddle
::
platform
::
float16
>
(
expect1
,
actual1
,
expect_lod1
,
actual_lod1
,
numel1
);
CheckValues
<
float
,
paddle
::
platform
::
float16
>
(
expect2
,
actual2
,
expect_lod2
,
actual_lod2
,
numel2
);
CheckValues
<
float
,
paddle
::
platform
::
float16
>
(
expect3
,
actual3
,
expect_lod3
,
actual_lod3
,
numel3
);
CheckValues
<
float
,
paddle
::
platform
::
float16
>
(
expect4
,
actual4
,
expect_lod4
,
actual_lod4
,
numel4
);
}
// Test with original SaveLoadTest
TEST
(
SaveLoadTestWithCombineOp
,
CPU
)
{
paddle
::
framework
::
Scope
scope
;
...
...
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