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
csdn_franckjun
Paddle
提交
d0ef6825
P
Paddle
项目概览
csdn_franckjun
/
Paddle
与 Fork 源项目一致
Fork自
PaddlePaddle / Paddle
通知
1
Star
1
Fork
0
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1
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
1
Issue
1
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
d0ef6825
编写于
3月 19, 2019
作者:
Z
Zeng Jinle
提交者:
GitHub
3月 19, 2019
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #16274 from sneaxiy/fix_grad_maker
Remove unused variables in op grad maker
上级
804afc51
023a3a3d
变更
6
隐藏空白更改
内联
并排
Showing
6 changed file
with
87 addition
and
32 deletion
+87
-32
paddle/fluid/framework/scope.cc
paddle/fluid/framework/scope.cc
+1
-1
paddle/fluid/operators/conv_transpose_op.cc
paddle/fluid/operators/conv_transpose_op.cc
+26
-3
paddle/fluid/operators/dropout_op.cc
paddle/fluid/operators/dropout_op.cc
+22
-11
paddle/fluid/operators/layer_norm_op.cc
paddle/fluid/operators/layer_norm_op.cc
+30
-2
paddle/fluid/operators/layer_norm_op.h
paddle/fluid/operators/layer_norm_op.h
+7
-14
paddle/fluid/operators/softmax_with_cross_entropy_op.cc
paddle/fluid/operators/softmax_with_cross_entropy_op.cc
+1
-1
未找到文件。
paddle/fluid/framework/scope.cc
浏览文件 @
d0ef6825
...
...
@@ -34,7 +34,7 @@ DEFINE_double(
"Memory size threshold (GB) when the garbage collector clear tensors."
"Disabled when this value is less than 0"
);
DEFINE_bool
(
fast_eager_deletion_mode
,
fals
e
,
DEFINE_bool
(
fast_eager_deletion_mode
,
tru
e
,
"Fast eager deletion mode. If enabled, memory would release "
"immediately without waiting GPU kernel ends."
);
...
...
paddle/fluid/operators/conv_transpose_op.cc
浏览文件 @
d0ef6825
...
...
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/conv_transpose_op.h"
#include <memory>
#include <string>
#include <vector>
...
...
@@ -344,6 +345,28 @@ framework::OpKernelType ConvTransposeOpGrad::GetExpectedKernelType(
ctx
.
GetPlace
(),
layout_
,
library_
);
}
class
ConvTransposeGradOpDescMaker
:
public
framework
::
SingleGradOpDescMaker
{
public:
using
framework
::
SingleGradOpDescMaker
::
SingleGradOpDescMaker
;
protected:
std
::
unique_ptr
<
framework
::
OpDesc
>
Apply
()
const
override
{
std
::
unique_ptr
<
framework
::
OpDesc
>
op
(
new
framework
::
OpDesc
());
op
->
SetType
(
ForwardOp
().
Type
()
+
"_grad"
);
op
->
SetInput
(
"Input"
,
Input
(
"Input"
));
op
->
SetInput
(
"Filter"
,
Input
(
"Filter"
));
op
->
SetOutput
(
framework
::
GradVarName
(
"Input"
),
InputGrad
(
"Input"
));
op
->
SetOutput
(
framework
::
GradVarName
(
"Filter"
),
InputGrad
(
"Filter"
));
if
(
ForwardOp
().
Inputs
().
count
(
"Bias"
)
>
0
)
{
op
->
SetInput
(
"Bias"
,
Input
(
"Bias"
));
op
->
SetOutput
(
framework
::
GradVarName
(
"Bias"
),
InputGrad
(
"Bias"
));
}
op
->
SetInput
(
framework
::
GradVarName
(
"Output"
),
OutputGrad
(
"Output"
));
op
->
SetAttrMap
(
Attrs
());
return
op
;
}
};
}
// namespace operators
}
// namespace paddle
...
...
@@ -352,7 +375,7 @@ namespace ops = paddle::operators;
// conv2d_transpose
REGISTER_OPERATOR
(
conv2d_transpose
,
ops
::
ConvTransposeOp
,
ops
::
Conv2DTransposeOpMaker
,
paddle
::
framework
::
DefaultGradOpDescMaker
<
true
>
);
ops
::
ConvTransposeGradOpDescMaker
);
REGISTER_OPERATOR
(
conv2d_transpose_grad
,
ops
::
ConvTransposeOpGrad
);
REGISTER_OP_CPU_KERNEL
(
...
...
@@ -368,7 +391,7 @@ REGISTER_OP_CPU_KERNEL(
// conv3d_transpose
REGISTER_OPERATOR
(
conv3d_transpose
,
ops
::
ConvTransposeOp
,
ops
::
Conv3DTransposeOpMaker
,
paddle
::
framework
::
DefaultGradOpDescMaker
<
true
>
);
ops
::
ConvTransposeGradOpDescMaker
);
REGISTER_OPERATOR
(
conv3d_transpose_grad
,
ops
::
ConvTransposeOpGrad
);
REGISTER_OP_CPU_KERNEL
(
...
...
@@ -384,7 +407,7 @@ REGISTER_OP_CPU_KERNEL(
// depthwise conv2d_transpose
REGISTER_OPERATOR
(
depthwise_conv2d_transpose
,
ops
::
ConvTransposeOp
,
ops
::
Conv2DTransposeOpMaker
,
paddle
::
framework
::
DefaultGradOpDescMaker
<
true
>
);
ops
::
ConvTransposeGradOpDescMaker
);
REGISTER_OPERATOR
(
depthwise_conv2d_transpose_grad
,
ops
::
ConvTransposeOpGrad
);
REGISTER_OP_CPU_KERNEL
(
...
...
paddle/fluid/operators/dropout_op.cc
浏览文件 @
d0ef6825
...
...
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/dropout_op.h"
#include <memory>
#include <string>
namespace
paddle
{
...
...
@@ -106,21 +107,31 @@ class DropoutOpGrad : public framework::OperatorWithKernel {
PADDLE_ENFORCE_EQ
(
ctx
->
Attrs
().
Get
<
bool
>
(
"is_test"
),
false
,
"GradOp is only callable when is_test is false"
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"X"
),
"Input(X) must not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
"Mask"
),
"Mask must not be null."
);
PADDLE_ENFORCE
(
ctx
->
HasInput
(
framework
::
GradVarName
(
"Out"
)),
"Input(Out@GRAD) must not be null."
);
auto
x_dims
=
ctx
->
GetInputDim
(
"X"
);
auto
out_dims
=
ctx
->
GetInputDim
(
framework
::
GradVarName
(
"Out"
));
PADDLE_ENFORCE_EQ
(
x_dims
,
out_dims
,
"Dimensions of Input(X) and Out@Grad must be the same."
);
auto
mask_dims
=
ctx
->
GetInputDim
(
"Mask"
);
PADDLE_ENFORCE_EQ
(
x_dims
,
mask_dims
,
"Dimensions of Input(X) and Mask must be the same."
);
ctx
->
SetOutputDim
(
framework
::
GradVarName
(
"X"
),
x_dims
);
ctx
->
ShareLoD
(
"X"
,
/*->*/
framework
::
GradVarName
(
"X"
));
ctx
->
SetOutputDim
(
framework
::
GradVarName
(
"X"
),
out_dims
);
ctx
->
ShareLoD
(
framework
::
GradVarName
(
"Out"
),
/*->*/
framework
::
GradVarName
(
"X"
));
}
};
class
DropoutGradOpDescMaker
:
public
framework
::
SingleGradOpDescMaker
{
public:
using
framework
::
SingleGradOpDescMaker
::
SingleGradOpDescMaker
;
protected:
std
::
unique_ptr
<
framework
::
OpDesc
>
Apply
()
const
override
{
std
::
unique_ptr
<
framework
::
OpDesc
>
op
(
new
framework
::
OpDesc
());
op
->
SetType
(
"dropout_grad"
);
op
->
SetInput
(
framework
::
GradVarName
(
"Out"
),
OutputGrad
(
"Out"
));
op
->
SetInput
(
"Mask"
,
Output
(
"Mask"
));
op
->
SetOutput
(
framework
::
GradVarName
(
"X"
),
InputGrad
(
"X"
));
op
->
SetAttrMap
(
Attrs
());
return
op
;
}
};
...
...
@@ -129,7 +140,7 @@ class DropoutOpGrad : public framework::OperatorWithKernel {
namespace
ops
=
paddle
::
operators
;
REGISTER_OPERATOR
(
dropout
,
ops
::
DropoutOp
,
ops
::
DropoutOpMaker
,
paddle
::
framework
::
DefaultGradOpDescMaker
<
true
>
);
ops
::
DropoutGradOpDescMaker
);
REGISTER_OPERATOR
(
dropout_grad
,
ops
::
DropoutOpGrad
);
REGISTER_OP_CPU_KERNEL
(
dropout
,
ops
::
CPUDropoutKernel
<
paddle
::
platform
::
CPUDeviceContext
,
float
>
,
...
...
paddle/fluid/operators/layer_norm_op.cc
浏览文件 @
d0ef6825
...
...
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/layer_norm_op.h"
#include <memory>
namespace
paddle
{
namespace
operators
{
...
...
@@ -133,7 +134,7 @@ class LayerNormGradOp : public framework::OperatorWithKernel {
}
if
(
ctx
->
HasOutput
(
framework
::
GradVarName
(
"Bias"
)))
{
ctx
->
SetOutputDim
(
framework
::
GradVarName
(
"Bias"
),
ctx
->
GetInputDim
(
"
Bias
"
));
ctx
->
GetInputDim
(
"
Scale
"
));
}
}
...
...
@@ -157,12 +158,39 @@ class LayerNormGradOp : public framework::OperatorWithKernel {
}
};
class
LayerNormGradOpDescMaker
:
public
framework
::
SingleGradOpDescMaker
{
public:
using
framework
::
SingleGradOpDescMaker
::
SingleGradOpDescMaker
;
protected:
std
::
unique_ptr
<
framework
::
OpDesc
>
Apply
()
const
override
{
std
::
unique_ptr
<
framework
::
OpDesc
>
op
(
new
framework
::
OpDesc
());
op
->
SetType
(
"layer_norm_grad"
);
op
->
SetInput
(
"X"
,
Input
(
"X"
));
op
->
SetInput
(
"Mean"
,
Output
(
"Mean"
));
op
->
SetInput
(
"Variance"
,
Output
(
"Variance"
));
if
(
ForwardOp
().
Inputs
().
count
(
"Scale"
)
>
0
)
{
op
->
SetInput
(
"Scale"
,
Input
(
"Scale"
));
op
->
SetOutput
(
framework
::
GradVarName
(
"Scale"
),
InputGrad
(
"Scale"
));
}
if
(
ForwardOp
().
Inputs
().
count
(
"Bias"
)
>
0
)
{
op
->
SetOutput
(
framework
::
GradVarName
(
"Bias"
),
InputGrad
(
"Bias"
));
}
op
->
SetInput
(
framework
::
GradVarName
(
"Y"
),
OutputGrad
(
"Y"
));
op
->
SetOutput
(
framework
::
GradVarName
(
"X"
),
InputGrad
(
"X"
));
op
->
SetAttrMap
(
Attrs
());
return
op
;
}
};
}
// namespace operators
}
// namespace paddle
namespace
ops
=
paddle
::
operators
;
REGISTER_OPERATOR
(
layer_norm
,
ops
::
LayerNormOp
,
ops
::
LayerNormOpMaker
,
paddle
::
framework
::
DefaultGradOpDescMaker
<
true
>
);
ops
::
LayerNormGradOpDescMaker
);
REGISTER_OPERATOR
(
layer_norm_grad
,
ops
::
LayerNormGradOp
);
REGISTER_OP_CPU_KERNEL
(
layer_norm
,
ops
::
LayerNormKernel
<
paddle
::
platform
::
CPUDeviceContext
,
float
>
,
...
...
paddle/fluid/operators/layer_norm_op.h
浏览文件 @
d0ef6825
...
...
@@ -245,11 +245,9 @@ class LayerNormGradKernel : public framework::OpKernel<T> {
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
const
float
epsilon
=
ctx
.
Attr
<
float
>
(
"epsilon"
);
auto
x
=
*
ctx
.
Input
<
Tensor
>
(
"X"
);
auto
*
y
=
ctx
.
Input
<
Tensor
>
(
"Y"
);
auto
*
mean
=
ctx
.
Input
<
Tensor
>
(
"Mean"
);
auto
*
var
=
ctx
.
Input
<
Tensor
>
(
"Variance"
);
auto
*
scale
=
ctx
.
Input
<
Tensor
>
(
"Scale"
);
auto
*
bias
=
ctx
.
Input
<
Tensor
>
(
"Bias"
);
auto
d_y
=
*
ctx
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"Y"
));
const
auto
begin_norm_axis
=
ctx
.
Attr
<
int
>
(
"begin_norm_axis"
);
...
...
@@ -275,18 +273,13 @@ class LayerNormGradKernel : public framework::OpKernel<T> {
x
.
Resize
(
matrix_shape
);
temp
.
mutable_data
<
T
>
(
matrix_shape
,
ctx
.
GetPlace
());
if
(
!
(
bias
&&
scale
))
{
temp_norm
.
ShareDataWith
(
*
y
);
temp_norm
.
Resize
(
matrix_shape
);
}
else
{
temp_norm
.
mutable_data
<
T
>
(
matrix_shape
,
ctx
.
GetPlace
());
// get x_norm
ElementwiseComputeEx
<
SubFunctor
<
T
>
,
DeviceContext
,
T
>
(
ctx
,
&
x
,
mean
,
/*axis*/
0
,
SubFunctor
<
T
>
(),
&
temp_norm
);
ElementwiseComputeEx
<
DivAndSqrtFunctor
<
T
>
,
DeviceContext
,
T
>
(
ctx
,
&
temp_norm
,
var
,
/*axis*/
0
,
DivAndSqrtFunctor
<
T
>
(
static_cast
<
T
>
(
epsilon
)),
&
temp_norm
);
}
temp_norm
.
mutable_data
<
T
>
(
matrix_shape
,
ctx
.
GetPlace
());
// get x_norm
ElementwiseComputeEx
<
SubFunctor
<
T
>
,
DeviceContext
,
T
>
(
ctx
,
&
x
,
mean
,
/*axis*/
0
,
SubFunctor
<
T
>
(),
&
temp_norm
);
ElementwiseComputeEx
<
DivAndSqrtFunctor
<
T
>
,
DeviceContext
,
T
>
(
ctx
,
&
temp_norm
,
var
,
/*axis*/
0
,
DivAndSqrtFunctor
<
T
>
(
static_cast
<
T
>
(
epsilon
)),
&
temp_norm
);
}
if
(
d_bias
)
{
...
...
paddle/fluid/operators/softmax_with_cross_entropy_op.cc
浏览文件 @
d0ef6825
...
...
@@ -13,6 +13,7 @@ See the License for the specific language governing permissions and
limitations under the License. */
#include "paddle/fluid/operators/softmax_with_cross_entropy_op.h"
#include <memory>
namespace
paddle
{
namespace
operators
{
...
...
@@ -187,7 +188,6 @@ class SoftmaxGradMaker : public framework::SingleGradOpDescMaker {
grad_op
->
SetType
(
"softmax_with_cross_entropy_grad"
);
grad_op
->
SetInput
(
"Label"
,
Input
(
"Label"
));
grad_op
->
SetInput
(
"Softmax"
,
Output
(
"Softmax"
));
grad_op
->
SetInput
(
"Loss"
,
Output
(
"Loss"
));
grad_op
->
SetInput
(
framework
::
GradVarName
(
"Softmax"
),
OutputGrad
(
"Softmax"
));
grad_op
->
SetInput
(
framework
::
GradVarName
(
"Loss"
),
OutputGrad
(
"Loss"
));
grad_op
->
SetOutput
(
framework
::
GradVarName
(
"Logits"
),
InputGrad
(
"Logits"
));
...
...
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