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
Paddle
提交
d7f422c9
P
Paddle
项目概览
PaddlePaddle
/
Paddle
大约 1 年 前同步成功
通知
2297
Star
20931
Fork
5422
代码
文件
提交
分支
Tags
贡献者
分支图
Diff
Issue
1423
列表
看板
标记
里程碑
合并请求
543
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
1,423
Issue
1,423
列表
看板
标记
里程碑
合并请求
543
合并请求
543
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
d7f422c9
编写于
9月 24, 2020
作者:
K
Kaipeng Deng
提交者:
GitHub
9月 24, 2020
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
fix error message in conv/conv_transpose. test=develop (#27464)
* fix error message in conv/conv_transpose. test=develop
上级
59c04999
变更
4
显示空白变更内容
内联
并排
Showing
4 changed file
with
70 addition
and
40 deletion
+70
-40
paddle/fluid/operators/conv_cudnn_op.cu
paddle/fluid/operators/conv_cudnn_op.cu
+36
-24
paddle/fluid/operators/conv_op.h
paddle/fluid/operators/conv_op.h
+14
-4
paddle/fluid/operators/conv_transpose_cudnn_op.cu
paddle/fluid/operators/conv_transpose_cudnn_op.cu
+10
-10
paddle/fluid/operators/conv_transpose_op.h
paddle/fluid/operators/conv_transpose_op.h
+10
-2
未找到文件。
paddle/fluid/operators/conv_cudnn_op.cu
浏览文件 @
d7f422c9
...
...
@@ -50,8 +50,9 @@ class CUDNNConvOpKernel : public framework::OpKernel<T> {
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
&
dev_ctx
=
ctx
.
template
device_context
<
platform
::
CUDADeviceContext
>();
PADDLE_ENFORCE_EQ
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()),
true
,
"It must use CUDAPlace."
);
PADDLE_ENFORCE_EQ
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()),
true
,
paddle
::
platform
::
errors
::
PreconditionNotMet
(
"It must use CUDAPlace."
));
const
Tensor
*
input
=
ctx
.
Input
<
Tensor
>
(
"Input"
);
auto
*
filter
=
ctx
.
Input
<
Tensor
>
(
"Filter"
);
auto
*
output
=
ctx
.
Output
<
Tensor
>
(
"Output"
);
...
...
@@ -60,14 +61,16 @@ class CUDNNConvOpKernel : public framework::OpKernel<T> {
std
::
vector
<
int
>
paddings
=
ctx
.
Attr
<
std
::
vector
<
int
>>
(
"paddings"
);
std
::
vector
<
int
>
dilations
=
ctx
.
Attr
<
std
::
vector
<
int
>>
(
"dilations"
);
int
groups
=
ctx
.
Attr
<
int
>
(
"groups"
);
bool
exhaustive_search
=
FLAGS_cudnn_exhaustive_search
||
ctx
.
Attr
<
bool
>
(
"exhaustive_search"
);
if
(
exhaustive_search
&&
FLAGS_cudnn_deterministic
)
{
PADDLE_THROW
(
bool
deterministic
=
FLAGS_cudnn_deterministic
;
auto
exhaustive_deterministic
=
exhaustive_search
&&
deterministic
;
PADDLE_ENFORCE_EQ
(
exhaustive_deterministic
,
false
,
platform
::
errors
::
InvalidArgument
(
"Cann't set exhaustive_search True and "
"FLAGS_cudnn_deterministic True at same time."
);
}
"FLAGS_cudnn_deterministic True at same time."
)
);
const
std
::
string
padding_algorithm
=
ctx
.
Attr
<
std
::
string
>
(
"padding_algorithm"
);
const
std
::
string
data_format
=
ctx
.
Attr
<
std
::
string
>
(
"data_format"
);
...
...
@@ -197,7 +200,8 @@ class CUDNNConvOpKernel : public framework::OpKernel<T> {
&
transformed_input
);
}
break
;
default:
PADDLE_THROW
(
"ConvOp only support tensors with 4 or 5 dimensions."
);
PADDLE_THROW
(
platform
::
errors
::
InvalidArgument
(
"ConvOp only support tensors with 4 or 5 dimensions."
));
}
}
else
{
...
...
@@ -317,8 +321,9 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
&
dev_ctx
=
ctx
.
template
device_context
<
platform
::
CUDADeviceContext
>();
PADDLE_ENFORCE_EQ
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()),
true
,
"It must use CUDAPlace."
);
PADDLE_ENFORCE_EQ
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()),
true
,
paddle
::
platform
::
errors
::
PreconditionNotMet
(
"It must use CUDAPlace."
));
auto
input
=
ctx
.
Input
<
Tensor
>
(
"Input"
);
auto
filter
=
ctx
.
Input
<
Tensor
>
(
"Filter"
);
auto
output_grad
=
ctx
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"Output"
));
...
...
@@ -337,14 +342,16 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
std
::
vector
<
int
>
paddings
=
ctx
.
Attr
<
std
::
vector
<
int
>>
(
"paddings"
);
std
::
string
padding_algorithm
=
ctx
.
Attr
<
std
::
string
>
(
"padding_algorithm"
);
int
groups
=
ctx
.
Attr
<
int
>
(
"groups"
);
bool
exhaustive_search
=
FLAGS_cudnn_exhaustive_search
||
ctx
.
Attr
<
bool
>
(
"exhaustive_search"
);
bool
deterministic
=
FLAGS_cudnn_deterministic
;
if
(
exhaustive_search
&&
deterministic
)
{
PADDLE_THROW
(
"Can't set exhaustive_search True and "
"FLAGS_cudnn_deterministic True at same time."
);
}
auto
exhaustive_deterministic
=
exhaustive_search
&&
deterministic
;
PADDLE_ENFORCE_EQ
(
exhaustive_deterministic
,
false
,
platform
::
errors
::
InvalidArgument
(
"Cann't set exhaustive_search True and "
"FLAGS_cudnn_deterministic True at same time."
));
const
std
::
string
data_format
=
ctx
.
Attr
<
std
::
string
>
(
"data_format"
);
const
bool
channel_last
=
(
data_format
==
"NHWC"
||
data_format
==
"NDHWC"
);
...
...
@@ -495,7 +502,8 @@ class CUDNNConvGradOpKernel : public framework::OpKernel<T> {
&
transformed_input
);
}
break
;
default:
PADDLE_THROW
(
"ConvOp only support tensors with 4 or 5 dimensions."
);
PADDLE_THROW
(
platform
::
errors
::
InvalidArgument
(
"ConvOp only support tensors with 4 or 5 dimensions."
));
}
}
else
{
transformed_input
.
ShareDataWith
(
transformed_input_channel
);
...
...
@@ -701,8 +709,9 @@ class CUDNNConvDoubleGradOpKernel : public framework::OpKernel<T> {
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
&
dev_ctx
=
ctx
.
template
device_context
<
platform
::
CUDADeviceContext
>();
PADDLE_ENFORCE_EQ
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()),
true
,
"It must use CUDAPlace."
);
PADDLE_ENFORCE_EQ
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()),
true
,
paddle
::
platform
::
errors
::
PreconditionNotMet
(
"It must use CUDAPlace."
));
auto
X
=
ctx
.
Input
<
Tensor
>
(
"Input"
);
auto
W
=
ctx
.
Input
<
Tensor
>
(
"Filter"
);
auto
dO
=
ctx
.
Input
<
Tensor
>
(
"DOutput"
);
...
...
@@ -736,14 +745,16 @@ class CUDNNConvDoubleGradOpKernel : public framework::OpKernel<T> {
const
std
::
vector
<
int
>&
strides
=
ctx
.
Attr
<
std
::
vector
<
int
>>
(
"strides"
);
std
::
vector
<
int
>
dilations
=
ctx
.
Attr
<
std
::
vector
<
int
>>
(
"dilations"
);
int
groups
=
ctx
.
Attr
<
int
>
(
"groups"
);
bool
exhaustive_search
=
FLAGS_cudnn_exhaustive_search
||
ctx
.
Attr
<
bool
>
(
"exhaustive_search"
);
bool
deterministic
=
FLAGS_cudnn_deterministic
;
if
(
exhaustive_search
&&
deterministic
)
{
PADDLE_THROW
(
"Can't set exhaustive_search True and "
"FLAGS_cudnn_deterministic True at same time."
);
}
auto
exhaustive_deterministic
=
exhaustive_search
&&
deterministic
;
PADDLE_ENFORCE_EQ
(
exhaustive_deterministic
,
false
,
platform
::
errors
::
InvalidArgument
(
"Cann't set exhaustive_search True and "
"FLAGS_cudnn_deterministic True at same time."
));
std
::
vector
<
int
>
paddings
=
ctx
.
Attr
<
std
::
vector
<
int
>>
(
"paddings"
);
std
::
string
padding_algorithm
=
ctx
.
Attr
<
std
::
string
>
(
"padding_algorithm"
);
...
...
@@ -878,7 +889,8 @@ class CUDNNConvDoubleGradOpKernel : public framework::OpKernel<T> {
}
}
break
;
default:
PADDLE_THROW
(
"ConvOp only support tensors with 4 or 5 dimensions."
);
PADDLE_THROW
(
platform
::
errors
::
InvalidArgument
(
"ConvOp only support tensors with 4 or 5 dimensions."
));
}
}
else
{
...
...
paddle/fluid/operators/conv_op.h
浏览文件 @
d7f422c9
...
...
@@ -685,8 +685,9 @@ class GemmConvDoubleGradKernel : public framework::OpKernel<T> {
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
auto
&
dev_ctx
=
ctx
.
template
device_context
<
platform
::
CPUDeviceContext
>();
PADDLE_ENFORCE_EQ
(
platform
::
is_cpu_place
(
ctx
.
GetPlace
()),
true
,
"It must use CPUPlace."
);
PADDLE_ENFORCE_EQ
(
platform
::
is_cpu_place
(
ctx
.
GetPlace
()),
true
,
paddle
::
platform
::
errors
::
PreconditionNotMet
(
"It must use CPUPlace."
));
const
Tensor
*
X
=
ctx
.
Input
<
Tensor
>
(
"Input"
);
const
Tensor
*
dY
=
ctx
.
Input
<
Tensor
>
(
"DOutput"
);
const
Tensor
*
ddX
=
ctx
.
Input
<
Tensor
>
(
"DDInput"
);
...
...
@@ -982,11 +983,20 @@ class DepthwiseConvKernel : public framework::OpKernel<T> {
PADDLE_ENFORCE_EQ
(
output
->
dims
()[
output
->
dims
().
size
()
-
1
]
%
input
->
dims
()[
input
->
dims
().
size
()
-
1
],
0
,
"The output channels must be a multiple of the input channels"
);
0
,
platform
::
errors
::
InvalidArgument
(
"ShapeError: The output channels must be a multiple of the "
"input channels. But receivced output channel number is %d "
"and input channel number is %d"
,
output
->
dims
()[
output
->
dims
().
size
()
-
1
],
input
->
dims
()[
input
->
dims
().
size
()
-
1
]));
}
else
{
PADDLE_ENFORCE_EQ
(
output
->
dims
()[
1
]
%
input
->
dims
()[
1
],
0
,
"The output channels must be a multiple of the input channels"
);
platform
::
errors
::
InvalidArgument
(
"ShapeError: The output channels must be a multiple of the "
"input channels. But receivced output channel number is %d "
"and input channel number is %d"
,
output
->
dims
()[
1
],
input
->
dims
()[
1
]));
}
// transform tensor
Tensor
transformed_input
(
input
->
type
());
...
...
paddle/fluid/operators/conv_transpose_cudnn_op.cu
浏览文件 @
d7f422c9
...
...
@@ -51,8 +51,9 @@ template <typename T>
class
CUDNNConvTransposeOpKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
PADDLE_ENFORCE_EQ
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()),
true
,
"It must use CUDAPlace."
);
PADDLE_ENFORCE_EQ
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()),
true
,
paddle
::
platform
::
errors
::
PreconditionNotMet
(
"It must use CUDAPlace."
));
auto
*
input
=
ctx
.
Input
<
Tensor
>
(
"Input"
);
auto
*
filter
=
ctx
.
Input
<
Tensor
>
(
"Filter"
);
auto
*
output
=
ctx
.
Output
<
Tensor
>
(
"Output"
);
...
...
@@ -145,9 +146,8 @@ class CUDNNConvTransposeOpKernel : public framework::OpKernel<T> {
ctx
,
input_pad
,
input_transpose
,
pad_value
,
&
transformed_input
);
}
break
;
default:
PADDLE_ENFORCE_EQ
(
rank
==
4
||
rank
==
5
,
true
,
"Op(ConvTranspose) only supports 4-D or 5-D input Tensor."
);
PADDLE_THROW
(
platform
::
errors
::
InvalidArgument
(
"Op(ConvTranspose) only supports 4-D or 5-D input Tensor."
));
}
}
else
{
transformed_input
=
input_transpose
;
...
...
@@ -290,8 +290,9 @@ template <typename T>
class
CUDNNConvTransposeGradOpKernel
:
public
framework
::
OpKernel
<
T
>
{
public:
void
Compute
(
const
framework
::
ExecutionContext
&
ctx
)
const
override
{
PADDLE_ENFORCE
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()),
"It must use CUDAPlace."
);
PADDLE_ENFORCE_EQ
(
platform
::
is_gpu_place
(
ctx
.
GetPlace
()),
true
,
paddle
::
platform
::
errors
::
PreconditionNotMet
(
"It must use CUDAPlace."
));
auto
input
=
ctx
.
Input
<
Tensor
>
(
"Input"
);
auto
filter
=
ctx
.
Input
<
Tensor
>
(
"Filter"
);
auto
output_grad
=
ctx
.
Input
<
Tensor
>
(
framework
::
GradVarName
(
"Output"
));
...
...
@@ -393,9 +394,8 @@ class CUDNNConvTransposeGradOpKernel : public framework::OpKernel<T> {
&
transformed_output_grad
);
}
break
;
default:
PADDLE_ENFORCE_EQ
(
rank
==
4
||
rank
==
5
,
true
,
"Op(ConvTranspose) only supports 4-D or 5-D input Tensor."
);
PADDLE_THROW
(
platform
::
errors
::
InvalidArgument
(
"Op(ConvTranspose) only supports 4-D or 5-D input Tensor."
));
}
}
else
{
transformed_output_grad
=
output_grad_transpose
;
...
...
paddle/fluid/operators/conv_transpose_op.h
浏览文件 @
d7f422c9
...
...
@@ -580,7 +580,12 @@ class DepthwiseConvTransposeKernel : public framework::OpKernel<T> {
output
->
mutable_data
<
T
>
(
context
.
GetPlace
());
int
groups
=
context
.
Attr
<
int
>
(
"groups"
);
PADDLE_ENFORCE_EQ
(
groups
,
filter
.
dims
()[
0
]);
PADDLE_ENFORCE_EQ
(
groups
,
filter
.
dims
()[
0
],
platform
::
errors
::
InvalidArgument
(
"groups should be error to the 1st dimension of filter. But "
"received groups is %d and filter dimension[0] is %d"
,
groups
,
filter
.
dims
()[
0
]));
std
::
vector
<
int
>
strides
=
context
.
Attr
<
std
::
vector
<
int
>>
(
"strides"
);
std
::
vector
<
int
>
paddings
=
context
.
Attr
<
std
::
vector
<
int
>>
(
"paddings"
);
...
...
@@ -588,7 +593,10 @@ class DepthwiseConvTransposeKernel : public framework::OpKernel<T> {
std
::
string
padding_algorithm
=
context
.
Attr
<
std
::
string
>
(
"padding_algorithm"
);
for
(
auto
v
:
dilations
)
{
PADDLE_ENFORCE_EQ
(
v
,
1
);
PADDLE_ENFORCE_EQ
(
v
,
1
,
platform
::
errors
::
InvalidArgument
(
"dilations should be 1 in depthwise conv. "
"But received dilations is %d"
,
v
));
}
auto
in_dims
=
input
->
dims
();
...
...
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