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
98ab2433
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
98ab2433
编写于
12月 09, 2022
作者:
Z
Zhang Jun
提交者:
GitHub
12月 09, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[inference][trt] upgrade prelu op (#48528)
* add prelu
上级
c1cadcca
变更
2
隐藏空白更改
内联
并排
Showing
2 changed file
with
81 addition
and
44 deletion
+81
-44
paddle/fluid/inference/tensorrt/convert/prelu_op.cc
paddle/fluid/inference/tensorrt/convert/prelu_op.cc
+75
-38
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_prelu.py
...id/tests/unittests/ir/inference/test_trt_convert_prelu.py
+6
-6
未找到文件。
paddle/fluid/inference/tensorrt/convert/prelu_op.cc
浏览文件 @
98ab2433
...
@@ -31,8 +31,8 @@ class PReluOpConverter : public OpConverter {
...
@@ -31,8 +31,8 @@ class PReluOpConverter : public OpConverter {
framework
::
OpDesc
op_desc
(
op
,
nullptr
);
framework
::
OpDesc
op_desc
(
op
,
nullptr
);
// Declare inputs
// Declare inputs
size_t
input_num
=
op_desc
.
Input
(
"X"
).
size
();
auto
*
input
=
engine_
->
GetITensor
(
op_desc
.
Input
(
"X"
)[
0
]);
auto
*
input
=
engine_
->
GetITensor
(
op_desc
.
Input
(
"X"
)[
0
]);
auto
input_dims
=
input
->
getDimensions
();
// Get attrs
// Get attrs
std
::
string
mode
=
PADDLE_GET_CONST
(
std
::
string
,
op_desc
.
GetAttr
(
"mode"
));
std
::
string
mode
=
PADDLE_GET_CONST
(
std
::
string
,
op_desc
.
GetAttr
(
"mode"
));
std
::
string
data_format
=
"NCHW"
;
std
::
string
data_format
=
"NCHW"
;
...
@@ -40,50 +40,87 @@ class PReluOpConverter : public OpConverter {
...
@@ -40,50 +40,87 @@ class PReluOpConverter : public OpConverter {
data_format
=
data_format
=
PADDLE_GET_CONST
(
std
::
string
,
op_desc
.
GetAttr
(
"data_format"
));
PADDLE_GET_CONST
(
std
::
string
,
op_desc
.
GetAttr
(
"data_format"
));
}
}
auto
*
alpha_var
=
scope
.
FindVar
(
op_desc
.
Input
(
"Alpha"
)[
0
]);
auto
*
alpha_tensor
=
alpha_var
->
GetMutable
<
phi
::
DenseTensor
>
();
auto
alpha_weight
=
auto
*
alpha_var
=
scope
.
FindVar
(
op_desc
.
Input
(
"Alpha"
)[
0
]);
engine_
->
GetFp32TrtWeight
(
op_desc
.
Input
(
"Alpha"
)[
0
],
*
alpha_tensor
);
auto
*
alpha_weight
=
alpha_var
->
GetMutable
<
phi
::
DenseTensor
>
();
auto
w_dims
=
alpha_weight
->
dims
();
auto
alpha_data
=
engine_
->
GetFp32TrtWeight
(
op_desc
.
Input
(
"Alpha"
)[
0
],
*
alpha_weight
);
platform
::
CPUPlace
cpu_place
;
nvinfer1
::
Dims
trt_w_dims
;
trt_w_dims
.
nbDims
=
w_dims
.
size
();
for
(
int
i
=
0
;
i
<
trt_w_dims
.
nbDims
;
i
++
)
{
trt_w_dims
.
d
[
i
]
=
w_dims
[
i
];
}
nvinfer1
::
ILayer
*
layer
=
nullptr
;
// The `element` or `channel` mode contains the batch using static shape.
if
(
engine_
->
with_dynamic_shape
())
{
if
((
mode
==
"element"
||
mode
==
"channel"
)
&&
plugin
::
PReluPluginDynamic
*
plugin
=
new
plugin
::
PReluPluginDynamic
(
!
engine_
->
with_dynamic_shape
()
&&
static_cast
<
const
float
*>
(
alpha_weight
.
get
().
values
),
(
trt_w_dims
.
nbDims
-
1
==
input_dims
.
nbDims
))
{
alpha_tensor
->
numel
(),
trt_w_dims
.
nbDims
--
;
mode
,
for
(
int
i
=
0
;
i
<
trt_w_dims
.
nbDims
;
i
++
)
{
data_format
);
trt_w_dims
.
d
[
i
]
=
trt_w_dims
.
d
[
i
+
1
];
layer
=
engine_
->
AddDynamicPlugin
(
&
input
,
input_num
,
plugin
);
}
else
{
#if IS_TRT_VERSION_GE(7000)
nvinfer1
::
Dims
dims
;
dims
.
nbDims
=
0
;
// jump batch dim
for
(
int
i
=
1
;
i
<
alpha_tensor
->
dims
().
size
();
i
++
)
{
dims
.
d
[
dims
.
nbDims
++
]
=
alpha_tensor
->
dims
()[
i
];
}
for
(;
dims
.
nbDims
<
input
->
getDimensions
().
nbDims
;
dims
.
nbDims
++
)
{
dims
.
d
[
dims
.
nbDims
]
=
1
;
}
}
}
auto
alpha_layer
=
nvinfer1
::
ITensor
*
alpha_tensor
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Constant
,
dims
,
alpha_weight
.
get
());
TRT_ENGINE_ADD_LAYER
(
engine_
,
Constant
,
trt_w_dims
,
alpha_data
.
get
())
auto
alpha_layer_output
=
alpha_layer
->
getOutput
(
0
);
->
getOutput
(
0
);
layer
=
TRT_ENGINE_ADD_LAYER
(
auto
alpha_dims
=
alpha_tensor
->
getDimensions
();
engine_
,
ParametricReLU
,
*
input
,
*
alpha_layer_output
);
nvinfer1
::
ITensor
*
real_alpha_tensor
=
alpha_tensor
;
#else
if
(
alpha_dims
.
nbDims
!=
input_dims
.
nbDims
)
{
plugin
::
PReluPlugin
*
plugin
=
new
plugin
::
PReluPlugin
(
auto
*
reshape_layer
=
static_cast
<
const
float
*>
(
alpha_weight
.
get
().
values
),
TRT_ENGINE_ADD_LAYER
(
engine_
,
Shuffle
,
*
alpha_tensor
);
alpha_tensor
->
numel
(),
int
c
=
alpha_dims
.
d
[
0
];
mode
,
if
(
engine_
->
with_dynamic_shape
())
{
data_format
);
std
::
vector
<
nvinfer1
::
ITensor
*>
itensors
;
layer
=
engine_
->
AddPlugin
(
&
input
,
input_num
,
plugin
);
auto
*
n_tensor
=
Add1DConstantLayer
(
1
);
#endif
auto
*
c_tensor
=
Add1DConstantLayer
(
c
);
nvinfer1
::
ITensor
*
hw_tensor
=
nullptr
;
nvinfer1
::
ITensor
*
shape_tensor
=
nullptr
;
if
(
input_dims
.
nbDims
-
2
>
0
)
{
hw_tensor
=
Add1DConstantLayer
(
std
::
vector
<
int32_t
>
(
input_dims
.
nbDims
-
2
,
1
));
}
if
(
data_format
==
"NCHW"
)
{
if
(
hw_tensor
!=
nullptr
)
{
shape_tensor
=
Concat
(
std
::
vector
<
nvinfer1
::
ITensor
*>
{
n_tensor
,
c_tensor
,
hw_tensor
});
}
else
{
shape_tensor
=
Concat
(
std
::
vector
<
nvinfer1
::
ITensor
*>
{
n_tensor
,
c_tensor
});
}
}
else
{
if
(
hw_tensor
!=
nullptr
)
{
shape_tensor
=
Concat
(
std
::
vector
<
nvinfer1
::
ITensor
*>
{
n_tensor
,
hw_tensor
,
c_tensor
});
}
else
{
shape_tensor
=
Concat
(
std
::
vector
<
nvinfer1
::
ITensor
*>
{
n_tensor
,
c_tensor
});
}
}
reshape_layer
->
setInput
(
1
,
*
shape_tensor
);
}
else
{
nvinfer1
::
Dims
reshape_dim
;
reshape_dim
.
nbDims
=
input_dims
.
nbDims
;
std
::
fill
(
reshape_dim
.
d
,
reshape_dim
.
d
+
input_dims
.
nbDims
,
1
);
if
(
data_format
==
"NCHW"
)
{
reshape_dim
.
d
[
0
]
=
c
;
}
else
if
(
data_format
==
"NHWC"
)
{
reshape_dim
.
d
[
input_dims
.
nbDims
-
1
]
=
c
;
}
reshape_layer
->
setReshapeDimensions
(
reshape_dim
);
}
real_alpha_tensor
=
reshape_layer
->
getOutput
(
0
);
}
}
nvinfer1
::
ILayer
*
layer
=
nullptr
;
layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
ParametricReLU
,
*
input
,
*
real_alpha_tensor
);
auto
output_name
=
op_desc
.
Output
(
"Out"
)[
0
];
auto
output_name
=
op_desc
.
Output
(
"Out"
)[
0
];
RreplenishLayerAndOutput
(
layer
,
"prelu"
,
{
output_name
},
test_mode
);
RreplenishLayerAndOutput
(
layer
,
"prelu"
,
{
output_name
},
test_mode
);
}
}
...
...
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_prelu.py
浏览文件 @
98ab2433
...
@@ -49,22 +49,22 @@ class TrtConvertPreluTest(TrtLayerAutoScanTest):
...
@@ -49,22 +49,22 @@ class TrtConvertPreluTest(TrtLayerAutoScanTest):
if
dim1
!=
0
:
if
dim1
!=
0
:
shape
.
append
(
dim1
)
shape
.
append
(
dim1
)
if
dim2
!=
0
:
if
dim2
!=
0
:
shape
.
append
(
1
)
shape
.
append
(
dim2
)
if
dim3
!=
0
:
if
dim3
!=
0
:
shape
.
append
(
1
)
shape
.
append
(
dim3
)
return
np
.
random
.
random
(
size
=
shape
).
astype
(
np
.
float32
)
return
np
.
random
.
random
(
size
=
shape
[
1
]
).
astype
(
np
.
float32
)
elif
(
elif
(
attrs
[
0
][
"mode"
]
==
"channel"
attrs
[
0
][
"mode"
]
==
"channel"
and
attrs
[
0
][
"data_format"
]
==
"NHWC"
and
attrs
[
0
][
"data_format"
]
==
"NHWC"
):
):
shape
=
[
1
]
shape
=
[
1
]
if
dim1
!=
0
:
if
dim1
!=
0
:
shape
.
append
(
1
)
shape
.
append
(
dim
1
)
if
dim2
!=
0
:
if
dim2
!=
0
:
shape
.
append
(
1
)
shape
.
append
(
dim2
)
if
dim3
!=
0
:
if
dim3
!=
0
:
shape
.
append
(
dim3
)
shape
.
append
(
dim3
)
return
np
.
random
.
random
(
size
=
shape
).
astype
(
np
.
float32
)
return
np
.
random
.
random
(
size
=
shape
[
-
1
]
).
astype
(
np
.
float32
)
elif
attrs
[
0
][
"mode"
]
==
"element"
:
elif
attrs
[
0
][
"mode"
]
==
"element"
:
shape
=
[
1
]
shape
=
[
1
]
if
dim1
!=
0
:
if
dim1
!=
0
:
...
...
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