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
dd304f31
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
dd304f31
编写于
12月 06, 2022
作者:
Z
Zhang Jun
提交者:
GitHub
12月 06, 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] add reduce max for trt (#48684)
* add reduce max for trt
上级
0c7f3575
变更
5
隐藏空白更改
内联
并排
Showing
5 changed file
with
83 addition
and
215 deletion
+83
-215
paddle/fluid/inference/api/analysis_predictor.cc
paddle/fluid/inference/api/analysis_predictor.cc
+2
-1
paddle/fluid/inference/tensorrt/convert/reduce_op.cc
paddle/fluid/inference/tensorrt/convert/reduce_op.cc
+28
-10
paddle/fluid/inference/tensorrt/op_teller.cc
paddle/fluid/inference/tensorrt/op_teller.cc
+6
-3
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_reduce.py
...d/tests/unittests/ir/inference/test_trt_convert_reduce.py
+47
-40
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_reduce_sum.py
...sts/unittests/ir/inference/test_trt_convert_reduce_sum.py
+0
-161
未找到文件。
paddle/fluid/inference/api/analysis_predictor.cc
浏览文件 @
dd304f31
...
...
@@ -2318,9 +2318,10 @@ USE_TRT_CONVERTER(nearest_interp_v2);
USE_TRT_CONVERTER
(
bilinear_interp_v2
);
USE_TRT_CONVERTER
(
reshape
);
USE_TRT_CONVERTER
(
reshape2
);
USE_TRT_CONVERTER
(
reduce_sum
);
USE_TRT_CONVERTER
(
gather_nd
);
USE_TRT_CONVERTER
(
reduce_mean
);
USE_TRT_CONVERTER
(
reduce_max
);
USE_TRT_CONVERTER
(
reduce_sum
);
USE_TRT_CONVERTER
(
tile
);
USE_TRT_CONVERTER
(
conv3d
);
USE_TRT_CONVERTER
(
conv3d_transpose
);
...
...
paddle/fluid/inference/tensorrt/convert/reduce_op.cc
浏览文件 @
dd304f31
...
...
@@ -42,12 +42,7 @@ class ReduceOpConverter : public OpConverter {
bool
test_mode
)
override
{
VLOG
(
4
)
<<
"convert a paddle "
<<
op_type
<<
" op to tensorrt reduce layer"
;
framework
::
OpDesc
op_desc
(
op
,
nullptr
);
nvinfer1
::
ReduceOperation
reduce_type
=
nvinfer1
::
ReduceOperation
::
kSUM
;
if
(
op_type
==
"reduce_sum"
)
{
reduce_type
=
nvinfer1
::
ReduceOperation
::
kSUM
;
}
else
if
(
op_type
==
"reduce_mean"
)
{
reduce_type
=
nvinfer1
::
ReduceOperation
::
kAVG
;
}
auto
reduce_type
=
ops_
.
find
(
op_type
);
auto
*
x
=
engine_
->
GetITensor
(
op_desc
.
Input
(
"X"
).
front
());
nvinfer1
::
Dims
input_shape
=
x
->
getDimensions
();
...
...
@@ -64,8 +59,12 @@ class ReduceOpConverter : public OpConverter {
for
(
int
i
=
0
;
i
<
input_dims
;
++
i
)
{
reduce_dim
|=
1
<<
i
;
}
layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Reduce
,
*
x
,
reduce_type
,
reduce_dim
,
keep_dim
);
layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Reduce
,
*
x
,
reduce_type
->
second
.
front
(),
reduce_dim
,
keep_dim
);
}
else
{
auto
CvtToBitMask
=
[
&
](
const
std
::
vector
<
int32_t
>&
dims
)
->
uint32_t
{
uint32_t
res
=
0
;
...
...
@@ -79,8 +78,12 @@ class ReduceOpConverter : public OpConverter {
}
return
res
;
};
layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Reduce
,
*
x
,
reduce_type
,
CvtToBitMask
(
dim
),
keep_dim
);
layer
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Reduce
,
*
x
,
reduce_type
->
second
.
front
(),
CvtToBitMask
(
dim
),
keep_dim
);
}
auto
output_name
=
op_desc
.
Output
(
"Out"
)[
0
];
...
...
@@ -91,6 +94,16 @@ class ReduceOpConverter : public OpConverter {
protected:
std
::
string
op_type
;
static
const
std
::
unordered_map
<
std
::
string
,
std
::
vector
<
nvinfer1
::
ReduceOperation
>>
ops_
;
};
const
std
::
unordered_map
<
std
::
string
,
std
::
vector
<
nvinfer1
::
ReduceOperation
>>
ReduceOpConverter
::
ops_
=
{
{
"reduce_mean"
,
{
nvinfer1
::
ReduceOperation
::
kAVG
}},
{
"reduce_sum"
,
{
nvinfer1
::
ReduceOperation
::
kSUM
}},
{
"reduce_max"
,
{
nvinfer1
::
ReduceOperation
::
kMAX
}},
};
class
ReduceSumOpConverter
:
public
ReduceOpConverter
{
...
...
@@ -103,9 +116,14 @@ class ReduceMeanOpConverter : public ReduceOpConverter {
ReduceMeanOpConverter
()
{
op_type
=
"reduce_mean"
;
}
};
class
ReduceMaxOpConverter
:
public
ReduceOpConverter
{
public:
ReduceMaxOpConverter
()
{
op_type
=
"reduce_max"
;
}
};
}
// namespace tensorrt
}
// namespace inference
}
// namespace paddle
REGISTER_TRT_OP_CONVERTER
(
reduce_sum
,
ReduceSumOpConverter
);
REGISTER_TRT_OP_CONVERTER
(
reduce_mean
,
ReduceMeanOpConverter
);
REGISTER_TRT_OP_CONVERTER
(
reduce_max
,
ReduceMaxOpConverter
);
paddle/fluid/inference/tensorrt/op_teller.cc
浏览文件 @
dd304f31
...
...
@@ -2038,7 +2038,8 @@ struct SimpleOpTypeSetTeller : public Teller {
const
auto
x_shape
=
x_var_desc
->
GetShape
();
}
if
(
op_type
==
"reduce_sum"
||
op_type
==
"reduce_mean"
)
{
if
(
op_type
==
"reduce_sum"
||
op_type
==
"reduce_mean"
||
op_type
==
"reduce_max"
)
{
if
(
!
desc
.
HasAttr
(
"dim"
,
/*with_attr_var=*/
false
))
{
VLOG
(
3
)
<<
"Skip to convert into TRT while found Attribute('dim') is "
"Variable type in "
...
...
@@ -2470,8 +2471,9 @@ struct SimpleOpTypeSetTeller : public Teller {
"affine_channel"
,
"nearest_interp"
,
"anchor_generator"
,
"reduce_
sum
"
,
"reduce_
max
"
,
"reduce_mean"
,
"reduce_sum"
,
"conv3d"
,
"conv3d_transpose"
,
"mish"
,
...
...
@@ -2610,8 +2612,9 @@ struct SimpleOpTypeSetTeller : public Teller {
"affine_channel"
,
"nearest_interp"
,
"anchor_generator"
,
"reduce_
sum
"
,
"reduce_
max
"
,
"reduce_mean"
,
"reduce_sum"
,
"conv3d"
,
"conv3d_transpose"
,
"mish"
,
...
...
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_reduce
_mean
.py
→
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_reduce.py
浏览文件 @
dd304f31
...
...
@@ -23,7 +23,7 @@ from trt_layer_auto_scan_test import TrtLayerAutoScanTest
import
paddle.inference
as
paddle_infer
class
TrtConvertReduce
Mean
Test
(
TrtLayerAutoScanTest
):
class
TrtConvertReduceTest
(
TrtLayerAutoScanTest
):
def
is_program_valid
(
self
,
program_config
:
ProgramConfig
)
->
bool
:
inputs
=
program_config
.
inputs
attrs
=
[
...
...
@@ -66,44 +66,51 @@ class TrtConvertReduceMeanTest(TrtLayerAutoScanTest):
]:
for
reduce_all
in
[
True
,
False
]:
for
out_dtype
in
[
-
1
,
2
,
5
]:
dics
=
[
{
"keep_dim"
:
keep_dim
,
"dim"
:
dim
,
"reduce_all"
:
reduce_all
,
"out_dtype"
:
out_dtype
,
"in_dtype"
:
out_dtype
,
},
{},
]
ops_config
=
[
{
"op_type"
:
"reduce_mean"
,
"op_inputs"
:
{
"X"
:
[
"input_data"
]},
"op_outputs"
:
{
"Out"
:
[
"reduce_output_data"
]},
"op_attrs"
:
dics
[
0
],
}
]
ops
=
self
.
generate_op_config
(
ops_config
)
program_config
=
ProgramConfig
(
ops
=
ops
,
weights
=
{},
inputs
=
{
"input_data"
:
TensorConfig
(
data_gen
=
partial
(
generate_input1
,
out_dtype
,
dics
for
op_type
in
[
"reduce_max"
,
"reduce_mean"
,
"reduce_sum"
,
]:
dics
=
[
{
"keep_dim"
:
keep_dim
,
"dim"
:
dim
,
"reduce_all"
:
reduce_all
,
"out_dtype"
:
out_dtype
,
"in_dtype"
:
out_dtype
,
},
{},
]
ops_config
=
[
{
"op_type"
:
op_type
,
"op_inputs"
:
{
"X"
:
[
"input_data"
]},
"op_outputs"
:
{
"Out"
:
[
"reduce_output_data"
]
},
"op_attrs"
:
dics
[
0
],
}
]
ops
=
self
.
generate_op_config
(
ops_config
)
program_config
=
ProgramConfig
(
ops
=
ops
,
weights
=
{},
inputs
=
{
"input_data"
:
TensorConfig
(
data_gen
=
partial
(
generate_input1
,
out_dtype
,
dics
)
)
)
},
outputs
=
[
"reduce_output_data"
],
)
},
outputs
=
[
"reduce_output_data"
],
)
if
not
self
.
is_program_valid
(
program_config
):
continue
if
not
self
.
is_program_valid
(
program_config
):
continue
yield
program_config
yield
program_config
def
sample_predictor_configs
(
self
,
program_config
...
...
@@ -139,22 +146,22 @@ class TrtConvertReduceMeanTest(TrtLayerAutoScanTest):
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Float32
yield
self
.
create_inference_config
(),
generate_trt_nodes_num
(
attrs
,
False
),
1e-5
),
(
1e-5
,
1e-5
)
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Half
yield
self
.
create_inference_config
(),
generate_trt_nodes_num
(
attrs
,
False
),
(
5e-4
,
5e-4
)
),
(
1e-3
,
1e-3
)
# for dynamic_shape
generate_dynamic_shape
(
attrs
)
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Float32
yield
self
.
create_inference_config
(),
generate_trt_nodes_num
(
attrs
,
True
),
1e-5
),
(
1e-5
,
1e-5
)
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Half
yield
self
.
create_inference_config
(),
generate_trt_nodes_num
(
attrs
,
True
),
(
5e-4
,
5e-4
)
),
(
1e-3
,
1e-3
)
def
add_skip_trt_case
(
self
):
pass
...
...
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_reduce_sum.py
已删除
100644 → 0
浏览文件 @
0c7f3575
# Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# 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.
import
unittest
from
functools
import
partial
from
typing
import
Any
,
Dict
,
List
import
numpy
as
np
from
program_config
import
ProgramConfig
,
TensorConfig
from
trt_layer_auto_scan_test
import
TrtLayerAutoScanTest
import
paddle.inference
as
paddle_infer
class
TrtConvertReduceSumTest
(
TrtLayerAutoScanTest
):
def
is_program_valid
(
self
,
program_config
:
ProgramConfig
)
->
bool
:
inputs
=
program_config
.
inputs
attrs
=
[
program_config
.
ops
[
i
].
attrs
for
i
in
range
(
len
(
program_config
.
ops
))
]
# dim should be in (-rank, rank), and not NONE
rank
=
len
(
inputs
[
'input_data'
].
shape
)
for
x
in
attrs
[
0
][
"dim"
]:
if
x
>=
rank
or
x
<=
-
rank
:
return
False
if
len
(
attrs
[
0
][
"dim"
])
==
0
:
return
False
return
True
def
sample_program_configs
(
self
):
def
generate_input1
(
dtype
,
attrs
:
List
[
Dict
[
str
,
Any
]]):
if
dtype
==
-
1
or
dtype
==
5
:
return
np
.
random
.
random
([
1
,
3
,
32
,
32
]).
astype
(
np
.
float32
)
elif
dtype
==
2
:
return
np
.
random
.
random
([
1
,
3
,
32
,
32
]).
astype
(
np
.
int32
)
for
keep_dim
in
[
True
,
False
]:
for
dim
in
[
[],
[
1
],
[
0
],
[
0
,
1
],
[
1
,
2
,
3
],
[
-
2
,
0
,
3
],
[
-
3
],
[
-
4
,
1
],
[
3
,
4
,
5
],
]:
for
reduce_all
in
[
True
,
False
]:
for
out_dtype
in
[
-
1
,
2
,
5
]:
dics
=
[
{
"keep_dim"
:
keep_dim
,
"dim"
:
dim
,
"reduce_all"
:
reduce_all
,
"out_dtype"
:
out_dtype
,
"in_dtype"
:
out_dtype
,
},
{},
]
ops_config
=
[
{
"op_type"
:
"reduce_sum"
,
"op_inputs"
:
{
"X"
:
[
"input_data"
]},
"op_outputs"
:
{
"Out"
:
[
"reduce_output_data"
]},
"op_attrs"
:
dics
[
0
],
}
]
ops
=
self
.
generate_op_config
(
ops_config
)
program_config
=
ProgramConfig
(
ops
=
ops
,
weights
=
{},
inputs
=
{
"input_data"
:
TensorConfig
(
data_gen
=
partial
(
generate_input1
,
out_dtype
,
dics
)
)
},
outputs
=
[
"reduce_output_data"
],
)
if
not
self
.
is_program_valid
(
program_config
):
continue
yield
program_config
def
sample_predictor_configs
(
self
,
program_config
):
def
generate_dynamic_shape
(
attrs
):
self
.
dynamic_shape
.
min_input_shape
=
{
"input_data"
:
[
1
,
3
,
32
,
32
]}
self
.
dynamic_shape
.
max_input_shape
=
{
"input_data"
:
[
4
,
3
,
64
,
64
]}
self
.
dynamic_shape
.
opt_input_shape
=
{
"input_data"
:
[
1
,
3
,
32
,
32
]}
def
clear_dynamic_shape
():
self
.
dynamic_shape
.
min_input_shape
=
{}
self
.
dynamic_shape
.
max_input_shape
=
{}
self
.
dynamic_shape
.
opt_input_shape
=
{}
def
generate_trt_nodes_num
(
attrs
,
dynamic_shape
):
if
dynamic_shape
:
if
(
not
attrs
[
0
][
'keep_dim'
])
and
attrs
[
0
][
'reduce_all'
]:
return
0
,
3
else
:
return
1
,
2
else
:
if
0
in
attrs
[
0
][
'dim'
]
or
attrs
[
0
][
'reduce_all'
]:
return
0
,
3
else
:
return
1
,
2
attrs
=
[
program_config
.
ops
[
i
].
attrs
for
i
in
range
(
len
(
program_config
.
ops
))
]
# for static_shape
clear_dynamic_shape
()
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Float32
yield
self
.
create_inference_config
(),
generate_trt_nodes_num
(
attrs
,
False
),
(
1e-5
,
1e-5
)
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Half
yield
self
.
create_inference_config
(),
generate_trt_nodes_num
(
attrs
,
False
),
(
1e-3
,
1e-3
)
# for dynamic_shape
generate_dynamic_shape
(
attrs
)
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Float32
yield
self
.
create_inference_config
(),
generate_trt_nodes_num
(
attrs
,
True
),
(
1e-5
,
1e-5
)
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Half
yield
self
.
create_inference_config
(),
generate_trt_nodes_num
(
attrs
,
True
),
(
1e-3
,
1e-3
)
def
add_skip_trt_case
(
self
):
pass
def
test
(
self
):
self
.
add_skip_trt_case
()
self
.
run_test
()
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