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
兽拳
Paddle
提交
5c291737
P
Paddle
项目概览
兽拳
/
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
5c291737
编写于
7月 18, 2022
作者:
Z
zhoutianzi666
提交者:
GitHub
7月 18, 2022
浏览文件
操作
浏览文件
下载
电子邮件补丁
差异文件
[Paddle-TRT] remove useless code in fc (#44382)
* remove useless code in fc
上级
0fd974b4
变更
2
显示空白变更内容
内联
并排
Showing
2 changed file
with
377 addition
and
84 deletion
+377
-84
paddle/fluid/inference/tensorrt/convert/fc_op.cc
paddle/fluid/inference/tensorrt/convert/fc_op.cc
+16
-84
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_fc.py
...fluid/tests/unittests/ir/inference/test_trt_convert_fc.py
+361
-0
未找到文件。
paddle/fluid/inference/tensorrt/convert/fc_op.cc
浏览文件 @
5c291737
...
@@ -333,74 +333,6 @@ class FcOpConverter : public OpConverter {
...
@@ -333,74 +333,6 @@ class FcOpConverter : public OpConverter {
if
(
!
engine_
->
with_dynamic_shape
())
{
if
(
!
engine_
->
with_dynamic_shape
())
{
x_num_col_dims
--
;
x_num_col_dims
--
;
}
}
// If use tensorrt'oss, the x_dim and x_num_col_dims need change, and can
// not add Shuffle layer in ernie's multihead.
if
(
x_dim
.
nbDims
==
4
&&
x_num_col_dims
==
1
)
{
if
(
enable_int8
||
support_int8
)
{
// add conv1x1 layer
nvinfer1
::
DimsHW
nv_ksize
(
1
,
1
);
auto
*
fc_layer_int8
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Convolution
,
*
X
,
n_output
,
nv_ksize
,
weight
.
get
(),
bias
.
get
());
if
(
activation_type
==
"relu"
)
{
fc_layer_int8
->
setName
(
(
"ernie_fc_op_int8: Convolution (Output: "
+
output_name
+
")"
)
.
c_str
());
PADDLE_ENFORCE_EQ
(
op_desc
.
HasAttr
(
"out_threshold"
),
true
,
platform
::
errors
::
InvalidArgument
(
"must have out threshold in fc layers in int8 mode"
));
float
out_scale
=
0
;
if
(
enable_int8
)
{
out_scale
=
BOOST_GET_CONST
(
float
,
op_desc
.
GetAttr
(
"out_threshold"
));
}
else
{
out_scale
=
BOOST_GET_CONST
(
float
,
op_desc
.
GetAttr
(
"Out"
));
}
engine_
->
SetTensorDynamicRange
(
fc_layer_int8
->
getOutput
(
0
),
out_scale
);
nvinfer1
::
IActivationLayer
*
relu_layer_int8
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Activation
,
*
(
fc_layer_int8
->
getOutput
(
0
)),
nvinfer1
::
ActivationType
::
kRELU
);
RreplenishLayerAndOutput
(
relu_layer_int8
,
"relu_after_ernie_fc_int8"
,
{
output_name
},
test_mode
);
}
else
{
RreplenishLayerAndOutput
(
fc_layer_int8
,
"ernie_fc_op_int8: Convolution"
,
{
output_name
},
test_mode
);
}
}
else
{
// add fc layer
auto
*
fc_layer_float
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
FullyConnected
,
*
X
,
n_output
,
weight
.
get
(),
bias
.
get
());
if
(
activation_type
==
"relu"
)
{
fc_layer_float
->
setName
(
(
"ernie_fc_op_float: (Output: "
+
output_name
+
")"
).
c_str
());
nvinfer1
::
IActivationLayer
*
relu_layer_float
=
TRT_ENGINE_ADD_LAYER
(
engine_
,
Activation
,
*
(
fc_layer_float
->
getOutput
(
0
)),
nvinfer1
::
ActivationType
::
kRELU
);
RreplenishLayerAndOutput
(
relu_layer_float
,
"relu_after_ernie_fc_float"
,
{
output_name
},
test_mode
);
}
else
{
RreplenishLayerAndOutput
(
fc_layer_float
,
"ernie_fc_op_float"
,
{
output_name
},
test_mode
);
}
}
}
else
{
// need reshape input before and after fc
PADDLE_ENFORCE_GT
(
PADDLE_ENFORCE_GT
(
x_dim
.
nbDims
,
x_dim
.
nbDims
,
x_num_col_dims
,
x_num_col_dims
,
...
@@ -410,6 +342,7 @@ class FcOpConverter : public OpConverter {
...
@@ -410,6 +342,7 @@ class FcOpConverter : public OpConverter {
"x_dim.nbDims : %d, x_num_col_dims : %d."
,
"x_dim.nbDims : %d, x_num_col_dims : %d."
,
x_dim
.
nbDims
,
x_dim
.
nbDims
,
x_num_col_dims
));
x_num_col_dims
));
// need reshape input before and after fc
auto
*
reshape_before_fc_layer
=
auto
*
reshape_before_fc_layer
=
reshape_before_fc
(
X
,
x_dim
,
x_num_col_dims
,
output_name
);
reshape_before_fc
(
X
,
x_dim
,
x_num_col_dims
,
output_name
);
auto
*
reshape_itensor
=
reshape_before_fc_layer
->
getOutput
(
0
);
auto
*
reshape_itensor
=
reshape_before_fc_layer
->
getOutput
(
0
);
...
@@ -418,7 +351,6 @@ class FcOpConverter : public OpConverter {
...
@@ -418,7 +351,6 @@ class FcOpConverter : public OpConverter {
}
}
regist_fc
(
reshape_itensor
,
n_output
,
weight
,
bias
);
regist_fc
(
reshape_itensor
,
n_output
,
weight
,
bias
);
}
}
}
};
};
}
// namespace tensorrt
}
// namespace tensorrt
...
...
python/paddle/fluid/tests/unittests/ir/inference/test_trt_convert_fc.py
0 → 100644
浏览文件 @
5c291737
# 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.
from
trt_layer_auto_scan_test
import
TrtLayerAutoScanTest
,
SkipReasons
from
program_config
import
TensorConfig
,
ProgramConfig
import
numpy
as
np
import
unittest
import
paddle.inference
as
paddle_infer
from
functools
import
partial
from
typing
import
Optional
,
List
,
Callable
,
Dict
,
Any
,
Set
import
os
class
TrtConvertFcTest
(
TrtLayerAutoScanTest
):
def
is_program_valid
(
self
,
program_config
:
ProgramConfig
)
->
bool
:
# The output has diff between gpu and trt in CI windows
if
(
os
.
name
==
'nt'
):
return
False
return
True
def
sample_program_configs
(
self
):
self
.
trt_param
.
workspace_size
=
1073741824
def
generate_input1
(
batch
,
attrs
:
List
[
Dict
[
str
,
Any
]]):
return
np
.
random
.
random
([
batch
,
3
,
64
,
(
int
)(
attrs
[
0
][
"m"
]
/
2
),
2
]).
astype
(
np
.
float32
)
def
generate_w
(
batch
,
attrs
:
List
[
Dict
[
str
,
Any
]]):
return
np
.
random
.
random
([
attrs
[
0
][
"m"
],
attrs
[
0
][
"n"
]]).
astype
(
np
.
float32
)
def
generate_bias
(
batch
,
attrs
:
List
[
Dict
[
str
,
Any
]]):
return
np
.
random
.
random
([
attrs
[
0
][
"n"
]]).
astype
(
np
.
float32
)
for
batch
in
[
1
,
4
]:
for
[
m
,
n
]
in
[[
32
,
23
]]:
dics
=
[
{
"in_num_col_dims"
:
3
,
# for my conveinence
"m"
:
m
,
"n"
:
n
,
},
{}
]
ops_config
=
[
{
"op_type"
:
"fc"
,
"op_inputs"
:
{
"Input"
:
[
"input_data"
],
"W"
:
[
"w_data"
],
"Bias"
:
[
"bias_data"
]
},
"op_outputs"
:
{
"Out"
:
[
"output_data"
]
},
"op_attrs"
:
dics
[
0
]
},
]
ops
=
self
.
generate_op_config
(
ops_config
)
program_config
=
ProgramConfig
(
ops
=
ops
,
weights
=
{
"w_data"
:
TensorConfig
(
data_gen
=
partial
(
generate_w
,
batch
,
dics
)),
"bias_data"
:
TensorConfig
(
data_gen
=
partial
(
generate_bias
,
batch
,
dics
))
},
inputs
=
{
"input_data"
:
TensorConfig
(
data_gen
=
partial
(
generate_input1
,
batch
,
dics
)),
},
outputs
=
[
"output_data"
])
yield
program_config
def
sample_predictor_configs
(
self
,
program_config
)
->
(
paddle_infer
.
Config
,
List
[
int
],
float
):
def
generate_dynamic_shape
(
attrs
):
self
.
dynamic_shape
.
min_input_shape
=
{
"input_data"
:
[
1
,
3
,
32
,
16
,
2
],
}
self
.
dynamic_shape
.
max_input_shape
=
{
"input_data"
:
[
4
,
3
,
64
,
16
,
2
],
}
self
.
dynamic_shape
.
opt_input_shape
=
{
"input_data"
:
[
1
,
3
,
64
,
16
,
2
],
}
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
):
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
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Half
yield
self
.
create_inference_config
(),
generate_trt_nodes_num
(
attrs
,
False
),
(
1e-5
,
1e-5
)
# 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
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Half
yield
self
.
create_inference_config
(),
generate_trt_nodes_num
(
attrs
,
True
),
(
1e-5
,
1e-5
)
def
test
(
self
):
self
.
run_test
()
def
test_quant
(
self
):
self
.
run_test
(
quant
=
True
)
class
TrtConvertFcTest2
(
TrtLayerAutoScanTest
):
def
is_program_valid
(
self
,
program_config
:
ProgramConfig
)
->
bool
:
# The output has diff between gpu and trt in CI windows
if
(
os
.
name
==
'nt'
):
return
False
return
True
def
sample_program_configs
(
self
):
self
.
trt_param
.
workspace_size
=
1073741824
def
generate_input1
(
batch
,
attrs
:
List
[
Dict
[
str
,
Any
]]):
return
np
.
random
.
random
([
batch
,
3
,
64
,
14
]).
astype
(
np
.
float32
)
def
generate_w
(
batch
,
attrs
:
List
[
Dict
[
str
,
Any
]]):
return
np
.
random
.
random
([
attrs
[
0
][
"m"
],
attrs
[
0
][
"n"
]]).
astype
(
np
.
float32
)
def
generate_bias
(
batch
,
attrs
:
List
[
Dict
[
str
,
Any
]]):
return
np
.
random
.
random
([
attrs
[
0
][
"n"
]]).
astype
(
np
.
float32
)
for
batch
in
[
1
,
4
]:
for
[
m
,
n
]
in
[[
14
,
43
]]:
dics
=
[
{
"in_num_col_dims"
:
3
,
# for my conveinence
"m"
:
m
,
"n"
:
n
,
},
{}
]
ops_config
=
[
{
"op_type"
:
"fc"
,
"op_inputs"
:
{
"Input"
:
[
"input_data"
],
"W"
:
[
"w_data"
],
"Bias"
:
[
"bias_data"
]
},
"op_outputs"
:
{
"Out"
:
[
"output_data"
]
},
"op_attrs"
:
dics
[
0
]
},
]
ops
=
self
.
generate_op_config
(
ops_config
)
program_config
=
ProgramConfig
(
ops
=
ops
,
weights
=
{
"w_data"
:
TensorConfig
(
data_gen
=
partial
(
generate_w
,
batch
,
dics
)),
"bias_data"
:
TensorConfig
(
data_gen
=
partial
(
generate_bias
,
batch
,
dics
))
},
inputs
=
{
"input_data"
:
TensorConfig
(
data_gen
=
partial
(
generate_input1
,
batch
,
dics
)),
},
outputs
=
[
"output_data"
])
yield
program_config
def
sample_predictor_configs
(
self
,
program_config
)
->
(
paddle_infer
.
Config
,
List
[
int
],
float
):
def
generate_dynamic_shape
():
self
.
dynamic_shape
.
min_input_shape
=
{
"input_data"
:
[
1
,
3
,
32
,
14
],
}
self
.
dynamic_shape
.
max_input_shape
=
{
"input_data"
:
[
4
,
3
,
64
,
14
],
}
self
.
dynamic_shape
.
opt_input_shape
=
{
"input_data"
:
[
1
,
3
,
64
,
14
],
}
def
clear_dynamic_shape
():
self
.
dynamic_shape
.
min_input_shape
=
{}
self
.
dynamic_shape
.
max_input_shape
=
{}
self
.
dynamic_shape
.
opt_input_shape
=
{}
# # for static_shape
clear_dynamic_shape
()
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Float32
yield
self
.
create_inference_config
(),
(
1
,
2
),
1e-5
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Half
yield
self
.
create_inference_config
(),
(
1
,
2
),
(
1e-5
,
1e-5
)
# for dynamic_shape
generate_dynamic_shape
()
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Float32
yield
self
.
create_inference_config
(),
(
1
,
2
),
1e-5
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Half
yield
self
.
create_inference_config
(),
(
1
,
2
),
(
1e-5
,
1e-5
)
def
test
(
self
):
self
.
run_test
()
# this is the special case when x_dim.nbDims == 4 && x_num_col_dims == 1
class
TrtConvertFcTest3
(
TrtLayerAutoScanTest
):
def
is_program_valid
(
self
,
program_config
:
ProgramConfig
)
->
bool
:
return
True
def
sample_program_configs
(
self
):
self
.
trt_param
.
workspace_size
=
1073741824
def
generate_input1
(
batch
,
attrs
:
List
[
Dict
[
str
,
Any
]]):
return
np
.
ones
([
batch
,
14
,
1
,
2
]).
astype
(
np
.
float32
)
def
generate_w
(
batch
,
attrs
:
List
[
Dict
[
str
,
Any
]]):
return
np
.
ones
([
attrs
[
0
][
"m"
],
attrs
[
0
][
"n"
]]).
astype
(
np
.
float32
)
def
generate_bias
(
batch
,
attrs
:
List
[
Dict
[
str
,
Any
]]):
return
np
.
ones
([
attrs
[
0
][
"n"
]]).
astype
(
np
.
float32
)
for
batch
in
[
1
,
4
]:
for
[
m
,
n
]
in
[[
28
,
43
]]:
dics
=
[
{
"in_num_col_dims"
:
1
,
"Input_scale"
:
0.1
,
"out_threshold"
:
0.1
,
"enable_int8"
:
True
,
# for my conveinence
"m"
:
m
,
"n"
:
n
,
},
{}
]
ops_config
=
[
{
"op_type"
:
"fc"
,
"op_inputs"
:
{
"Input"
:
[
"input_data"
],
"W"
:
[
"w_data"
],
"Bias"
:
[
"bias_data"
]
},
"op_outputs"
:
{
"Out"
:
[
"output_data"
]
},
"op_attrs"
:
dics
[
0
]
},
]
ops
=
self
.
generate_op_config
(
ops_config
)
program_config
=
ProgramConfig
(
ops
=
ops
,
weights
=
{
"w_data"
:
TensorConfig
(
data_gen
=
partial
(
generate_w
,
batch
,
dics
)),
"bias_data"
:
TensorConfig
(
data_gen
=
partial
(
generate_bias
,
batch
,
dics
))
},
inputs
=
{
"input_data"
:
TensorConfig
(
data_gen
=
partial
(
generate_input1
,
batch
,
dics
)),
},
outputs
=
[
"output_data"
])
yield
program_config
def
sample_predictor_configs
(
self
,
program_config
)
->
(
paddle_infer
.
Config
,
List
[
int
],
float
):
def
generate_dynamic_shape
():
self
.
dynamic_shape
.
min_input_shape
=
{
"input_data"
:
[
1
,
14
,
1
,
2
],
}
self
.
dynamic_shape
.
max_input_shape
=
{
"input_data"
:
[
4
,
14
,
1
,
2
],
}
self
.
dynamic_shape
.
opt_input_shape
=
{
"input_data"
:
[
1
,
14
,
1
,
2
],
}
def
clear_dynamic_shape
():
self
.
dynamic_shape
.
min_input_shape
=
{}
self
.
dynamic_shape
.
max_input_shape
=
{}
self
.
dynamic_shape
.
opt_input_shape
=
{}
# for static_shape
clear_dynamic_shape
()
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Float32
yield
self
.
create_inference_config
(),
(
1
,
2
),
1e-5
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Half
yield
self
.
create_inference_config
(),
(
1
,
2
),
(
1e-5
,
1e-5
)
# for dynamic_shape
generate_dynamic_shape
()
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Float32
yield
self
.
create_inference_config
(),
(
1
,
2
),
1e-5
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Half
yield
self
.
create_inference_config
(),
(
1
,
2
),
(
1e-5
,
1e-5
)
self
.
trt_param
.
precision
=
paddle_infer
.
PrecisionType
.
Int8
yield
self
.
create_inference_config
(),
(
1
,
2
),
(
1e-5
,
1e-5
)
def
test
(
self
):
self
.
run_test
()
def
test_quant
(
self
):
self
.
run_test
(
quant
=
True
)
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