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
马赛克医院
Paddle
提交
94b66bdb
P
Paddle
项目概览
马赛克医院
/
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
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
94b66bdb
编写于
9月 11, 2018
作者:
T
Tao Luo
提交者:
GitHub
9月 11, 2018
浏览文件
操作
浏览文件
下载
差异文件
Merge pull request #13326 from luotao1/analysis_test_refine
refine Analysis test
上级
76e92274
9664c53c
变更
8
隐藏空白更改
内联
并排
Showing
8 changed file
with
368 addition
and
345 deletion
+368
-345
paddle/fluid/inference/CMakeLists.txt
paddle/fluid/inference/CMakeLists.txt
+2
-1
paddle/fluid/inference/analysis/CMakeLists.txt
paddle/fluid/inference/analysis/CMakeLists.txt
+1
-64
paddle/fluid/inference/analysis/analyzer_tester.cc
paddle/fluid/inference/analysis/analyzer_tester.cc
+2
-280
paddle/fluid/inference/tests/api/CMakeLists.txt
paddle/fluid/inference/tests/api/CMakeLists.txt
+57
-0
paddle/fluid/inference/tests/api/analyzer_lac_tester.cc
paddle/fluid/inference/tests/api/analyzer_lac_tester.cc
+0
-0
paddle/fluid/inference/tests/api/analyzer_ner_tester.cc
paddle/fluid/inference/tests/api/analyzer_ner_tester.cc
+0
-0
paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc
paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc
+306
-0
paddle/fluid/inference/tests/api/analyzer_text_classification_tester.cc
...nference/tests/api/analyzer_text_classification_tester.cc
+0
-0
未找到文件。
paddle/fluid/inference/CMakeLists.txt
浏览文件 @
94b66bdb
...
...
@@ -55,6 +55,7 @@ if(NOT APPLE)
endif
()
if
(
WITH_TESTING
)
#
both tests/book and analysis
depends the models that generated by python/paddle/fluid/tests/book
#
tests/book
depends the models that generated by python/paddle/fluid/tests/book
add_subdirectory
(
tests/book
)
add_subdirectory
(
tests/api
)
endif
()
paddle/fluid/inference/analysis/CMakeLists.txt
浏览文件 @
94b66bdb
...
...
@@ -40,27 +40,7 @@ function (inference_analysis_test TARGET)
endif
(
WITH_TESTING
)
endfunction
(
inference_analysis_test
)
function
(
inference_download_and_uncompress install_dir url gz_filename
)
message
(
STATUS
"Download inference test stuff
${
gz_filename
}
from
${
url
}
"
)
execute_process
(
COMMAND bash -c
"mkdir -p
${
install_dir
}
"
)
execute_process
(
COMMAND bash -c
"cd
${
install_dir
}
&& wget -q
${
url
}
"
)
execute_process
(
COMMAND bash -c
"cd
${
install_dir
}
&& tar xzf
${
gz_filename
}
"
)
message
(
STATUS
"finish downloading
${
gz_filename
}
"
)
endfunction
(
inference_download_and_uncompress
)
set
(
RNN1_MODEL_URL
"http://paddle-inference-dist.bj.bcebos.com/rnn1%2Fmodel.tar.gz"
)
set
(
RNN1_DATA_URL
"http://paddle-inference-dist.bj.bcebos.com/rnn1%2Fdata.txt.tar.gz"
)
set
(
RNN1_INSTALL_DIR
"
${
THIRD_PARTY_PATH
}
/inference_demo/rnn1"
CACHE PATH
"RNN1 model and data root."
FORCE
)
if
(
NOT EXISTS
${
RNN1_INSTALL_DIR
}
AND WITH_TESTING
)
inference_download_and_uncompress
(
${
RNN1_INSTALL_DIR
}
${
RNN1_MODEL_URL
}
"rnn1%2Fmodel.tar.gz"
)
inference_download_and_uncompress
(
${
RNN1_INSTALL_DIR
}
${
RNN1_DATA_URL
}
"rnn1%2Fdata.txt.tar.gz"
)
endif
()
inference_analysis_test
(
test_analyzer SRCS analyzer_tester.cc
EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor
ARGS --infer_model=
${
RNN1_INSTALL_DIR
}
/model
--infer_data=
${
RNN1_INSTALL_DIR
}
/data.txt
)
inference_analysis_test
(
test_analyzer SRCS analyzer_tester.cc EXTRA_DEPS paddle_inference_api
)
inference_analysis_test
(
test_data_flow_graph SRCS data_flow_graph_tester.cc
)
inference_analysis_test
(
test_data_flow_graph_to_fluid_pass SRCS data_flow_graph_to_fluid_pass_tester.cc
)
inference_analysis_test
(
test_fluid_to_ir_pass SRCS fluid_to_ir_pass_tester.cc
)
...
...
@@ -71,46 +51,3 @@ inference_analysis_test(test_tensorrt_subgraph_pass SRCS tensorrt_subgraph_pass_
inference_analysis_test
(
test_pass_manager SRCS pass_manager_tester.cc
)
inference_analysis_test
(
test_tensorrt_subgraph_node_mark_pass SRCS tensorrt_subgraph_node_mark_pass_tester.cc
)
inference_analysis_test
(
test_model_store_pass SRCS model_store_pass_tester.cc
)
set
(
CHINESE_NER_MODEL_URL
"http://paddle-inference-dist.bj.bcebos.com/chinese_ner_model.tar.gz"
)
set
(
CHINESE_NER_DATA_URL
"http://paddle-inference-dist.bj.bcebos.com/chinese_ner-data.txt.tar.gz"
)
set
(
CHINESE_NER_INSTALL_DIR
"
${
THIRD_PARTY_PATH
}
/inference_demo/chinese_ner"
CACHE PATH
"Chinese ner model and data root."
FORCE
)
if
(
NOT EXISTS
${
CHINESE_NER_INSTALL_DIR
}
AND WITH_TESTING AND WITH_INFERENCE
)
inference_download_and_uncompress
(
${
CHINESE_NER_INSTALL_DIR
}
${
CHINESE_NER_MODEL_URL
}
"chinese_ner_model.tar.gz"
)
inference_download_and_uncompress
(
${
CHINESE_NER_INSTALL_DIR
}
${
CHINESE_NER_DATA_URL
}
"chinese_ner-data.txt.tar.gz"
)
endif
()
inference_analysis_test
(
test_analyzer_ner SRCS analyzer_ner_tester.cc
EXTRA_DEPS paddle_inference_api paddle_fluid_api analysis_predictor
ARGS --infer_model=
${
CHINESE_NER_INSTALL_DIR
}
/model
--infer_data=
${
CHINESE_NER_INSTALL_DIR
}
/data.txt
)
set
(
LAC_MODEL_URL
"http://paddle-inference-dist.bj.bcebos.com/lac_model.tar.gz"
)
set
(
LAC_DATA_URL
"http://paddle-inference-dist.bj.bcebos.com/lac_data.txt.tar.gz"
)
set
(
LAC_INSTALL_DIR
"
${
THIRD_PARTY_PATH
}
/inference_demo/lac"
CACHE PATH
"LAC model and data root."
FORCE
)
if
(
NOT EXISTS
${
LAC_INSTALL_DIR
}
AND WITH_TESTING AND WITH_INFERENCE
)
inference_download_and_uncompress
(
${
LAC_INSTALL_DIR
}
${
LAC_MODEL_URL
}
"lac_model.tar.gz"
)
inference_download_and_uncompress
(
${
LAC_INSTALL_DIR
}
${
LAC_DATA_URL
}
"lac_data.txt.tar.gz"
)
endif
()
inference_analysis_test
(
test_analyzer_lac SRCS analyzer_lac_tester.cc
EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor
ARGS --infer_model=
${
LAC_INSTALL_DIR
}
/model
--infer_data=
${
LAC_INSTALL_DIR
}
/data.txt
)
set
(
TEXT_CLASSIFICATION_MODEL_URL
"http://paddle-inference-dist.bj.bcebos.com/text-classification-Senta.tar.gz"
)
set
(
TEXT_CLASSIFICATION_DATA_URL
"http://paddle-inference-dist.bj.bcebos.com/text_classification_data.txt.tar.gz"
)
set
(
TEXT_CLASSIFICATION_INSTALL_DIR
"
${
THIRD_PARTY_PATH
}
/inference_demo/text_classification"
CACHE PATH
"Text Classification model and data root."
FORCE
)
if
(
NOT EXISTS
${
TEXT_CLASSIFICATION_INSTALL_DIR
}
AND WITH_TESTING AND WITH_INFERENCE
)
inference_download_and_uncompress
(
${
TEXT_CLASSIFICATION_INSTALL_DIR
}
${
TEXT_CLASSIFICATION_MODEL_URL
}
"text-classification-Senta.tar.gz"
)
inference_download_and_uncompress
(
${
TEXT_CLASSIFICATION_INSTALL_DIR
}
${
TEXT_CLASSIFICATION_DATA_URL
}
"text_classification_data.txt.tar.gz"
)
endif
()
inference_analysis_test
(
test_text_classification SRCS analyzer_text_classification_tester.cc
EXTRA_DEPS paddle_inference_api paddle_fluid_api analysis_predictor
ARGS --infer_model=
${
TEXT_CLASSIFICATION_INSTALL_DIR
}
/text-classification-Senta
--infer_data=
${
TEXT_CLASSIFICATION_INSTALL_DIR
}
/data.txt
--topn=1
# Just run top 1 batch.
)
paddle/fluid/inference/analysis/analyzer_tester.cc
浏览文件 @
94b66bdb
...
...
@@ -16,21 +16,9 @@
#include <google/protobuf/text_format.h>
#include <gtest/gtest.h>
#include <thread> // NOLINT
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/pass.h"
#include "paddle/fluid/inference/analysis/ut_helper.h"
#include "paddle/fluid/inference/api/analysis_predictor.h"
#include "paddle/fluid/inference/api/helper.h"
#include "paddle/fluid/inference/api/paddle_inference_api.h"
#include "paddle/fluid/inference/api/paddle_inference_pass.h"
#include "paddle/fluid/inference/utils/singleton.h"
DEFINE_string
(
infer_model
,
""
,
"model path"
);
DEFINE_string
(
infer_data
,
""
,
"data path"
);
DEFINE_int32
(
batch_size
,
10
,
"batch size."
);
DEFINE_int32
(
repeat
,
1
,
"Running the inference program repeat times."
);
DEFINE_int32
(
num_threads
,
1
,
"Running the inference program in multi-threads."
);
namespace
paddle
{
namespace
inference
{
...
...
@@ -91,274 +79,8 @@ void TestWord2vecPrediction(const std::string &model_path) {
}
}
namespace
{
struct
DataRecord
{
std
::
vector
<
std
::
vector
<
std
::
vector
<
float
>>>
link_step_data_all
;
std
::
vector
<
std
::
vector
<
float
>>
week_data_all
,
minute_data_all
;
std
::
vector
<
size_t
>
lod1
,
lod2
,
lod3
;
std
::
vector
<
std
::
vector
<
float
>>
rnn_link_data
,
rnn_week_datas
,
rnn_minute_datas
;
size_t
batch_iter
{
0
};
size_t
batch_size
{
1
};
DataRecord
()
=
default
;
explicit
DataRecord
(
const
std
::
string
&
path
,
int
batch_size
=
1
)
:
batch_size
(
batch_size
)
{
Load
(
path
);
}
DataRecord
NextBatch
()
{
DataRecord
data
;
size_t
batch_end
=
batch_iter
+
batch_size
;
// NOTE skip the final batch, if no enough data is provided.
if
(
batch_end
<=
link_step_data_all
.
size
())
{
data
.
link_step_data_all
.
assign
(
link_step_data_all
.
begin
()
+
batch_iter
,
link_step_data_all
.
begin
()
+
batch_end
);
data
.
week_data_all
.
assign
(
week_data_all
.
begin
()
+
batch_iter
,
week_data_all
.
begin
()
+
batch_end
);
data
.
minute_data_all
.
assign
(
minute_data_all
.
begin
()
+
batch_iter
,
minute_data_all
.
begin
()
+
batch_end
);
// Prepare LoDs
data
.
lod1
.
push_back
(
0
);
data
.
lod2
.
push_back
(
0
);
data
.
lod3
.
push_back
(
0
);
CHECK
(
!
data
.
link_step_data_all
.
empty
())
<<
"empty"
;
CHECK
(
!
data
.
week_data_all
.
empty
());
CHECK
(
!
data
.
minute_data_all
.
empty
());
CHECK_EQ
(
data
.
link_step_data_all
.
size
(),
data
.
week_data_all
.
size
());
CHECK_EQ
(
data
.
minute_data_all
.
size
(),
data
.
link_step_data_all
.
size
());
for
(
size_t
j
=
0
;
j
<
data
.
link_step_data_all
.
size
();
j
++
)
{
for
(
const
auto
&
d
:
data
.
link_step_data_all
[
j
])
{
data
.
rnn_link_data
.
push_back
(
d
);
}
data
.
rnn_week_datas
.
push_back
(
data
.
week_data_all
[
j
]);
data
.
rnn_minute_datas
.
push_back
(
data
.
minute_data_all
[
j
]);
// calculate lod
data
.
lod1
.
push_back
(
data
.
lod1
.
back
()
+
data
.
link_step_data_all
[
j
].
size
());
data
.
lod3
.
push_back
(
data
.
lod3
.
back
()
+
1
);
for
(
size_t
i
=
1
;
i
<
data
.
link_step_data_all
[
j
].
size
()
+
1
;
i
++
)
{
data
.
lod2
.
push_back
(
data
.
lod2
.
back
()
+
data
.
link_step_data_all
[
j
].
size
());
}
}
}
batch_iter
+=
batch_size
;
return
data
;
}
void
Load
(
const
std
::
string
&
path
)
{
std
::
ifstream
file
(
path
);
std
::
string
line
;
int
num_lines
=
0
;
while
(
std
::
getline
(
file
,
line
))
{
num_lines
++
;
std
::
vector
<
std
::
string
>
data
;
split
(
line
,
':'
,
&
data
);
std
::
vector
<
std
::
vector
<
float
>>
link_step_data
;
std
::
vector
<
std
::
string
>
link_datas
;
split
(
data
[
0
],
'|'
,
&
link_datas
);
for
(
auto
&
step_data
:
link_datas
)
{
std
::
vector
<
float
>
tmp
;
split_to_float
(
step_data
,
','
,
&
tmp
);
link_step_data
.
push_back
(
tmp
);
}
// load week data
std
::
vector
<
float
>
week_data
;
split_to_float
(
data
[
2
],
','
,
&
week_data
);
// load minute data
std
::
vector
<
float
>
minute_data
;
split_to_float
(
data
[
1
],
','
,
&
minute_data
);
link_step_data_all
.
push_back
(
std
::
move
(
link_step_data
));
week_data_all
.
push_back
(
std
::
move
(
week_data
));
minute_data_all
.
push_back
(
std
::
move
(
minute_data
));
}
}
};
void
PrepareInputs
(
std
::
vector
<
PaddleTensor
>
*
input_slots
,
DataRecord
*
data
,
int
batch_size
)
{
PaddleTensor
lod_attention_tensor
,
init_zero_tensor
,
lod_tensor_tensor
,
week_tensor
,
minute_tensor
;
lod_attention_tensor
.
name
=
"data_lod_attention"
;
init_zero_tensor
.
name
=
"cell_init"
;
lod_tensor_tensor
.
name
=
"data"
;
week_tensor
.
name
=
"week"
;
minute_tensor
.
name
=
"minute"
;
auto
one_batch
=
data
->
NextBatch
();
std
::
vector
<
int
>
rnn_link_data_shape
(
{
static_cast
<
int
>
(
one_batch
.
rnn_link_data
.
size
()),
static_cast
<
int
>
(
one_batch
.
rnn_link_data
.
front
().
size
())});
lod_attention_tensor
.
shape
.
assign
({
1
,
2
});
lod_attention_tensor
.
lod
.
assign
({
one_batch
.
lod1
,
one_batch
.
lod2
});
init_zero_tensor
.
shape
.
assign
({
batch_size
,
15
});
init_zero_tensor
.
lod
.
assign
({
one_batch
.
lod3
});
lod_tensor_tensor
.
shape
=
rnn_link_data_shape
;
lod_tensor_tensor
.
lod
.
assign
({
one_batch
.
lod1
});
// clang-format off
week_tensor
.
shape
.
assign
(
{
static_cast
<
int
>
(
one_batch
.
rnn_week_datas
.
size
()),
static_cast
<
int
>
(
one_batch
.
rnn_week_datas
.
front
().
size
())});
week_tensor
.
lod
.
assign
({
one_batch
.
lod3
});
minute_tensor
.
shape
.
assign
(
{
static_cast
<
int
>
(
one_batch
.
rnn_minute_datas
.
size
()),
static_cast
<
int
>
(
one_batch
.
rnn_minute_datas
.
front
().
size
())});
minute_tensor
.
lod
.
assign
({
one_batch
.
lod3
});
// clang-format on
// assign data
TensorAssignData
<
float
>
(
&
lod_attention_tensor
,
std
::
vector
<
std
::
vector
<
float
>>
({{
0
,
0
}}));
std
::
vector
<
float
>
tmp_zeros
(
batch_size
*
15
,
0.
);
TensorAssignData
<
float
>
(
&
init_zero_tensor
,
{
tmp_zeros
});
TensorAssignData
<
float
>
(
&
lod_tensor_tensor
,
one_batch
.
rnn_link_data
);
TensorAssignData
<
float
>
(
&
week_tensor
,
one_batch
.
rnn_week_datas
);
TensorAssignData
<
float
>
(
&
minute_tensor
,
one_batch
.
rnn_minute_datas
);
// Set inputs.
auto
init_zero_tensor1
=
init_zero_tensor
;
init_zero_tensor1
.
name
=
"hidden_init"
;
input_slots
->
assign
({
week_tensor
,
init_zero_tensor
,
minute_tensor
,
init_zero_tensor1
,
lod_attention_tensor
,
lod_tensor_tensor
});
for
(
auto
&
tensor
:
*
input_slots
)
{
tensor
.
dtype
=
PaddleDType
::
FLOAT32
;
}
}
}
// namespace
void
CompareResult
(
const
std
::
vector
<
PaddleTensor
>
&
outputs
,
const
std
::
vector
<
PaddleTensor
>
&
base_outputs
)
{
PADDLE_ENFORCE_GT
(
outputs
.
size
(),
0
);
PADDLE_ENFORCE_EQ
(
outputs
.
size
(),
base_outputs
.
size
());
for
(
size_t
i
=
0
;
i
<
outputs
.
size
();
i
++
)
{
auto
&
out
=
outputs
[
i
];
auto
&
base_out
=
base_outputs
[
i
];
size_t
size
=
std
::
accumulate
(
out
.
shape
.
begin
(),
out
.
shape
.
end
(),
1
,
[](
int
a
,
int
b
)
{
return
a
*
b
;
});
size_t
size1
=
std
::
accumulate
(
base_out
.
shape
.
begin
(),
base_out
.
shape
.
end
(),
1
,
[](
int
a
,
int
b
)
{
return
a
*
b
;
});
PADDLE_ENFORCE_EQ
(
size
,
size1
);
PADDLE_ENFORCE_GT
(
size
,
0
);
float
*
data
=
static_cast
<
float
*>
(
out
.
data
.
data
());
float
*
base_data
=
static_cast
<
float
*>
(
base_out
.
data
.
data
());
for
(
size_t
i
=
0
;
i
<
size
;
i
++
)
{
EXPECT_NEAR
(
data
[
i
],
base_data
[
i
],
1e-3
);
}
}
}
// Test with a really complicate model.
void
TestRNN1Prediction
(
bool
use_analysis
,
bool
activate_ir
,
int
num_threads
)
{
AnalysisConfig
config
;
config
.
prog_file
=
FLAGS_infer_model
+
"/__model__"
;
config
.
param_file
=
FLAGS_infer_model
+
"/param"
;
config
.
use_gpu
=
false
;
config
.
device
=
0
;
config
.
specify_input_name
=
true
;
config
.
enable_ir_optim
=
activate_ir
;
PADDLE_ENFORCE
(
config
.
ir_mode
==
AnalysisConfig
::
IrPassMode
::
kExclude
);
// default
config
.
ir_passes
.
clear
();
// Do not exclude any pass.
int
batch_size
=
FLAGS_batch_size
;
int
num_times
=
FLAGS_repeat
;
auto
base_predictor
=
CreatePaddlePredictor
<
NativeConfig
,
PaddleEngineKind
::
kNative
>
(
config
);
auto
predictor
=
CreatePaddlePredictor
<
AnalysisConfig
,
PaddleEngineKind
::
kAnalysis
>
(
config
);
std
::
vector
<
PaddleTensor
>
input_slots
;
DataRecord
data
(
FLAGS_infer_data
,
batch_size
);
// Prepare inputs.
PrepareInputs
(
&
input_slots
,
&
data
,
batch_size
);
std
::
vector
<
PaddleTensor
>
outputs
,
base_outputs
;
base_predictor
->
Run
(
input_slots
,
&
base_outputs
);
if
(
num_threads
==
1
)
{
// Prepare inputs.
Timer
timer
;
timer
.
tic
();
for
(
int
i
=
0
;
i
<
num_times
;
i
++
)
{
predictor
->
Run
(
input_slots
,
&
outputs
);
}
PrintTime
(
batch_size
,
num_times
,
1
,
0
,
timer
.
toc
()
/
num_times
);
CompareResult
(
outputs
,
base_outputs
);
}
else
{
std
::
vector
<
std
::
thread
>
threads
;
std
::
vector
<
std
::
unique_ptr
<
PaddlePredictor
>>
predictors
;
// TODO(yanchunwei): Bug here, the analyzer phase can't be parallelled
// because AttentionLSTM's hard code nodeid will be damanged.
for
(
int
tid
=
0
;
tid
<
num_threads
;
++
tid
)
{
predictors
.
emplace_back
(
CreatePaddlePredictor
<
AnalysisConfig
,
PaddleEngineKind
::
kAnalysis
>
(
config
));
}
for
(
int
tid
=
0
;
tid
<
num_threads
;
++
tid
)
{
threads
.
emplace_back
([
&
,
tid
]()
{
// Each thread should have local input_slots and outputs.
std
::
vector
<
PaddleTensor
>
input_slots
;
DataRecord
data
(
FLAGS_infer_data
,
batch_size
);
PrepareInputs
(
&
input_slots
,
&
data
,
batch_size
);
std
::
vector
<
PaddleTensor
>
outputs
;
Timer
timer
;
timer
.
tic
();
for
(
int
i
=
0
;
i
<
num_times
;
i
++
)
{
predictors
[
tid
]
->
Run
(
input_slots
,
&
outputs
);
}
PrintTime
(
batch_size
,
num_times
,
num_threads
,
tid
,
timer
.
toc
()
/
num_times
);
CompareResult
(
outputs
,
base_outputs
);
});
}
for
(
int
i
=
0
;
i
<
num_threads
;
++
i
)
{
threads
[
i
].
join
();
}
}
if
(
use_analysis
&&
activate_ir
)
{
AnalysisPredictor
*
analysis_predictor
=
dynamic_cast
<
AnalysisPredictor
*>
(
predictor
.
get
());
auto
&
fuse_statis
=
analysis_predictor
->
analysis_argument
()
.
Get
<
std
::
unordered_map
<
std
::
string
,
int
>>
(
framework
::
ir
::
kFuseStatisAttr
);
for
(
auto
&
item
:
fuse_statis
)
{
LOG
(
INFO
)
<<
"fused "
<<
item
.
first
<<
" "
<<
item
.
second
;
}
int
num_ops
=
0
;
for
(
auto
&
node
:
analysis_predictor
->
analysis_argument
().
main_dfg
->
nodes
.
nodes
())
{
if
(
node
->
IsFunction
())
{
++
num_ops
;
}
}
LOG
(
INFO
)
<<
"has num ops: "
<<
num_ops
;
ASSERT_TRUE
(
fuse_statis
.
count
(
"fc_fuse"
));
EXPECT_EQ
(
fuse_statis
.
at
(
"fc_fuse"
),
1
);
EXPECT_EQ
(
fuse_statis
.
at
(
"fc_nobias_lstm_fuse"
),
2
);
// bi-directional LSTM
EXPECT_EQ
(
fuse_statis
.
at
(
"seq_concat_fc_fuse"
),
1
);
EXPECT_EQ
(
num_ops
,
13
);
// After graph optimization, only 13 operators exists.
}
}
// Inference with analysis and IR, easy for profiling independently.
TEST
(
Analyzer
,
rnn1
)
{
TestRNN1Prediction
(
true
,
true
,
FLAGS_num_threads
);
}
// Other unit-tests of RNN1, test different options of use_analysis,
// activate_ir and multi-threads.
TEST
(
Analyzer
,
RNN_tests
)
{
int
num_threads
[
2
]
=
{
1
,
4
};
for
(
auto
i
:
num_threads
)
{
// Directly infer with the original model.
TestRNN1Prediction
(
false
,
false
,
i
);
// Inference with the original model with the analysis turned on, the
// analysis
// module will transform the program to a data flow graph.
TestRNN1Prediction
(
true
,
false
,
i
);
// Inference with analysis and IR. The IR module will fuse some large
// kernels.
TestRNN1Prediction
(
true
,
true
,
i
);
}
TEST
(
Analyzer
,
word2vec_without_analysis
)
{
TestWord2vecPrediction
(
FLAGS_inference_model_dir
);
}
}
// namespace analysis
...
...
paddle/fluid/inference/tests/api/CMakeLists.txt
0 → 100644
浏览文件 @
94b66bdb
function
(
inference_download_and_uncompress install_dir url
)
get_filename_component
(
filename
${
url
}
NAME
)
message
(
STATUS
"Download inference test stuff
${
filename
}
from
${
url
}
"
)
execute_process
(
COMMAND bash -c
"mkdir -p
${
install_dir
}
"
)
execute_process
(
COMMAND bash -c
"cd
${
install_dir
}
&& wget -q
${
url
}
"
)
execute_process
(
COMMAND bash -c
"cd
${
install_dir
}
&& tar xzf
${
filename
}
"
)
message
(
STATUS
"finish downloading
${
filename
}
"
)
endfunction
(
inference_download_and_uncompress
)
function
(
download_model_and_data install_dir model_url data_url
)
if
(
NOT EXISTS
${
install_dir
}
AND WITH_INFERENCE
)
inference_download_and_uncompress
(
${
install_dir
}
${
model_url
}
)
inference_download_and_uncompress
(
${
install_dir
}
${
data_url
}
)
endif
()
endfunction
()
# RNN1
set
(
RNN1_MODEL_URL
"http://paddle-inference-dist.bj.bcebos.com/rnn1%2Fmodel.tar.gz"
)
set
(
RNN1_DATA_URL
"http://paddle-inference-dist.bj.bcebos.com/rnn1%2Fdata.txt.tar.gz"
)
set
(
RNN1_INSTALL_DIR
"
${
THIRD_PARTY_PATH
}
/inference_demo/rnn1"
)
download_model_and_data
(
${
RNN1_INSTALL_DIR
}
${
RNN1_MODEL_URL
}
${
RNN1_DATA_URL
}
)
inference_analysis_test
(
test_analyzer_rnn1 SRCS analyzer_rnn1_tester.cc
EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor
ARGS --infer_model=
${
RNN1_INSTALL_DIR
}
/model
--infer_data=
${
RNN1_INSTALL_DIR
}
/data.txt
)
# chinese_ner
set
(
CHINESE_NER_MODEL_URL
"http://paddle-inference-dist.bj.bcebos.com/chinese_ner_model.tar.gz"
)
set
(
CHINESE_NER_DATA_URL
"http://paddle-inference-dist.bj.bcebos.com/chinese_ner-data.txt.tar.gz"
)
set
(
CHINESE_NER_INSTALL_DIR
"
${
THIRD_PARTY_PATH
}
/inference_demo/chinese_ner"
)
download_model_and_data
(
${
CHINESE_NER_INSTALL_DIR
}
${
CHINESE_NER_MODEL_URL
}
${
CHINESE_NER_DATA_URL
}
)
inference_analysis_test
(
test_analyzer_ner SRCS analyzer_ner_tester.cc
EXTRA_DEPS paddle_inference_api paddle_fluid_api analysis_predictor
ARGS --infer_model=
${
CHINESE_NER_INSTALL_DIR
}
/model
--infer_data=
${
CHINESE_NER_INSTALL_DIR
}
/data.txt
)
# lac
set
(
LAC_MODEL_URL
"http://paddle-inference-dist.bj.bcebos.com/lac_model.tar.gz"
)
set
(
LAC_DATA_URL
"http://paddle-inference-dist.bj.bcebos.com/lac_data.txt.tar.gz"
)
set
(
LAC_INSTALL_DIR
"
${
THIRD_PARTY_PATH
}
/inference_demo/lac"
)
download_model_and_data
(
${
LAC_INSTALL_DIR
}
${
LAC_MODEL_URL
}
${
LAC_DATA_URL
}
)
inference_analysis_test
(
test_analyzer_lac SRCS analyzer_lac_tester.cc
EXTRA_DEPS paddle_inference_api paddle_fluid_api ir_pass_manager analysis_predictor
ARGS --infer_model=
${
LAC_INSTALL_DIR
}
/model
--infer_data=
${
LAC_INSTALL_DIR
}
/data.txt
)
# text_classification
set
(
TEXT_CLASSIFICATION_MODEL_URL
"http://paddle-inference-dist.bj.bcebos.com/text-classification-Senta.tar.gz"
)
set
(
TEXT_CLASSIFICATION_DATA_URL
"http://paddle-inference-dist.bj.bcebos.com/text_classification_data.txt.tar.gz"
)
set
(
TEXT_CLASSIFICATION_INSTALL_DIR
"
${
THIRD_PARTY_PATH
}
/inference_demo/text_classification"
)
download_model_and_data
(
${
TEXT_CLASSIFICATION_INSTALL_DIR
}
${
TEXT_CLASSIFICATION_MODEL_URL
}
${
TEXT_CLASSIFICATION_DATA_URL
}
)
inference_analysis_test
(
test_text_classification SRCS analyzer_text_classification_tester.cc
EXTRA_DEPS paddle_inference_api paddle_fluid_api analysis_predictor
ARGS --infer_model=
${
TEXT_CLASSIFICATION_INSTALL_DIR
}
/text-classification-Senta
--infer_data=
${
TEXT_CLASSIFICATION_INSTALL_DIR
}
/data.txt
--topn=1
# Just run top 1 batch.
)
paddle/fluid/inference/
analysis
/analyzer_lac_tester.cc
→
paddle/fluid/inference/
tests/api
/analyzer_lac_tester.cc
浏览文件 @
94b66bdb
文件已移动
paddle/fluid/inference/
analysis
/analyzer_ner_tester.cc
→
paddle/fluid/inference/
tests/api
/analyzer_ner_tester.cc
浏览文件 @
94b66bdb
文件已移动
paddle/fluid/inference/tests/api/analyzer_rnn1_tester.cc
0 → 100644
浏览文件 @
94b66bdb
// Copyright (c) 2018 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.
#include "paddle/fluid/inference/analysis/analyzer.h"
#include <google/protobuf/text_format.h>
#include <gtest/gtest.h>
#include <thread> // NOLINT
#include "paddle/fluid/framework/ir/fuse_pass_base.h"
#include "paddle/fluid/framework/ir/pass.h"
#include "paddle/fluid/inference/analysis/ut_helper.h"
#include "paddle/fluid/inference/api/analysis_predictor.h"
#include "paddle/fluid/inference/api/helper.h"
#include "paddle/fluid/inference/api/paddle_inference_api.h"
#include "paddle/fluid/inference/api/paddle_inference_pass.h"
DEFINE_string
(
infer_model
,
""
,
"model path"
);
DEFINE_string
(
infer_data
,
""
,
"data path"
);
DEFINE_int32
(
batch_size
,
10
,
"batch size."
);
DEFINE_int32
(
repeat
,
1
,
"Running the inference program repeat times."
);
DEFINE_int32
(
num_threads
,
1
,
"Running the inference program in multi-threads."
);
namespace
paddle
{
namespace
inference
{
using
namespace
framework
;
// NOLINT
struct
DataRecord
{
std
::
vector
<
std
::
vector
<
std
::
vector
<
float
>>>
link_step_data_all
;
std
::
vector
<
std
::
vector
<
float
>>
week_data_all
,
minute_data_all
;
std
::
vector
<
size_t
>
lod1
,
lod2
,
lod3
;
std
::
vector
<
std
::
vector
<
float
>>
rnn_link_data
,
rnn_week_datas
,
rnn_minute_datas
;
size_t
batch_iter
{
0
};
size_t
batch_size
{
1
};
DataRecord
()
=
default
;
explicit
DataRecord
(
const
std
::
string
&
path
,
int
batch_size
=
1
)
:
batch_size
(
batch_size
)
{
Load
(
path
);
}
DataRecord
NextBatch
()
{
DataRecord
data
;
size_t
batch_end
=
batch_iter
+
batch_size
;
// NOTE skip the final batch, if no enough data is provided.
if
(
batch_end
<=
link_step_data_all
.
size
())
{
data
.
link_step_data_all
.
assign
(
link_step_data_all
.
begin
()
+
batch_iter
,
link_step_data_all
.
begin
()
+
batch_end
);
data
.
week_data_all
.
assign
(
week_data_all
.
begin
()
+
batch_iter
,
week_data_all
.
begin
()
+
batch_end
);
data
.
minute_data_all
.
assign
(
minute_data_all
.
begin
()
+
batch_iter
,
minute_data_all
.
begin
()
+
batch_end
);
// Prepare LoDs
data
.
lod1
.
push_back
(
0
);
data
.
lod2
.
push_back
(
0
);
data
.
lod3
.
push_back
(
0
);
CHECK
(
!
data
.
link_step_data_all
.
empty
())
<<
"empty"
;
CHECK
(
!
data
.
week_data_all
.
empty
());
CHECK
(
!
data
.
minute_data_all
.
empty
());
CHECK_EQ
(
data
.
link_step_data_all
.
size
(),
data
.
week_data_all
.
size
());
CHECK_EQ
(
data
.
minute_data_all
.
size
(),
data
.
link_step_data_all
.
size
());
for
(
size_t
j
=
0
;
j
<
data
.
link_step_data_all
.
size
();
j
++
)
{
for
(
const
auto
&
d
:
data
.
link_step_data_all
[
j
])
{
data
.
rnn_link_data
.
push_back
(
d
);
}
data
.
rnn_week_datas
.
push_back
(
data
.
week_data_all
[
j
]);
data
.
rnn_minute_datas
.
push_back
(
data
.
minute_data_all
[
j
]);
// calculate lod
data
.
lod1
.
push_back
(
data
.
lod1
.
back
()
+
data
.
link_step_data_all
[
j
].
size
());
data
.
lod3
.
push_back
(
data
.
lod3
.
back
()
+
1
);
for
(
size_t
i
=
1
;
i
<
data
.
link_step_data_all
[
j
].
size
()
+
1
;
i
++
)
{
data
.
lod2
.
push_back
(
data
.
lod2
.
back
()
+
data
.
link_step_data_all
[
j
].
size
());
}
}
}
batch_iter
+=
batch_size
;
return
data
;
}
void
Load
(
const
std
::
string
&
path
)
{
std
::
ifstream
file
(
path
);
std
::
string
line
;
int
num_lines
=
0
;
while
(
std
::
getline
(
file
,
line
))
{
num_lines
++
;
std
::
vector
<
std
::
string
>
data
;
split
(
line
,
':'
,
&
data
);
std
::
vector
<
std
::
vector
<
float
>>
link_step_data
;
std
::
vector
<
std
::
string
>
link_datas
;
split
(
data
[
0
],
'|'
,
&
link_datas
);
for
(
auto
&
step_data
:
link_datas
)
{
std
::
vector
<
float
>
tmp
;
split_to_float
(
step_data
,
','
,
&
tmp
);
link_step_data
.
push_back
(
tmp
);
}
// load week data
std
::
vector
<
float
>
week_data
;
split_to_float
(
data
[
2
],
','
,
&
week_data
);
// load minute data
std
::
vector
<
float
>
minute_data
;
split_to_float
(
data
[
1
],
','
,
&
minute_data
);
link_step_data_all
.
push_back
(
std
::
move
(
link_step_data
));
week_data_all
.
push_back
(
std
::
move
(
week_data
));
minute_data_all
.
push_back
(
std
::
move
(
minute_data
));
}
}
};
void
PrepareInputs
(
std
::
vector
<
PaddleTensor
>
*
input_slots
,
DataRecord
*
data
,
int
batch_size
)
{
PaddleTensor
lod_attention_tensor
,
init_zero_tensor
,
lod_tensor_tensor
,
week_tensor
,
minute_tensor
;
lod_attention_tensor
.
name
=
"data_lod_attention"
;
init_zero_tensor
.
name
=
"cell_init"
;
lod_tensor_tensor
.
name
=
"data"
;
week_tensor
.
name
=
"week"
;
minute_tensor
.
name
=
"minute"
;
auto
one_batch
=
data
->
NextBatch
();
std
::
vector
<
int
>
rnn_link_data_shape
(
{
static_cast
<
int
>
(
one_batch
.
rnn_link_data
.
size
()),
static_cast
<
int
>
(
one_batch
.
rnn_link_data
.
front
().
size
())});
lod_attention_tensor
.
shape
.
assign
({
1
,
2
});
lod_attention_tensor
.
lod
.
assign
({
one_batch
.
lod1
,
one_batch
.
lod2
});
init_zero_tensor
.
shape
.
assign
({
batch_size
,
15
});
init_zero_tensor
.
lod
.
assign
({
one_batch
.
lod3
});
lod_tensor_tensor
.
shape
=
rnn_link_data_shape
;
lod_tensor_tensor
.
lod
.
assign
({
one_batch
.
lod1
});
// clang-format off
week_tensor
.
shape
.
assign
(
{
static_cast
<
int
>
(
one_batch
.
rnn_week_datas
.
size
()),
static_cast
<
int
>
(
one_batch
.
rnn_week_datas
.
front
().
size
())});
week_tensor
.
lod
.
assign
({
one_batch
.
lod3
});
minute_tensor
.
shape
.
assign
(
{
static_cast
<
int
>
(
one_batch
.
rnn_minute_datas
.
size
()),
static_cast
<
int
>
(
one_batch
.
rnn_minute_datas
.
front
().
size
())});
minute_tensor
.
lod
.
assign
({
one_batch
.
lod3
});
// clang-format on
// assign data
TensorAssignData
<
float
>
(
&
lod_attention_tensor
,
std
::
vector
<
std
::
vector
<
float
>>
({{
0
,
0
}}));
std
::
vector
<
float
>
tmp_zeros
(
batch_size
*
15
,
0.
);
TensorAssignData
<
float
>
(
&
init_zero_tensor
,
{
tmp_zeros
});
TensorAssignData
<
float
>
(
&
lod_tensor_tensor
,
one_batch
.
rnn_link_data
);
TensorAssignData
<
float
>
(
&
week_tensor
,
one_batch
.
rnn_week_datas
);
TensorAssignData
<
float
>
(
&
minute_tensor
,
one_batch
.
rnn_minute_datas
);
// Set inputs.
auto
init_zero_tensor1
=
init_zero_tensor
;
init_zero_tensor1
.
name
=
"hidden_init"
;
input_slots
->
assign
({
week_tensor
,
init_zero_tensor
,
minute_tensor
,
init_zero_tensor1
,
lod_attention_tensor
,
lod_tensor_tensor
});
for
(
auto
&
tensor
:
*
input_slots
)
{
tensor
.
dtype
=
PaddleDType
::
FLOAT32
;
}
}
void
CompareResult
(
const
std
::
vector
<
PaddleTensor
>
&
outputs
,
const
std
::
vector
<
PaddleTensor
>
&
base_outputs
)
{
PADDLE_ENFORCE_GT
(
outputs
.
size
(),
0
);
PADDLE_ENFORCE_EQ
(
outputs
.
size
(),
base_outputs
.
size
());
for
(
size_t
i
=
0
;
i
<
outputs
.
size
();
i
++
)
{
auto
&
out
=
outputs
[
i
];
auto
&
base_out
=
base_outputs
[
i
];
size_t
size
=
std
::
accumulate
(
out
.
shape
.
begin
(),
out
.
shape
.
end
(),
1
,
[](
int
a
,
int
b
)
{
return
a
*
b
;
});
size_t
size1
=
std
::
accumulate
(
base_out
.
shape
.
begin
(),
base_out
.
shape
.
end
(),
1
,
[](
int
a
,
int
b
)
{
return
a
*
b
;
});
PADDLE_ENFORCE_EQ
(
size
,
size1
);
PADDLE_ENFORCE_GT
(
size
,
0
);
float
*
data
=
static_cast
<
float
*>
(
out
.
data
.
data
());
float
*
base_data
=
static_cast
<
float
*>
(
base_out
.
data
.
data
());
for
(
size_t
i
=
0
;
i
<
size
;
i
++
)
{
EXPECT_NEAR
(
data
[
i
],
base_data
[
i
],
1e-3
);
}
}
}
// Test with a really complicate model.
void
TestRNN1Prediction
(
bool
use_analysis
,
bool
activate_ir
,
int
num_threads
)
{
AnalysisConfig
config
;
config
.
prog_file
=
FLAGS_infer_model
+
"/__model__"
;
config
.
param_file
=
FLAGS_infer_model
+
"/param"
;
config
.
use_gpu
=
false
;
config
.
device
=
0
;
config
.
specify_input_name
=
true
;
config
.
enable_ir_optim
=
activate_ir
;
PADDLE_ENFORCE
(
config
.
ir_mode
==
AnalysisConfig
::
IrPassMode
::
kExclude
);
// default
config
.
ir_passes
.
clear
();
// Do not exclude any pass.
int
batch_size
=
FLAGS_batch_size
;
int
num_times
=
FLAGS_repeat
;
auto
base_predictor
=
CreatePaddlePredictor
<
NativeConfig
,
PaddleEngineKind
::
kNative
>
(
config
);
auto
predictor
=
CreatePaddlePredictor
<
AnalysisConfig
,
PaddleEngineKind
::
kAnalysis
>
(
config
);
std
::
vector
<
PaddleTensor
>
input_slots
;
DataRecord
data
(
FLAGS_infer_data
,
batch_size
);
// Prepare inputs.
PrepareInputs
(
&
input_slots
,
&
data
,
batch_size
);
std
::
vector
<
PaddleTensor
>
outputs
,
base_outputs
;
base_predictor
->
Run
(
input_slots
,
&
base_outputs
);
if
(
num_threads
==
1
)
{
// Prepare inputs.
Timer
timer
;
timer
.
tic
();
for
(
int
i
=
0
;
i
<
num_times
;
i
++
)
{
predictor
->
Run
(
input_slots
,
&
outputs
);
}
PrintTime
(
batch_size
,
num_times
,
1
,
0
,
timer
.
toc
()
/
num_times
);
CompareResult
(
outputs
,
base_outputs
);
}
else
{
std
::
vector
<
std
::
thread
>
threads
;
std
::
vector
<
std
::
unique_ptr
<
PaddlePredictor
>>
predictors
;
// TODO(yanchunwei): Bug here, the analyzer phase can't be parallelled
// because AttentionLSTM's hard code nodeid will be damanged.
for
(
int
tid
=
0
;
tid
<
num_threads
;
++
tid
)
{
predictors
.
emplace_back
(
CreatePaddlePredictor
<
AnalysisConfig
,
PaddleEngineKind
::
kAnalysis
>
(
config
));
}
for
(
int
tid
=
0
;
tid
<
num_threads
;
++
tid
)
{
threads
.
emplace_back
([
&
,
tid
]()
{
// Each thread should have local input_slots and outputs.
std
::
vector
<
PaddleTensor
>
input_slots
;
DataRecord
data
(
FLAGS_infer_data
,
batch_size
);
PrepareInputs
(
&
input_slots
,
&
data
,
batch_size
);
std
::
vector
<
PaddleTensor
>
outputs
;
Timer
timer
;
timer
.
tic
();
for
(
int
i
=
0
;
i
<
num_times
;
i
++
)
{
predictors
[
tid
]
->
Run
(
input_slots
,
&
outputs
);
}
PrintTime
(
batch_size
,
num_times
,
num_threads
,
tid
,
timer
.
toc
()
/
num_times
);
CompareResult
(
outputs
,
base_outputs
);
});
}
for
(
int
i
=
0
;
i
<
num_threads
;
++
i
)
{
threads
[
i
].
join
();
}
}
if
(
use_analysis
&&
activate_ir
)
{
AnalysisPredictor
*
analysis_predictor
=
dynamic_cast
<
AnalysisPredictor
*>
(
predictor
.
get
());
auto
&
fuse_statis
=
analysis_predictor
->
analysis_argument
()
.
Get
<
std
::
unordered_map
<
std
::
string
,
int
>>
(
framework
::
ir
::
kFuseStatisAttr
);
for
(
auto
&
item
:
fuse_statis
)
{
LOG
(
INFO
)
<<
"fused "
<<
item
.
first
<<
" "
<<
item
.
second
;
}
int
num_ops
=
0
;
for
(
auto
&
node
:
analysis_predictor
->
analysis_argument
().
main_dfg
->
nodes
.
nodes
())
{
if
(
node
->
IsFunction
())
{
++
num_ops
;
}
}
LOG
(
INFO
)
<<
"has num ops: "
<<
num_ops
;
ASSERT_TRUE
(
fuse_statis
.
count
(
"fc_fuse"
));
EXPECT_EQ
(
fuse_statis
.
at
(
"fc_fuse"
),
1
);
EXPECT_EQ
(
fuse_statis
.
at
(
"fc_nobias_lstm_fuse"
),
2
);
// bi-directional LSTM
EXPECT_EQ
(
fuse_statis
.
at
(
"seq_concat_fc_fuse"
),
1
);
EXPECT_EQ
(
num_ops
,
13
);
// After graph optimization, only 13 operators exists.
}
}
// Inference with analysis and IR, easy for profiling independently.
TEST
(
Analyzer
,
rnn1
)
{
TestRNN1Prediction
(
true
,
true
,
FLAGS_num_threads
);
}
// Other unit-tests of RNN1, test different options of use_analysis,
// activate_ir and multi-threads.
TEST
(
Analyzer
,
RNN_tests
)
{
int
num_threads
[
2
]
=
{
1
,
4
};
for
(
auto
i
:
num_threads
)
{
// Directly infer with the original model.
TestRNN1Prediction
(
false
,
false
,
i
);
// Inference with the original model with the analysis turned on, the
// analysis
// module will transform the program to a data flow graph.
TestRNN1Prediction
(
true
,
false
,
i
);
// Inference with analysis and IR. The IR module will fuse some large
// kernels.
TestRNN1Prediction
(
true
,
true
,
i
);
}
}
}
// namespace inference
}
// namespace paddle
paddle/fluid/inference/
analysis
/analyzer_text_classification_tester.cc
→
paddle/fluid/inference/
tests/api
/analyzer_text_classification_tester.cc
浏览文件 @
94b66bdb
文件已移动
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