!16729 mass & transformer & advanced_east 310 infer

From: @zhangxiaoxiao16
Reviewed-by: @c_34,@linqingke
Signed-off-by: @c_34
This commit is contained in:
mindspore-ci-bot 2021-05-24 19:28:05 +08:00 committed by Gitee
commit 38afecfddc
29 changed files with 1988 additions and 0 deletions

View File

@ -26,6 +26,10 @@
- [Pre-training](#pre-training)
- [Fine-tuning](#fine-tuning)
- [Inference](#inference)
- [Mindir Inference Process](#mindir-inference-process)
- [Export MindIR](#export-mindir)
- [Infer on Ascend310](#infer-on-ascend310)
- [result](#result)
- [Performance](#performance)
- [Results](#results)
- [Fine-Tuning on Text Summarization](#fine-tuning-on-text-summarization)
@ -587,6 +591,33 @@ You can also run the shell script `run_gpu.sh` on gpu as followed:
sh run_gpu.sh -t i -n 1 -i 1 -c config/config.json -o {outputfile}
```
## Mindir Inference Process
### [Export MindIR](#contents)
```shell
python export.py --ckpt_file [CKPT_PATH] --file_name [FILE_NAME] --file_format [FILE_FORMAT]
```
The ckpt_file parameter is required,
`EXPORT_FORMAT` should be in ["AIR", "MINDIR"]
### Infer on Ascend310
Before performing inference, the mindir file must be exported by `export.py` script. We only provide an example of inference using MINDIR model.
```shell
# Ascend310 inference
bash run_infer_310.sh [MINDIR_PATH] [CONFIG] [VOCAB] [OUTPUT] [NEED_PREPROCESS] [DEVICE_ID]
```
- `NEED_PREPROCESS` means weather need preprocess or not, it's value is 'y' or 'n'.
- `DEVICE_ID` is optional, default value is 0.
### result
Inference result is saved in current path, you can find result in acc.log file.
# Performance
## Results

View File

@ -27,6 +27,10 @@
- [预训练](#预训练)
- [微调](#微调)
- [推理](#推理)
- [Mindir推理](#Mindir推理)
- [导出模型](#导出模型)
- [在Ascend310执行推理](#在Ascend310执行推理)
- [结果](#结果)
- [性能](#性能)
- [结果](#结果)
- [文本摘要微调](#文本摘要微调)
@ -586,6 +590,33 @@ sh run_ascend.sh -t i -n 1 -i 1 -c config/config.json -o {outputfile}
sh run_gpu.sh -t i -n 1 -i 1 -c config/config.json -o {outputfile}
```
## Mindir推理
### [导出模型](#contents)
```shell
python export.py --ckpt_file [CKPT_PATH] --file_name [FILE_NAME] --file_format [FILE_FORMAT]
```
参数ckpt_file为必填项
`EXPORT_FORMAT` 必须在 ["AIR", "MINDIR"]中选择。
### 在Ascend310执行推理
在执行推理前mindir文件必须通过`export.py`脚本导出。以下展示了使用minir模型执行推理的示例。
```shell
# Ascend310推理
bash run_infer_310.sh [MINDIR_PATH] [CONFIG] [VOCAB] [OUTPUT] [NEED_PREPROCESS] [DEVICE_ID]
```
- `NEED_PREPROCESS` 表示数据是否需要预处理,取值范围为:'y' 或者 'n'。
- `DEVICE_ID` 可选默认值为0。
### 结果
推理结果保存在脚本执行的当前路径精度计算结果可以在acc.log中看到。
# 性能
## 结果

View File

@ -0,0 +1,14 @@
cmake_minimum_required(VERSION 3.14.1)
project(Ascend310Infer)
add_compile_definitions(_GLIBCXX_USE_CXX11_ABI=0)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O0 -g -std=c++17 -Werror -Wall -fPIE -Wl,--allow-shlib-undefined")
set(PROJECT_SRC_ROOT ${CMAKE_CURRENT_LIST_DIR}/)
option(MINDSPORE_PATH "mindspore install path" "")
include_directories(${MINDSPORE_PATH})
include_directories(${MINDSPORE_PATH}/include)
include_directories(${PROJECT_SRC_ROOT})
find_library(MS_LIB libmindspore.so ${MINDSPORE_PATH}/lib)
file(GLOB_RECURSE MD_LIB ${MINDSPORE_PATH}/_c_dataengine*)
add_executable(main src/main.cc src/utils.cc)
target_link_libraries(main ${MS_LIB} ${MD_LIB} gflags)

View File

@ -0,0 +1,29 @@
#!/bin/bash
# Copyright 2021 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
if [ -d out ]; then
rm -rf out
fi
mkdir out
cd out || exit
if [ -f "Makefile" ]; then
make clean
fi
cmake .. \
-DMINDSPORE_PATH="`pip3.7 show mindspore-ascend | grep Location | awk '{print $2"/mindspore"}' | xargs realpath`"
make

View File

@ -0,0 +1,32 @@
/**
* Copyright 2021 Huawei Technologies Co., Ltd
*
* 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.
*/
#ifndef MINDSPORE_INFERENCE_UTILS_H_
#define MINDSPORE_INFERENCE_UTILS_H_
#include <sys/stat.h>
#include <dirent.h>
#include <vector>
#include <string>
#include <memory>
#include "include/api/types.h"
std::vector<std::string> GetAllFiles(std::string_view dirName);
DIR *OpenDir(std::string_view dirName);
std::string RealPath(std::string_view path);
mindspore::MSTensor ReadFileToTensor(const std::string &file);
int WriteResult(const std::string& imageFile, const std::vector<mindspore::MSTensor> &outputs);
#endif

View File

@ -0,0 +1,137 @@
/**
* Copyright 2021 Huawei Technologies Co., Ltd
*
* 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 <sys/time.h>
#include <gflags/gflags.h>
#include <dirent.h>
#include <iostream>
#include <string>
#include <algorithm>
#include <iosfwd>
#include <vector>
#include <fstream>
#include <sstream>
#include "include/api/model.h"
#include "include/api/context.h"
#include "include/api/types.h"
#include "include/api/serialization.h"
#include "include/dataset/execute.h"
#include "include/dataset/vision.h"
#include "inc/utils.h"
using mindspore::Context;
using mindspore::Serialization;
using mindspore::Model;
using mindspore::Status;
using mindspore::MSTensor;
using mindspore::dataset::Execute;
using mindspore::ModelType;
using mindspore::GraphCell;
using mindspore::kSuccess;
DEFINE_string(mindir_path, "", "mindir path");
DEFINE_string(input0_path, ".", "input0 path");
DEFINE_string(input1_path, ".", "input1 path");
DEFINE_int32(device_id, 0, "device id");
int main(int argc, char **argv) {
gflags::ParseCommandLineFlags(&argc, &argv, true);
if (RealPath(FLAGS_mindir_path).empty()) {
std::cout << "Invalid mindir" << std::endl;
return 1;
}
auto context = std::make_shared<Context>();
auto ascend310 = std::make_shared<mindspore::Ascend310DeviceInfo>();
ascend310->SetDeviceID(FLAGS_device_id);
ascend310->SetPrecisionMode("allow_fp32_to_fp16");
ascend310->SetOpSelectImplMode("high_precision");
context->MutableDeviceInfo().push_back(ascend310);
mindspore::Graph graph;
Serialization::Load(FLAGS_mindir_path, ModelType::kMindIR, &graph);
Model model;
Status ret = model.Build(GraphCell(graph), context);
if (ret != kSuccess) {
std::cout << "ERROR: Build failed." << std::endl;
return 1;
}
std::vector<MSTensor> model_inputs = model.GetInputs();
if (model_inputs.empty()) {
std::cout << "Invalid model, inputs is empty." << std::endl;
return 1;
}
auto input0_files = GetAllFiles(FLAGS_input0_path);
auto input1_files = GetAllFiles(FLAGS_input1_path);
if (input0_files.empty() || input1_files.empty()) {
std::cout << "ERROR: input data empty." << std::endl;
return 1;
}
std::map<double, double> costTime_map;
size_t size = input0_files.size();
for (size_t i = 0; i < size; ++i) {
struct timeval start = {0};
struct timeval end = {0};
double startTimeMs;
double endTimeMs;
std::vector<MSTensor> inputs;
std::vector<MSTensor> outputs;
std::cout << "Start predict input files:" << input0_files[i] << std::endl;
auto input0 = ReadFileToTensor(input0_files[i]);
auto input1 = ReadFileToTensor(input1_files[i]);
inputs.emplace_back(model_inputs[0].Name(), model_inputs[0].DataType(), model_inputs[0].Shape(),
input0.Data().get(), input0.DataSize());
inputs.emplace_back(model_inputs[1].Name(), model_inputs[1].DataType(), model_inputs[1].Shape(),
input1.Data().get(), input1.DataSize());
gettimeofday(&start, nullptr);
ret = model.Predict(inputs, &outputs);
gettimeofday(&end, nullptr);
if (ret != kSuccess) {
std::cout << "Predict " << input0_files[i] << " failed." << std::endl;
return 1;
}
startTimeMs = (1.0 * start.tv_sec * 1000000 + start.tv_usec) / 1000;
endTimeMs = (1.0 * end.tv_sec * 1000000 + end.tv_usec) / 1000;
costTime_map.insert(std::pair<double, double>(startTimeMs, endTimeMs));
WriteResult(input0_files[i], outputs);
}
double average = 0.0;
int inferCount = 0;
for (auto iter = costTime_map.begin(); iter != costTime_map.end(); iter++) {
double diff = 0.0;
diff = iter->second - iter->first;
average += diff;
inferCount++;
}
average = average / inferCount;
std::stringstream timeCost;
timeCost << "NN inference cost average time: "<< average << " ms of infer_count " << inferCount << std::endl;
std::cout << "NN inference cost average time: "<< average << "ms of infer_count " << inferCount << std::endl;
std::string fileName = "./time_Result" + std::string("/test_perform_static.txt");
std::ofstream fileStream(fileName.c_str(), std::ios::trunc);
fileStream << timeCost.str();
fileStream.close();
costTime_map.clear();
return 0;
}

View File

@ -0,0 +1,129 @@
/**
* Copyright 2021 Huawei Technologies Co., Ltd
*
* 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 <fstream>
#include <algorithm>
#include <iostream>
#include "inc/utils.h"
using mindspore::MSTensor;
using mindspore::DataType;
std::vector<std::string> GetAllFiles(std::string_view dirName) {
struct dirent *filename;
DIR *dir = OpenDir(dirName);
if (dir == nullptr) {
return {};
}
std::vector<std::string> res;
while ((filename = readdir(dir)) != nullptr) {
std::string dName = std::string(filename->d_name);
if (dName == "." || dName == ".." || filename->d_type != DT_REG) {
continue;
}
res.emplace_back(std::string(dirName) + "/" + filename->d_name);
}
std::sort(res.begin(), res.end());
for (auto &f : res) {
std::cout << "image file: " << f << std::endl;
}
return res;
}
int WriteResult(const std::string& imageFile, const std::vector<MSTensor> &outputs) {
std::string homePath = "./result_Files";
for (size_t i = 0; i < outputs.size(); ++i) {
size_t outputSize;
std::shared_ptr<const void> netOutput;
netOutput = outputs[i].Data();
outputSize = outputs[i].DataSize();
int pos = imageFile.rfind('/');
std::string fileName(imageFile, pos + 1);
fileName.replace(fileName.find('.'), fileName.size() - fileName.find('.'), '_' + std::to_string(i) + ".bin");
std::string outFileName = homePath + "/" + fileName;
FILE * outputFile = fopen(outFileName.c_str(), "wb");
fwrite(netOutput.get(), outputSize, sizeof(char), outputFile);
fclose(outputFile);
outputFile = nullptr;
}
return 0;
}
mindspore::MSTensor ReadFileToTensor(const std::string &file) {
if (file.empty()) {
std::cout << "Pointer file is nullptr" << std::endl;
return mindspore::MSTensor();
}
std::ifstream ifs(file);
if (!ifs.good()) {
std::cout << "File: " << file << " is not exist" << std::endl;
return mindspore::MSTensor();
}
if (!ifs.is_open()) {
std::cout << "File: " << file << "open failed" << std::endl;
return mindspore::MSTensor();
}
ifs.seekg(0, std::ios::end);
size_t size = ifs.tellg();
mindspore::MSTensor buffer(file, mindspore::DataType::kNumberTypeUInt8, {static_cast<int64_t>(size)}, nullptr, size);
ifs.seekg(0, std::ios::beg);
ifs.read(reinterpret_cast<char *>(buffer.MutableData()), size);
ifs.close();
return buffer;
}
DIR *OpenDir(std::string_view dirName) {
if (dirName.empty()) {
std::cout << " dirName is null ! " << std::endl;
return nullptr;
}
std::string realPath = RealPath(dirName);
struct stat s;
lstat(realPath.c_str(), &s);
if (!S_ISDIR(s.st_mode)) {
std::cout << "dirName is not a valid directory !" << std::endl;
return nullptr;
}
DIR *dir;
dir = opendir(realPath.c_str());
if (dir == nullptr) {
std::cout << "Can not open dir " << dirName << std::endl;
return nullptr;
}
std::cout << "Successfully opened the dir " << dirName << std::endl;
return dir;
}
std::string RealPath(std::string_view path) {
char realPathMem[PATH_MAX] = {0};
char *realPathRet = nullptr;
realPathRet = realpath(path.data(), realPathMem);
if (realPathRet == nullptr) {
std::cout << "File: " << path << " is not exist.";
return "";
}
std::string realPath(realPathMem);
std::cout << path << " realpath is: " << realPath << std::endl;
return realPath;
}

View File

@ -0,0 +1,99 @@
# Copyright 2021 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
"""Evaluation api."""
import os
import argparse
import pickle
import numpy as np
from config import TransformerConfig
from src.utils import Dictionary
from src.utils import get_score
parser = argparse.ArgumentParser(description='postprocess.')
parser.add_argument("--config", type=str, required=True,
help="Model config json file path.")
parser.add_argument("--vocab", type=str, required=True,
help="Vocabulary to use.")
parser.add_argument("--output", type=str, required=True,
help="Result file path.")
parser.add_argument("--metric", type=str, default='rouge',
help='Set eval method.')
parser.add_argument("--source_id_folder", type=str, default='',
help="source_eos_ids folder path.")
parser.add_argument("--target_id_folder", type=str, default='',
help="target_eos_ids folder path.")
parser.add_argument("--result_dir", type=str, default='./result_Files',
help="result dir path.")
args, _ = parser.parse_known_args()
def read_from_file(config):
'''
calculate accuraty.
'''
predictions = []
probs = []
source_sentences = []
target_sentences = []
file_num = len(os.listdir(args.source_id_folder))
for i in range(file_num):
f_name = "gigaword_bs_" + str(config.batch_size) + "_" + str(i)
source_ids = np.fromfile(os.path.join(args.source_id_folder, f_name + ".bin"), np.int32)
source_ids = source_ids.reshape(1, config.max_decode_length)
target_ids = np.fromfile(os.path.join(args.target_id_folder, f_name + ".bin"), np.int32)
target_ids = target_ids.reshape(1, config.max_decode_length)
predicted_ids = np.fromfile(os.path.join(args.result_dir, f_name + "_0.bin"), np.int32)
predicted_ids = predicted_ids.reshape(1, config.max_decode_length + 1)
entire_probs = np.fromfile(os.path.join(args.result_dir, f_name + "_1.bin"), np.float32)
entire_probs = entire_probs.reshape(1, config.beam_width, config.max_decode_length + 1)
source_sentences.append(source_ids)
target_sentences.append(target_ids)
predictions.append(predicted_ids)
probs.append(entire_probs)
output = []
for inputs, ref, batch_out, batch_probs in zip(source_sentences,
target_sentences,
predictions,
probs):
for i in range(config.batch_size):
if batch_out.ndim == 3:
batch_out = batch_out[:, 0]
example = {
"source": inputs[i].tolist(),
"target": ref[i].tolist(),
"prediction": batch_out[i].tolist(),
"prediction_prob": batch_probs[i].tolist()
}
output.append(example)
return output
if __name__ == '__main__':
conf = TransformerConfig.from_json_file(args.config)
result = read_from_file(conf)
vocab = Dictionary.load_from_persisted_dict(args.vocab)
with open(args.output, "wb") as f:
pickle.dump(result, f, 1)
# get score by given metric
score = get_score(result, vocab, metric=args.metric)
print(score)

View File

@ -0,0 +1,63 @@
# Copyright 2021 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
"""Evaluation api."""
import os
import argparse
from config import TransformerConfig
from src.dataset import load_dataset
parser = argparse.ArgumentParser(description='preprocess.')
parser.add_argument("--config", type=str, required=True,
help="Model config json file path.")
parser.add_argument("--result_path", type=str, default='./preprocess_Result/',
help="preprocess result path.")
args, _ = parser.parse_known_args()
def generate_bin():
'''
Generate bin files.
'''
config = TransformerConfig.from_json_file(args.config)
ds = load_dataset(data_files=config.test_dataset,
batch_size=config.batch_size,
epoch_count=1,
sink_mode=config.dataset_sink_mode,
shuffle=False) if config.test_dataset else None
cur_dir = args.result_path
source_eos_ids_path = os.path.join(cur_dir, "00_source_eos_ids")
source_eos_mask_path = os.path.join(cur_dir, "01_source_eos_mask")
target_eos_ids_path = os.path.join(cur_dir, " target_eos_ids")
if not os.path.isdir(source_eos_ids_path):
os.makedirs(source_eos_ids_path)
if not os.path.isdir(source_eos_mask_path):
os.makedirs(source_eos_mask_path)
if not os.path.isdir(target_eos_ids_path):
os.makedirs(target_eos_ids_path)
for i, data in enumerate(ds.create_dict_iterator(output_numpy=True, num_epochs=1)):
file_name = "gigaword_bs_" + str(config.batch_size) + "_" + str(i) + ".bin"
source_eos_ids = data['source_eos_ids']
source_eos_ids.tofile(os.path.join(source_eos_ids_path, file_name))
source_eos_mask = data['source_eos_mask']
source_eos_mask.tofile(os.path.join(source_eos_mask_path, file_name))
target_eos_ids = data['target_eos_ids']
target_eos_ids.tofile(os.path.join(target_eos_ids_path, file_name))
print("="*20, "export bin files finished", "="*20)
if __name__ == '__main__':
generate_bin()

View File

@ -0,0 +1,126 @@
#!/bin/bash
# Copyright 2021 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
if [[ $# -lt 5 || $# -gt 6 ]]; then
echo "Usage: bash run_infer_310.sh [MINDIR_PATH] [CONFIG] [VOCAB] [OUTPUT] [NEED_PREPROCESS] [DEVICE_ID]
NEED_PREPROCESS means weather need preprocess or not, it's value is 'y' or 'n'.
DEVICE_ID is optional, it can be set by environment variable device_id, otherwise the value is zero"
exit 1
fi
get_real_path(){
if [ "${1:0:1}" == "/" ]; then
echo "$1"
else
echo "$(realpath -m $PWD/$1)"
fi
}
model=$(get_real_path $1)
config=$(get_real_path $2)
vocab=$(get_real_path $3)
output=$(get_real_path $4)
if [ "$5" == "y" ] || [ "$5" == "n" ];then
need_preprocess=$5
else
echo "weather need preprocess or not, it's value must be in [y, n]"
exit 1
fi
device_id=0
if [ $# == 6 ]; then
device_id=$6
fi
echo "mindir name: "$model
echo "config: "$config
echo "vocab: "$vocab
echo "output: "$output
echo "need preprocess: "$need_preprocess
echo "device id: "$device_id
export ASCEND_HOME=/usr/local/Ascend/
if [ -d ${ASCEND_HOME}/ascend-toolkit ]; then
export PATH=$ASCEND_HOME/ascend-toolkit/latest/fwkacllib/ccec_compiler/bin:$ASCEND_HOME/ascend-toolkit/latest/atc/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/lib:$ASCEND_HOME/ascend-toolkit/latest/atc/lib64:$ASCEND_HOME/ascend-toolkit/latest/fwkacllib/lib64:$ASCEND_HOME/driver/lib64:$ASCEND_HOME/add-ons:$LD_LIBRARY_PATH
export TBE_IMPL_PATH=$ASCEND_HOME/ascend-toolkit/latest/opp/op_impl/built-in/ai_core/tbe
export PYTHONPATH=${TBE_IMPL_PATH}:$ASCEND_HOME/ascend-toolkit/latest/fwkacllib/python/site-packages:$PYTHONPATH
export ASCEND_OPP_PATH=$ASCEND_HOME/ascend-toolkit/latest/opp
else
export PATH=$ASCEND_HOME/atc/ccec_compiler/bin:$ASCEND_HOME/atc/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/lib:$ASCEND_HOME/atc/lib64:$ASCEND_HOME/acllib/lib64:$ASCEND_HOME/driver/lib64:$ASCEND_HOME/add-ons:$LD_LIBRARY_PATH
export PYTHONPATH=$ASCEND_HOME/atc/python/site-packages:$PYTHONPATH
export ASCEND_OPP_PATH=$ASCEND_HOME/opp
fi
function preprocess_data()
{
if [ -d preprocess_Result ]; then
rm -rf ./preprocess_Result
fi
mkdir preprocess_Result
python3.7 ../preprocess.py --config=$config --result_path=./preprocess_Result/
}
function compile_app()
{
cd ../ascend310_infer || exit
bash build.sh &> build.log
}
function infer()
{
cd - || exit
if [ -d result_Files ]; then
rm -rf ./result_Files
fi
if [ -d time_Result ]; then
rm -rf ./time_Result
fi
mkdir result_Files
mkdir time_Result
../ascend310_infer/out/main --mindir_path=$model --input0_path=./preprocess_Result/00_source_eos_ids --input1_path=./preprocess_Result/01_source_eos_mask --device_id=$device_id &> infer.log
}
function cal_acc()
{
python3.7 ../postprocess.py --config=$config --vocab=$vocab --output=$output --source_id_folder=./preprocess_Result/00_source_eos_ids --target_id_folder=./preprocess_Result/01_source_eos_mask --result_dir=./result_Files &> acc.log
}
if [ $need_preprocess == "y" ]; then
preprocess_data
if [ $? -ne 0 ]; then
echo "preprocess dataset failed"
exit 1
fi
fi
compile_app
if [ $? -ne 0 ]; then
echo "compile app code failed"
exit 1
fi
infer
if [ $? -ne 0 ]; then
echo " execute inference failed"
exit 1
fi
cal_acc
if [ $? -ne 0 ]; then
echo "calculate accuracy failed"
exit 1
fi

View File

@ -13,6 +13,10 @@
- [Dataset Preparation](#dataset-preparation)
- [Training Process](#training-process)
- [Evaluation Process](#evaluation-process)
- [Inference Process](#inference-process)
- [Export MindIR](#export-mindir)
- [Infer on Ascend310](#infer-on-ascend310)
- [result](#result)
- [Model Description](#model-description)
- [Performance](#performance)
- [Training Performance](#training-performance)
@ -232,6 +236,33 @@ Parameters for learning rate:
perl multi-bleu.perl REF_DATA.forbleu < EVAL_OUTPUT.forbleu
```
## Inference Process
### [Export MindIR](#contents)
```shell
python export.py --ckpt_file [CKPT_PATH] --file_name [FILE_NAME] --file_format [FILE_FORMAT]
```
The ckpt_file parameter is required,
`EXPORT_FORMAT` should be in ["AIR", "MINDIR"]
### Infer on Ascend310
Before performing inference, the mindir file must be exported by `export.py` script. We only provide an example of inference using MINDIR model.
```shell
# Ascend310 inference
bash run_infer_310.sh [MINDIR_PATH] [NEED_PREPROCESS] [DEVICE_ID]
```
- `NEED_PREPROCESS` means weather need preprocess or not, it's value is 'y' or 'n'.
- `DEVICE_ID` is optional, default value is 0.
### result
Inference result is saved in current path, 'output_file' will generate in path specified, For details about how to get BLEU score, see [Evaluation Process](#evaluation-process).
## [Model Description](#contents)
### [Performance](#contents)

View File

@ -19,6 +19,10 @@
- [准备数据集](#准备数据集)
- [训练过程](#训练过程)
- [评估过程](#评估过程)
- [推理过程](#推理过程)
- [导出MindIR](#导出mindir)
- [在Ascend310执行推理](#在ascend310执行推理)
- [结果](#结果)
- [模型描述](#模型描述)
- [性能](#性能)
- [训练性能](#训练性能)
@ -239,6 +243,33 @@ Parameters for learning rate:
perl multi-bleu.perl REF_DATA.forbleu < EVAL_OUTPUT.forbleu
```
## 推理过程
### [导出MindIR](#contents)
```shell
python export.py --ckpt_file [CKPT_PATH] --file_name [FILE_NAME] --file_format [FILE_FORMAT]
```
参数ckpt_file为必填项
`EXPORT_FORMAT` 必须在 ["AIR", "MINDIR"]中选择。
### 在Ascend310执行推理
在执行推理前mindir文件必须通过`export.py`脚本导出。以下展示了使用minir模型执行推理的示例。
```shell
# Ascend310 推理
bash run_infer_310.sh [MINDIR_PATH] [NEED_PREPROCESS] [DEVICE_ID]
```
- `NEED_PREPROCESS` 表示是否需要对数据集进行预处理, 取值为'y' 或者 'n'。
- `DEVICE_ID` 可选默认值为0。
### 结果
推理结果保存在脚本执行的当前路径,'output_file' 将会生成在指定路径生成BLEU分数的过程请参照[评估过程](#评估过程).
## 模型描述
### 性能

View File

@ -0,0 +1,14 @@
cmake_minimum_required(VERSION 3.14.1)
project(Ascend310Infer)
add_compile_definitions(_GLIBCXX_USE_CXX11_ABI=0)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O0 -g -std=c++17 -Werror -Wall -fPIE -Wl,--allow-shlib-undefined")
set(PROJECT_SRC_ROOT ${CMAKE_CURRENT_LIST_DIR}/)
option(MINDSPORE_PATH "mindspore install path" "")
include_directories(${MINDSPORE_PATH})
include_directories(${MINDSPORE_PATH}/include)
include_directories(${PROJECT_SRC_ROOT})
find_library(MS_LIB libmindspore.so ${MINDSPORE_PATH}/lib)
file(GLOB_RECURSE MD_LIB ${MINDSPORE_PATH}/_c_dataengine*)
add_executable(main src/main.cc src/utils.cc)
target_link_libraries(main ${MS_LIB} ${MD_LIB} gflags)

View File

@ -0,0 +1,29 @@
#!/bin/bash
# Copyright 2021 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
if [ -d out ]; then
rm -rf out
fi
mkdir out
cd out || exit
if [ -f "Makefile" ]; then
make clean
fi
cmake .. \
-DMINDSPORE_PATH="`pip3.7 show mindspore-ascend | grep Location | awk '{print $2"/mindspore"}' | xargs realpath`"
make

View File

@ -0,0 +1,32 @@
/**
* Copyright 2021 Huawei Technologies Co., Ltd
*
* 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.
*/
#ifndef MINDSPORE_INFERENCE_UTILS_H_
#define MINDSPORE_INFERENCE_UTILS_H_
#include <sys/stat.h>
#include <dirent.h>
#include <vector>
#include <string>
#include <memory>
#include "include/api/types.h"
std::vector<std::string> GetAllFiles(std::string_view dirName);
DIR *OpenDir(std::string_view dirName);
std::string RealPath(std::string_view path);
mindspore::MSTensor ReadFileToTensor(const std::string &file);
int WriteResult(const std::string& imageFile, const std::vector<mindspore::MSTensor> &outputs);
#endif

View File

@ -0,0 +1,137 @@
/**
* Copyright 2021 Huawei Technologies Co., Ltd
*
* 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 <sys/time.h>
#include <gflags/gflags.h>
#include <dirent.h>
#include <iostream>
#include <string>
#include <algorithm>
#include <iosfwd>
#include <vector>
#include <fstream>
#include <sstream>
#include "include/api/model.h"
#include "include/api/context.h"
#include "include/api/types.h"
#include "include/api/serialization.h"
#include "include/dataset/execute.h"
#include "include/dataset/vision.h"
#include "inc/utils.h"
using mindspore::Context;
using mindspore::Serialization;
using mindspore::Model;
using mindspore::Status;
using mindspore::MSTensor;
using mindspore::dataset::Execute;
using mindspore::ModelType;
using mindspore::GraphCell;
using mindspore::kSuccess;
DEFINE_string(mindir_path, "", "mindir path");
DEFINE_string(input0_path, ".", "input0 path");
DEFINE_string(input1_path, ".", "input1 path");
DEFINE_int32(device_id, 0, "device id");
int main(int argc, char **argv) {
gflags::ParseCommandLineFlags(&argc, &argv, true);
if (RealPath(FLAGS_mindir_path).empty()) {
std::cout << "Invalid mindir" << std::endl;
return 1;
}
auto context = std::make_shared<Context>();
auto ascend310 = std::make_shared<mindspore::Ascend310DeviceInfo>();
ascend310->SetDeviceID(FLAGS_device_id);
ascend310->SetPrecisionMode("allow_fp32_to_fp16");
ascend310->SetOpSelectImplMode("high_precision");
context->MutableDeviceInfo().push_back(ascend310);
mindspore::Graph graph;
Serialization::Load(FLAGS_mindir_path, ModelType::kMindIR, &graph);
Model model;
Status ret = model.Build(GraphCell(graph), context);
if (ret != kSuccess) {
std::cout << "ERROR: Build failed." << std::endl;
return 1;
}
std::vector<MSTensor> model_inputs = model.GetInputs();
if (model_inputs.empty()) {
std::cout << "Invalid model, inputs is empty." << std::endl;
return 1;
}
auto input0_files = GetAllFiles(FLAGS_input0_path);
auto input1_files = GetAllFiles(FLAGS_input1_path);
if (input0_files.empty() || input1_files.empty()) {
std::cout << "ERROR: input data empty." << std::endl;
return 1;
}
std::map<double, double> costTime_map;
size_t size = input0_files.size();
for (size_t i = 0; i < size; ++i) {
struct timeval start = {0};
struct timeval end = {0};
double startTimeMs;
double endTimeMs;
std::vector<MSTensor> inputs;
std::vector<MSTensor> outputs;
std::cout << "Start predict input files:" << input0_files[i] << std::endl;
auto input0 = ReadFileToTensor(input0_files[i]);
auto input1 = ReadFileToTensor(input1_files[i]);
inputs.emplace_back(model_inputs[0].Name(), model_inputs[0].DataType(), model_inputs[0].Shape(),
input0.Data().get(), input0.DataSize());
inputs.emplace_back(model_inputs[1].Name(), model_inputs[1].DataType(), model_inputs[1].Shape(),
input1.Data().get(), input1.DataSize());
gettimeofday(&start, nullptr);
ret = model.Predict(inputs, &outputs);
gettimeofday(&end, nullptr);
if (ret != kSuccess) {
std::cout << "Predict " << input0_files[i] << " failed." << std::endl;
return 1;
}
startTimeMs = (1.0 * start.tv_sec * 1000000 + start.tv_usec) / 1000;
endTimeMs = (1.0 * end.tv_sec * 1000000 + end.tv_usec) / 1000;
costTime_map.insert(std::pair<double, double>(startTimeMs, endTimeMs));
WriteResult(input0_files[i], outputs);
}
double average = 0.0;
int inferCount = 0;
for (auto iter = costTime_map.begin(); iter != costTime_map.end(); iter++) {
double diff = 0.0;
diff = iter->second - iter->first;
average += diff;
inferCount++;
}
average = average / inferCount;
std::stringstream timeCost;
timeCost << "NN inference cost average time: "<< average << " ms of infer_count " << inferCount << std::endl;
std::cout << "NN inference cost average time: "<< average << "ms of infer_count " << inferCount << std::endl;
std::string fileName = "./time_Result" + std::string("/test_perform_static.txt");
std::ofstream fileStream(fileName.c_str(), std::ios::trunc);
fileStream << timeCost.str();
fileStream.close();
costTime_map.clear();
return 0;
}

View File

@ -0,0 +1,129 @@
/**
* Copyright 2021 Huawei Technologies Co., Ltd
*
* 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 <fstream>
#include <algorithm>
#include <iostream>
#include "inc/utils.h"
using mindspore::MSTensor;
using mindspore::DataType;
std::vector<std::string> GetAllFiles(std::string_view dirName) {
struct dirent *filename;
DIR *dir = OpenDir(dirName);
if (dir == nullptr) {
return {};
}
std::vector<std::string> res;
while ((filename = readdir(dir)) != nullptr) {
std::string dName = std::string(filename->d_name);
if (dName == "." || dName == ".." || filename->d_type != DT_REG) {
continue;
}
res.emplace_back(std::string(dirName) + "/" + filename->d_name);
}
std::sort(res.begin(), res.end());
for (auto &f : res) {
std::cout << "image file: " << f << std::endl;
}
return res;
}
int WriteResult(const std::string& imageFile, const std::vector<MSTensor> &outputs) {
std::string homePath = "./result_Files";
for (size_t i = 0; i < outputs.size(); ++i) {
size_t outputSize;
std::shared_ptr<const void> netOutput;
netOutput = outputs[i].Data();
outputSize = outputs[i].DataSize();
int pos = imageFile.rfind('/');
std::string fileName(imageFile, pos + 1);
fileName.replace(fileName.find('.'), fileName.size() - fileName.find('.'), '_' + std::to_string(i) + ".bin");
std::string outFileName = homePath + "/" + fileName;
FILE * outputFile = fopen(outFileName.c_str(), "wb");
fwrite(netOutput.get(), outputSize, sizeof(char), outputFile);
fclose(outputFile);
outputFile = nullptr;
}
return 0;
}
mindspore::MSTensor ReadFileToTensor(const std::string &file) {
if (file.empty()) {
std::cout << "Pointer file is nullptr" << std::endl;
return mindspore::MSTensor();
}
std::ifstream ifs(file);
if (!ifs.good()) {
std::cout << "File: " << file << " is not exist" << std::endl;
return mindspore::MSTensor();
}
if (!ifs.is_open()) {
std::cout << "File: " << file << "open failed" << std::endl;
return mindspore::MSTensor();
}
ifs.seekg(0, std::ios::end);
size_t size = ifs.tellg();
mindspore::MSTensor buffer(file, mindspore::DataType::kNumberTypeUInt8, {static_cast<int64_t>(size)}, nullptr, size);
ifs.seekg(0, std::ios::beg);
ifs.read(reinterpret_cast<char *>(buffer.MutableData()), size);
ifs.close();
return buffer;
}
DIR *OpenDir(std::string_view dirName) {
if (dirName.empty()) {
std::cout << " dirName is null ! " << std::endl;
return nullptr;
}
std::string realPath = RealPath(dirName);
struct stat s;
lstat(realPath.c_str(), &s);
if (!S_ISDIR(s.st_mode)) {
std::cout << "dirName is not a valid directory !" << std::endl;
return nullptr;
}
DIR *dir;
dir = opendir(realPath.c_str());
if (dir == nullptr) {
std::cout << "Can not open dir " << dirName << std::endl;
return nullptr;
}
std::cout << "Successfully opened the dir " << dirName << std::endl;
return dir;
}
std::string RealPath(std::string_view path) {
char realPathMem[PATH_MAX] = {0};
char *realPathRet = nullptr;
realPathRet = realpath(path.data(), realPathMem);
if (realPathRet == nullptr) {
std::cout << "File: " << path << " is not exist.";
return "";
}
std::string realPath(realPathMem);
std::cout << path << " realpath is: " << realPath << std::endl;
return realPath;
}

View File

@ -0,0 +1,51 @@
# Copyright 2021 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
"""Transformer evaluation script."""
import os
import argparse
import numpy as np
from src.eval_config import cfg, transformer_net_cfg
parser = argparse.ArgumentParser(description='postprocess')
parser.add_argument("--result_dir", type=str, default="./result_Files",
help="infer result path.")
args = parser.parse_args()
def generate_output():
'''
Generate output.
'''
predictions = []
file_num = len(os.listdir(args.result_dir))
for i in range(file_num):
batch = "transformer_bs_" + str(transformer_net_cfg.batch_size) + "_" + str(i) + "_0.bin"
pred = np.fromfile(os.path.join(args.result_dir, batch), np.int32)
predictions.append(pred.reshape(1, 1, transformer_net_cfg.max_decode_length + 1))
# decode and write to file
f = open(cfg.output_file, 'w')
for batch_out in predictions:
for i in range(transformer_net_cfg.batch_size):
if batch_out.ndim == 3:
batch_out = batch_out[:, 0]
token_ids = [str(x) for x in batch_out[i].tolist()]
f.write(" ".join(token_ids) + "\n")
f.close()
if __name__ == "__main__":
generate_output()

View File

@ -0,0 +1,58 @@
# Copyright 2021 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
"""Transformer evaluation script."""
import os
import argparse
from src.eval_config import cfg, transformer_net_cfg
from eval import load_test_data
parser = argparse.ArgumentParser(description='preprocess')
parser.add_argument("--result_path", type=str, default="./preprocess_Result/",
help="preprocess result path.")
args = parser.parse_args()
def generate_bin():
'''
Generate bin files.
'''
dataset = load_test_data(batch_size=transformer_net_cfg.batch_size, data_file=cfg.data_file)
cur_dir = args.result_path
source_eos_ids_path = os.path.join(cur_dir, "00_source_eos_ids")
source_eos_mask_path = os.path.join(cur_dir, "01_source_eos_mask")
if not os.path.isdir(source_eos_ids_path):
os.makedirs(source_eos_ids_path)
if not os.path.isdir(source_eos_mask_path):
os.makedirs(source_eos_mask_path)
batch_size = transformer_net_cfg.batch_size
for i, data in enumerate(dataset.create_dict_iterator(output_numpy=True, num_epochs=1)):
file_name = "transformer_bs_" + str(batch_size) + "_" + str(i) + ".bin"
source_eos_ids = data['source_eos_ids']
source_eos_ids.tofile(os.path.join(source_eos_ids_path, file_name))
source_eos_mask = data['source_eos_mask']
source_eos_mask.tofile(os.path.join(source_eos_mask_path, file_name))
print("="*20, "export bin files finished", "="*20)
if __name__ == "__main__":
generate_bin()

View File

@ -0,0 +1,120 @@
#!/bin/bash
# Copyright 2021 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
if [[ $# -lt 2 || $# -gt 3 ]]; then
echo "Usage: bash run_infer_310.sh [MINDIR_PATH] [NEED_PREPROCESS] [DEVICE_ID]
NEED_PREPROCESS means weather need preprocess or not, it's value is 'y' or 'n'.
DEVICE_ID is optional, it can be set by environment variable device_id, otherwise the value is zero"
exit 1
fi
get_real_path(){
if [ "${1:0:1}" == "/" ]; then
echo "$1"
else
echo "$(realpath -m $PWD/$1)"
fi
}
model=$(get_real_path $1)
if [ "$2" == "y" ] || [ "$2" == "n" ];then
need_preprocess=$2
else
echo "weather need preprocess or not, it's value must be in [y, n]"
exit 1
fi
device_id=0
if [ $# == 3 ]; then
device_id=$3
fi
echo "mindir name: "$model
echo "need preprocess: "$need_preprocess
echo "device id: "$device_id
export ASCEND_HOME=/usr/local/Ascend/
if [ -d ${ASCEND_HOME}/ascend-toolkit ]; then
export PATH=$ASCEND_HOME/ascend-toolkit/latest/fwkacllib/ccec_compiler/bin:$ASCEND_HOME/ascend-toolkit/latest/atc/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/lib:$ASCEND_HOME/ascend-toolkit/latest/atc/lib64:$ASCEND_HOME/ascend-toolkit/latest/fwkacllib/lib64:$ASCEND_HOME/driver/lib64:$ASCEND_HOME/add-ons:$LD_LIBRARY_PATH
export TBE_IMPL_PATH=$ASCEND_HOME/ascend-toolkit/latest/opp/op_impl/built-in/ai_core/tbe
export PYTHONPATH=${TBE_IMPL_PATH}:$ASCEND_HOME/ascend-toolkit/latest/fwkacllib/python/site-packages:$PYTHONPATH
export ASCEND_OPP_PATH=$ASCEND_HOME/ascend-toolkit/latest/opp
else
export PATH=$ASCEND_HOME/atc/ccec_compiler/bin:$ASCEND_HOME/atc/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/lib:$ASCEND_HOME/atc/lib64:$ASCEND_HOME/acllib/lib64:$ASCEND_HOME/driver/lib64:$ASCEND_HOME/add-ons:$LD_LIBRARY_PATH
export PYTHONPATH=$ASCEND_HOME/atc/python/site-packages:$PYTHONPATH
export ASCEND_OPP_PATH=$ASCEND_HOME/opp
fi
function preprocess_data()
{
if [ -d preprocess_Result ]; then
rm -rf ./preprocess_Result
fi
mkdir preprocess_Result
python3.7 ../preprocess.py --result_path=./preprocess_Result/
}
function compile_app()
{
cd ../ascend310_infer || exit
bash build.sh &> build.log
}
function infer()
{
cd - || exit
if [ -d result_Files ]; then
rm -rf ./result_Files
fi
if [ -d time_Result ]; then
rm -rf ./time_Result
fi
mkdir result_Files
mkdir time_Result
../ascend310_infer/out/main --mindir_path=$model --input0_path=./preprocess_Result/00_source_eos_ids --input1_path=./preprocess_Result/01_source_eos_mask --device_id=$device_id &> infer.log
}
function cal_acc()
{
python3.7 ../postprocess.py --result_dir=./result_Files &> acc.log
}
if [ $need_preprocess == "y" ]; then
preprocess_data
if [ $? -ne 0 ]; then
echo "preprocess dataset failed"
exit 1
fi
fi
compile_app
if [ $? -ne 0 ]; then
echo "compile app code failed"
exit 1
fi
infer
if [ $? -ne 0 ]; then
echo " execute inference failed"
exit 1
fi
cal_acc
if [ $? -ne 0 ]; then
echo "calculate accuracy failed"
exit 1
fi

View File

@ -161,6 +161,34 @@ config.py
The above python command will run in the background, you can view the results through the file output.eval.log. You will get the accuracy as following:
## Inference Process
### [Export MindIR](#contents)
```shell
python export.py --ckpt_file [CKPT_PATH] --file_name [FILE_NAME] --file_format [FILE_FORMAT]
```
The ckpt_file parameter is required,
`EXPORT_FORMAT` should be in ["AIR", "MINDIR"]
### Infer on Ascend310
Before performing inference, the mindir file must be exported by `export.py` script. We only provide an example of inference using MINDIR model.
Current batch_size can only be set to 1.
```shell
# Ascend310 inference
bash run_infer_310.sh [MINDIR_PATH] [NEED_PREPROCESS] [DEVICE_ID]
```
`NEED_PREPROCESS` means weather need preprocess or not, it's value is 'y' or 'n'.
`DEVICE_ID` is optional, default value is 0.
### result
Inference result is saved in current path, you can find result in acc.log file.
## performance
### Training performance

View File

@ -0,0 +1,14 @@
cmake_minimum_required(VERSION 3.14.1)
project(Ascend310Infer)
add_compile_definitions(_GLIBCXX_USE_CXX11_ABI=0)
set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -O0 -g -std=c++17 -Werror -Wall -fPIE -Wl,--allow-shlib-undefined")
set(PROJECT_SRC_ROOT ${CMAKE_CURRENT_LIST_DIR}/)
option(MINDSPORE_PATH "mindspore install path" "")
include_directories(${MINDSPORE_PATH})
include_directories(${MINDSPORE_PATH}/include)
include_directories(${PROJECT_SRC_ROOT})
find_library(MS_LIB libmindspore.so ${MINDSPORE_PATH}/lib)
file(GLOB_RECURSE MD_LIB ${MINDSPORE_PATH}/_c_dataengine*)
add_executable(main src/main.cc src/utils.cc)
target_link_libraries(main ${MS_LIB} ${MD_LIB} gflags)

View File

@ -0,0 +1,29 @@
#!/bin/bash
# Copyright 2021 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
if [ -d out ]; then
rm -rf out
fi
mkdir out
cd out || exit
if [ -f "Makefile" ]; then
make clean
fi
cmake .. \
-DMINDSPORE_PATH="`pip3.7 show mindspore-ascend | grep Location | awk '{print $2"/mindspore"}' | xargs realpath`"
make

View File

@ -0,0 +1,35 @@
/**
* Copyright 2021 Huawei Technologies Co., Ltd
*
* 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.
*/
#ifndef MINDSPORE_INFERENCE_UTILS_H_
#define MINDSPORE_INFERENCE_UTILS_H_
#include <sys/stat.h>
#include <dirent.h>
#include <vector>
#include <string>
#include <memory>
#include "include/api/types.h"
std::vector<std::string> GetAllFiles(std::string_view dirName);
DIR *OpenDir(std::string_view dirName);
std::string RealPath(std::string_view path);
mindspore::MSTensor ReadFileToTensor(const std::string &file);
int WriteResult(const std::string& imageFile, const std::vector<mindspore::MSTensor> &outputs);
std::vector<std::string> GetAllFiles(std::string dir_name);
std::vector<std::vector<std::string>> GetAllInputData(std::string dir_name);
#endif

View File

@ -0,0 +1,144 @@
/**
* Copyright 2021 Huawei Technologies Co., Ltd
*
* 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 <sys/time.h>
#include <gflags/gflags.h>
#include <dirent.h>
#include <iostream>
#include <string>
#include <algorithm>
#include <iosfwd>
#include <vector>
#include <fstream>
#include <sstream>
#include "include/api/model.h"
#include "include/api/serialization.h"
#include "include/api/context.h"
#include "include/dataset/execute.h"
#include "include/dataset/vision.h"
#include "include/dataset/config.h"
#include "../inc/utils.h"
#include "include/api/types.h"
using mindspore::Context;
using mindspore::Serialization;
using mindspore::Model;
using mindspore::Status;
using mindspore::dataset::Execute;
using mindspore::MSTensor;
using mindspore::ModelType;
using mindspore::GraphCell;
using mindspore::kSuccess;
using mindspore::dataset::vision::Decode;
using mindspore::dataset::vision::Normalize;
using mindspore::dataset::vision::Resize;
using mindspore::dataset::vision::HWC2CHW;
using mindspore::dataset::InterpolationMode;
DEFINE_string(mindir_path, "", "mindir path");
DEFINE_string(dataset_path, ".", "dataset path");
DEFINE_int32(device_id, 0, "device id");
int main(int argc, char **argv) {
gflags::ParseCommandLineFlags(&argc, &argv, true);
if (RealPath(FLAGS_mindir_path).empty()) {
std::cout << "Invalid mindir" << std::endl;
return 1;
}
auto context = std::make_shared<Context>();
auto ascend310 = std::make_shared<mindspore::Ascend310DeviceInfo>();
ascend310->SetDeviceID(FLAGS_device_id);
ascend310->SetPrecisionMode("allow_fp32_to_fp16");
context->MutableDeviceInfo().push_back(ascend310);
mindspore::Graph graph;
Serialization::Load(FLAGS_mindir_path, ModelType::kMindIR, &graph);
Model model;
Status ret = model.Build(GraphCell(graph), context);
if (ret != kSuccess) {
std::cout << "ERROR: Build failed." << std::endl;
return 1;
}
std::vector<MSTensor> model_inputs = model.GetInputs();
auto all_files = GetAllFiles(FLAGS_dataset_path);
if (all_files.empty()) {
std::cout << "ERROR: no input data." << std::endl;
return 1;
}
std::map<double, double> costTime_map;
size_t size = all_files.size();
auto decode(new Decode());
auto resize(new Resize({448, 448}, InterpolationMode::kNearestNeighbour));
auto normalize(new Normalize({123.68, 116.779, 103.939}, {1.0, 1.0, 1.0}));
auto hwc2chw(new HWC2CHW());
Execute preprocess({decode, resize, normalize, hwc2chw});
for (size_t i = 0; i < size; ++i) {
struct timeval start = {0};
struct timeval end = {0};
double startTimeMs;
double endTimeMs;
std::vector<MSTensor> inputs;
std::vector<MSTensor> outputs;
std::cout << "Start predict input files:" << all_files[i] <<std::endl;
auto img = MSTensor();
ret = preprocess(ReadFileToTensor(all_files[i]), &img);
if (ret != kSuccess) {
std::cout << "preprocess " << all_files[i] << " failed." << std::endl;
return 1;
}
inputs.emplace_back(model_inputs[0].Name(), model_inputs[0].DataType(), model_inputs[0].Shape(),
img.Data().get(), img.DataSize());
gettimeofday(&start, nullptr);
ret = model.Predict(inputs, &outputs);
gettimeofday(&end, nullptr);
if (ret != kSuccess) {
std::cout << "Predict " << all_files[i] << " failed." << std::endl;
return 1;
}
startTimeMs = (1.0 * start.tv_sec * 1000000 + start.tv_usec) / 1000;
endTimeMs = (1.0 * end.tv_sec * 1000000 + end.tv_usec) / 1000;
costTime_map.insert(std::pair<double, double>(startTimeMs, endTimeMs));
WriteResult(all_files[i], outputs);
}
double average = 0.0;
int inferCount = 0;
for (auto iter = costTime_map.begin(); iter != costTime_map.end(); iter++) {
double diff = 0.0;
diff = iter->second - iter->first;
average += diff;
inferCount++;
}
average = average / inferCount;
std::stringstream timeCost;
timeCost << "NN inference cost average time: "<< average << " ms of infer_count " << inferCount << std::endl;
std::cout << "NN inference cost average time: "<< average << "ms of infer_count " << inferCount << std::endl;
std::string fileName = "./time_Result" + std::string("/test_perform_static.txt");
std::ofstream fileStream(fileName.c_str(), std::ios::trunc);
fileStream << timeCost.str();
fileStream.close();
costTime_map.clear();
return 0;
}

View File

@ -0,0 +1,185 @@
/**
* Copyright 2021 Huawei Technologies Co., Ltd
*
* 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 <fstream>
#include <algorithm>
#include <iostream>
#include "inc/utils.h"
using mindspore::MSTensor;
using mindspore::DataType;
std::vector<std::vector<std::string>> GetAllInputData(std::string dir_name) {
std::vector<std::vector<std::string>> ret;
DIR *dir = OpenDir(dir_name);
if (dir == nullptr) {
return {};
}
struct dirent *filename;
/* read all the files in the dir ~ */
std::vector<std::string> sub_dirs;
while ((filename = readdir(dir)) != nullptr) {
std::string d_name = std::string(filename->d_name);
// get rid of "." and ".."
if (d_name == "." || d_name == ".." || d_name.empty()) {
continue;
}
std::string dir_path = RealPath(std::string(dir_name) + "/" + filename->d_name);
struct stat s;
lstat(dir_path.c_str(), &s);
if (!S_ISDIR(s.st_mode)) {
continue;
}
sub_dirs.emplace_back(dir_path);
}
std::sort(sub_dirs.begin(), sub_dirs.end());
(void)std::transform(sub_dirs.begin(), sub_dirs.end(), std::back_inserter(ret),
[](const std::string &d) { return GetAllFiles(d); });
return ret;
}
std::vector<std::string> GetAllFiles(std::string dir_name) {
struct dirent *filename;
DIR *dir = OpenDir(dir_name);
if (dir == nullptr) {
return {};
}
std::vector<std::string> res;
while ((filename = readdir(dir)) != nullptr) {
std::string d_name = std::string(filename->d_name);
if (d_name == "." || d_name == ".." || d_name.size() <= 3) {
continue;
}
res.emplace_back(std::string(dir_name) + "/" + filename->d_name);
}
std::sort(res.begin(), res.end());
return res;
}
std::vector<std::string> GetAllFiles(std::string_view dirName) {
struct dirent *filename;
DIR *dir = OpenDir(dirName);
if (dir == nullptr) {
return {};
}
std::vector<std::string> res;
while ((filename = readdir(dir)) != nullptr) {
std::string dName = std::string(filename->d_name);
if (dName == "." || dName == ".." || filename->d_type != DT_REG) {
continue;
}
res.emplace_back(std::string(dirName) + "/" + filename->d_name);
}
std::sort(res.begin(), res.end());
for (auto &f : res) {
std::cout << "image file: " << f << std::endl;
}
return res;
}
int WriteResult(const std::string& imageFile, const std::vector<MSTensor> &outputs) {
std::string homePath = "./result_Files";
for (size_t i = 0; i < outputs.size(); ++i) {
size_t outputSize;
std::shared_ptr<const void> netOutput;
netOutput = outputs[i].Data();
outputSize = outputs[i].DataSize();
int pos = imageFile.rfind('/');
std::string fileName(imageFile, pos + 1);
fileName = fileName + '_' + std::to_string(i) + ".bin";
std::string outFileName = homePath + "/" + fileName;
FILE *outputFile = fopen(outFileName.c_str(), "wb");
fwrite(netOutput.get(), outputSize, sizeof(char), outputFile);
fclose(outputFile);
outputFile = nullptr;
}
return 0;
}
mindspore::MSTensor ReadFileToTensor(const std::string &file) {
if (file.empty()) {
std::cout << "Pointer file is nullptr" << std::endl;
return mindspore::MSTensor();
}
std::ifstream ifs(file);
if (!ifs.good()) {
std::cout << "File: " << file << " is not exist" << std::endl;
return mindspore::MSTensor();
}
if (!ifs.is_open()) {
std::cout << "File: " << file << "open failed" << std::endl;
return mindspore::MSTensor();
}
ifs.seekg(0, std::ios::end);
size_t size = ifs.tellg();
mindspore::MSTensor buffer(file, mindspore::DataType::kNumberTypeUInt8, {static_cast<int64_t>(size)}, nullptr, size);
ifs.seekg(0, std::ios::beg);
ifs.read(reinterpret_cast<char *>(buffer.MutableData()), size);
ifs.close();
return buffer;
}
DIR *OpenDir(std::string_view dirName) {
if (dirName.empty()) {
std::cout << " dirName is null ! " << std::endl;
return nullptr;
}
std::string realPath = RealPath(dirName);
struct stat s;
lstat(realPath.c_str(), &s);
if (!S_ISDIR(s.st_mode)) {
std::cout << "dirName is not a valid directory !" << std::endl;
return nullptr;
}
DIR *dir;
dir = opendir(realPath.c_str());
if (dir == nullptr) {
std::cout << "Can not open dir " << dirName << std::endl;
return nullptr;
}
std::cout << "Successfully opened the dir " << dirName << std::endl;
return dir;
}
std::string RealPath(std::string_view path) {
char realPathMem[PATH_MAX] = {0};
char *realPathRet = nullptr;
realPathRet = realpath(path.data(), realPathMem);
if (realPathRet == nullptr) {
std::cout << "File: " << path << " is not exist.";
return "";
}
std::string realPath(realPathMem);
std::cout << path << " realpath is: " << realPath << std::endl;
return realPath;
}

View File

@ -0,0 +1,60 @@
# Copyright 2021 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
"""
#################postprocess########################
"""
import os
import argparse
import numpy as np
from mindspore import Tensor
from src.config import config as cfg
from src.score import eval_pre_rec_f1
def parse_args(cloud_args=None):
"""parameters"""
parser = argparse.ArgumentParser('postprocess')
parser.add_argument('--rst_path', type=str, default='./result_Files/',
help='infer result path.')
args_opt = parser.parse_args()
args_opt.data_dir = cfg.data_dir
args_opt.train_image_dir_name = os.path.join(cfg.data_dir, cfg.train_image_dir_name)
args_opt.val_fname = cfg.val_fname
args_opt.train_label_dir_name = os.path.join(cfg.data_dir, cfg.train_label_dir_name)
args_opt.batch_size = 1
return args_opt
if __name__ == '__main__':
arg = parse_args()
obj = eval_pre_rec_f1()
with open(os.path.join(arg.data_dir, arg.val_fname), 'r') as f_val:
f_list = f_val.readlines()
batch_list = np.arange(0, len(f_list), arg.batch_size)
for idx in batch_list:
gt_list = []
for i in range(idx, min(idx + arg.batch_size, len(f_list))):
item = f_list[i]
img_filename = str(item).strip().split(',')[0]
gt_list.append(np.load(os.path.join(arg.train_label_dir_name, img_filename[:-4]) + '.npy'))
y = np.fromfile(os.path.join(arg.rst_path, img_filename + '_0.bin'), np.float32)
y = Tensor(y.reshape(1, 7, 112, 112))
obj.add(y, gt_list)
print(obj.val())

View File

@ -0,0 +1,50 @@
# Copyright 2021 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
"""
################preprocess########################
"""
import argparse
import os
from PIL import Image
from src.config import config as cfg
def parse_args(cloud_args=None):
"""parameters"""
parser = argparse.ArgumentParser('preprocess')
parser.add_argument('--result_path', type=str, default='./preprocess_Result/',
help='result path')
args_opt = parser.parse_args()
args_opt.data_dir = cfg.data_dir
args_opt.train_image_dir_name = os.path.join(cfg.data_dir, cfg.train_image_dir_name)
args_opt.val_fname = cfg.val_fname
return args_opt
def prepare_valset(arg):
"""generate validate dataset."""
with open(os.path.join(arg.data_dir, arg.val_fname), 'r') as f_val:
f_list = f_val.readlines()
for i, _ in enumerate(f_list):
item = f_list[i]
img_filename = str(item).strip().split(',')[0]
img_path = os.path.join(arg.train_image_dir_name, img_filename)
img = Image.open(img_path)
img.save(os.path.join(arg.result_path, img_filename))
if __name__ == '__main__':
args = parse_args()
prepare_valset(args)

View File

@ -0,0 +1,120 @@
#!/bin/bash
# Copyright 2021 Huawei Technologies Co., Ltd
#
# 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.
# ============================================================================
if [[ $# -lt 2 || $# -gt 3 ]]; then
echo "Usage: bash run_infer_310.sh [MINDIR_PATH] [NEED_PREPROCESS] [DEVICE_ID]
NEED_PREPROCESS means weather need preprocess or not, it's value is 'y' or 'n'.
DEVICE_ID is optional, it can be set by environment variable device_id, otherwise the value is zero"
exit 1
fi
get_real_path(){
if [ "${1:0:1}" == "/" ]; then
echo "$1"
else
echo "$(realpath -m $PWD/$1)"
fi
}
model=$(get_real_path $1)
if [ "$2" == "y" ] || [ "$2" == "n" ];then
need_preprocess=$2
else
echo "weather need preprocess or not, it's value must be in [y, n]"
exit 1
fi
device_id=0
if [ $# == 3 ]; then
device_id=$3
fi
echo "mindir name: "$model
echo "need preprocess: "$need_preprocess
echo "device id: "$device_id
export ASCEND_HOME=/usr/local/Ascend/
if [ -d ${ASCEND_HOME}/ascend-toolkit ]; then
export PATH=$ASCEND_HOME/ascend-toolkit/latest/fwkacllib/ccec_compiler/bin:$ASCEND_HOME/ascend-toolkit/latest/atc/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/lib:$ASCEND_HOME/ascend-toolkit/latest/atc/lib64:$ASCEND_HOME/ascend-toolkit/latest/fwkacllib/lib64:$ASCEND_HOME/driver/lib64:$ASCEND_HOME/add-ons:$LD_LIBRARY_PATH
export TBE_IMPL_PATH=$ASCEND_HOME/ascend-toolkit/latest/opp/op_impl/built-in/ai_core/tbe
export PYTHONPATH=${TBE_IMPL_PATH}:$ASCEND_HOME/ascend-toolkit/latest/fwkacllib/python/site-packages:$PYTHONPATH
export ASCEND_OPP_PATH=$ASCEND_HOME/ascend-toolkit/latest/opp
else
export PATH=$ASCEND_HOME/atc/ccec_compiler/bin:$ASCEND_HOME/atc/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/lib:$ASCEND_HOME/atc/lib64:$ASCEND_HOME/acllib/lib64:$ASCEND_HOME/driver/lib64:$ASCEND_HOME/add-ons:$LD_LIBRARY_PATH
export PYTHONPATH=$ASCEND_HOME/atc/python/site-packages:$PYTHONPATH
export ASCEND_OPP_PATH=$ASCEND_HOME/opp
fi
function preprocess_data()
{
if [ -d preprocess_Result ]; then
rm -rf ./preprocess_Result
fi
mkdir preprocess_Result
python3.7 ../preprocess.py --result_path=./preprocess_Result/
}
function compile_app()
{
cd ../ascend310_infer || exit
bash build.sh &> build.log
}
function infer()
{
cd - || exit
if [ -d result_Files ]; then
rm -rf ./result_Files
fi
if [ -d time_Result ]; then
rm -rf ./time_Result
fi
mkdir result_Files
mkdir time_Result
../ascend310_infer/out/main --mindir_path=$model --dataset_path=./preprocess_Result/ --device_id=$device_id &> infer.log
}
function cal_acc()
{
python3.7 ../postprocess.py --rst_path=./result_Files &> acc.log
}
if [ $need_preprocess == "y" ]; then
preprocess_data
if [ $? -ne 0 ]; then
echo "preprocess dataset failed"
exit 1
fi
fi
compile_app
if [ $? -ne 0 ]; then
echo "compile app code failed"
exit 1
fi
infer
if [ $? -ne 0 ]; then
echo " execute inference failed"
exit 1
fi
cal_acc
if [ $? -ne 0 ]; then
echo "calculate accuracy failed"
exit 1
fi