【benchmark】remove static benchmark model (#6505)

* fix becnhmark,delete run_process_type

* fix

* fix benchmarkk

* fix mask_rcnn_r50_1x_coco benchmark

* fix benchmark

* fix benchmark

* fix benchmark prepare.sh

* fix benchmark

* remove static benchmark model
Co-authored-by: Nzhengya01 <zhengya01@baidu.com>

test_tipc/static/README.MD

-# PaddleDetection 下静态图benchmark模型执行说明
-静态图benchmark测试脚本说明
-# 目录说明 
-# Docker 运行环境
-docker image: registry.baidubce.com/paddlepaddle/paddle:latest-dev-cuda11.2-cudnn8-gcc82
-paddle = 2.2.2
-python = 3.7
-# 运行benchmark测试步骤
-git clone https://github.com/PaddlePaddle/PaddleDetection.git
-cd PaddleDetection
-# 准备数据
-bash test_tipc/static/${model_item}/benchmark_common/prepare.sh
-# 运行模型
-## 单卡（自动运行打开Profiling）
-export CUDA_VISIBLE_DEVICES=0 
-bash  test_tipc/static/${model_item}/N1C1/yolov3_darknet53_270e_coco_bs8_fp32_SingleP_DP.sh 
-## 多卡
-export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 
-bash test_tipc/static/${model_item}/N1C8/yolov3_darknet53_270e_coco_bs8_fp32_MultiP_DP.sh

test_tipc/static/mask_rcnn_r50_1x_coco/N1C1/mask_rcnn_r50_1x_coco_bs2_fp32_DP.sh

-model_item=mask_rcnn_r50_1x_coco
-bs_item=2
-fp_item=fp32
-run_mode=DP
-device_num=N1C1
-max_iter=100
-num_workers=8
-
-# get data
-bash test_tipc/static/${model_item}/benchmark_common/prepare.sh
-# run
-bash test_tipc/static/${model_item}/benchmark_common/run_benchmark.sh ${model_item} ${bs_item} ${fp_item} ${run_mode} ${device_num} ${max_iter} ${num_workers} 2>&1;
-# run profiling
-sleep 10;
-export PROFILING=true
-bash test_tipc/static/${model_item}/benchmark_common/run_benchmark.sh ${model_item} ${bs_item} ${fp_item} ${run_mode} ${device_num} 11 ${num_workers} 2>&1;

test_tipc/static/mask_rcnn_r50_1x_coco/N1C8/mask_rcnn_r50_1x_coco_bs2_fp32_DP.sh

-model_item=mask_rcnn_r50_1x_coco
-bs_item=2
-fp_item=fp32
-run_mode=DP
-device_num=N1C8
-max_iter=100
-num_workers=8
-
-# get data
-bash test_tipc/static/${model_item}/benchmark_common/prepare.sh
-# run
-bash test_tipc/static/${model_item}/benchmark_common/run_benchmark.sh ${model_item} ${bs_item} ${fp_item} ${run_mode} ${device_num} ${max_iter} ${num_workers} 2>&1;

test_tipc/static/mask_rcnn_r50_1x_coco/benchmark_common/prepare.sh

-#!/usr/bin/env bash
-# 执行路径在模型库的根目录下
-################################# 安装框架 如:
-echo "*******prepare benchmark start ***********"
-pip install -U pip
-echo `pip --version`
-pip install Cython
-pip install -r requirements.txt
-
-################################# 准备训练数据 如:
-wget -nc -P static/dataset/coco/ https://paddledet.bj.bcebos.com/data/coco_benchmark.tar
-cd ./static/dataset/coco/ && tar -xf coco_benchmark.tar && mv -u coco_benchmark/* .
-rm -rf coco_benchmark/ && cd ../../../
-echo "*******prepare benchmark end***********"

test_tipc/static/mask_rcnn_r50_1x_coco/benchmark_common/run_benchmark.sh

-#!/usr/bin/env bash
-# Test training benchmark for a model.
-# Usage：bash run_benchmark.sh ${model_item} ${bs_item} ${fp_item} ${run_mode} ${device_num}
-function _set_params(){
-    model_item=${1:-"model_item"}   # (必选) 模型 item
-    base_batch_size=${2:-"2"}       # (必选) 如果是静态图单进程，则表示每张卡上的BS，需在训练时*卡数
-    fp_item=${3:-"fp32"}            # (必选) fp32|fp16
-    run_mode=${4:-"DP"}             # (必选) MP模型并行|DP数据并行|PP流水线并行|混合并行DP1-MP1-PP1|DP1-MP4-PP1
-    device_num=${5:-"N1C1"}         # (必选) 使用的卡数量，N1C1|N1C8|N4C32 （4机32卡）
-    profiling=${PROFILING:-"false"}      # (必选) Profiling  开关，默认关闭，通过全局变量传递
-    model_repo="PaddleDetection"          # (必选) 模型套件的名字
-    speed_unit="samples/sec"         # (必选)速度指标单位
-    skip_steps=10                  # (必选)解析日志，跳过模型前几个性能不稳定的step
-    keyword="ips:"                 # (必选)解析日志，筛选出性能数据所在行的关键字
-    convergence_key="loss:"        # (可选)解析日志，筛选出收敛数据所在行的关键字 如：convergence_key="loss:"
-    max_iter=${6:-"100"}           # （可选）需保证模型执行时间在5分钟内，需要修改代码提前中断的直接提PR 合入套件；或使用max_epoch参数
-    num_workers=${7:-"8"}                  # (可选)
-#   以下为通用执行命令，无特殊可不用修改
-    model_name=${model_item}_bs${base_batch_size}_${fp_item}_${run_mode}  # (必填) 且格式不要改动,与竞品名称对齐
-    device=${CUDA_VISIBLE_DEVICES//,/ }
-    arr=(${device})
-    num_gpu_devices=${#arr[*]}
-    run_log_path=${TRAIN_LOG_DIR:-$(pwd)}  # （必填） TRAIN_LOG_DIR  benchmark框架设置该参数为全局变量
-    profiling_log_path=${PROFILING_LOG_DIR:-$(pwd)}  # （必填） PROFILING_LOG_DIR benchmark框架设置该参数为全局变量
-    speed_log_path=${LOG_PATH_INDEX_DIR:-$(pwd)}
-
-    train_log_file=${run_log_path}/${model_repo}_${model_name}_${device_num}_log
-    profiling_log_file=${profiling_log_path}/${model_repo}_${model_name}_${device_num}_profiling
-    speed_log_file=${speed_log_path}/${model_repo}_${model_name}_${device_num}_speed
-}
-function _train(){
-    export FLAGS_eager_delete_tensor_gb=0.0
-    export FLAGS_fraction_of_gpu_memory_to_use=0.98
-    export FLAGS_memory_fraction_of_eager_deletion=1.0
-    export FLAGS_conv_workspace_size_limit=500
-    cd ./static
-    batch_size=${base_batch_size}  # 如果模型跑多卡单进程时,请在_train函数中计算出多卡需要的bs
-    echo "current CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES}, model_name=${model_name}, device_num=${device_num}, is profiling=${profiling}"
-
-    if [ ${profiling} = "true" ];then
-        log_file=${profiling_log_file}
-        is_profiler=1
-    else
-        log_file=${train_log_file}
-        is_profiler=0
-    fi
-    if [ ${fp_item} = "fp16" ]; then
-        use_fp16_cmd="--fp16"
-    else
-        use_fp16_cmd=""
-    fi
-
-    train_cmd="-c configs/mask_rcnn_r50_1x.yml  \
-         -o TrainReader.batch_size=${batch_size} \
-         max_iters=${max_iter} log_iter=1 \
-         TrainReader.worker_num=${num_workers} ${use_fp16_cmd} \
-         --is_profiler=${is_profiler} "
-#   以下为通用执行命令，无特殊可不用修改
-    case ${run_mode} in
-    DP) if [[ ${device_num} = "N1C1" ]];then
-            echo "run ${run_mode} ${device_num}"
-            train_cmd="python -u tools/train.py ${train_cmd}"
-        else
-            rm -rf ./mylog
-            train_cmd="python -m paddle.distributed.launch --log_dir=./mylog --gpus=$CUDA_VISIBLE_DEVICES \
-                  tools/train.py ${train_cmd}"
-        fi
-        ;;
-    DP1-MP1-PP1)  echo "run run_mode: DP1-MP1-PP1" ;;
-    *) echo "choose run_mode "; exit 1;
-    esac
-
-    echo "train_cmd: ${train_cmd}  log_file: ${log_file}"
-    timeout 15m ${train_cmd} > ${log_file} 2>&1
-    if [ $? -ne 0 ];then
-        echo -e "${model_name}, FAIL"
-    else
-        echo -e "${model_name}, SUCCESS"
-    fi
-    # kill -9 `ps -ef|grep 'python'|awk '{print $2}'`
-    if [ ${device_num} != "N1C1" -a -d mylog ]; then
-        rm ${log_file}
-        cp mylog/workerlog.0 ${log_file}
-    fi
-    cd ../
-}
-source ${BENCHMARK_ROOT}/scripts/run_model.sh   # 在该脚本中会对符合benchmark规范的log使用analysis.py 脚本进行性能数据解析;如果不联调只想要产出训练log可以注掉本行,提交时需打开
-_set_params $@
-#_train       # 如果只产出训练log,不解析,可取消注释
-_run     # 该函数在run_model.sh中,执行时会调用_train; 如果不联调只产出训练log可以注掉本行,提交时需打开

test_tipc/static/mask_rcnn_r50_fpn_1x_coco/N1C1/mask_rcnn_r50_fpn_1x_coco_bs2_fp32_DP.sh

-model_item=mask_rcnn_r50_fpn_1x_coco
-bs_item=2
-fp_item=fp32
-run_mode=DP
-device_num=N1C1
-max_iter=100
-num_workers=2
-
-# get data
-bash test_tipc/static/${model_item}/benchmark_common/prepare.sh
-# run
-bash test_tipc/static/${model_item}/benchmark_common/run_benchmark.sh ${model_item} ${bs_item} ${fp_item} ${run_mode} ${device_num} ${max_iter} ${num_workers} 2>&1;
-# run profiling
-sleep 10;
-export PROFILING=true
-bash test_tipc/static/${model_item}/benchmark_common/run_benchmark.sh ${model_item} ${bs_item} ${fp_item} ${run_mode} ${device_num} 11 ${num_workers} 2>&1;

test_tipc/static/mask_rcnn_r50_fpn_1x_coco/N1C8/mask_rcnn_r50_fpn_1x_coco_bs2_fp32_DP.sh

-model_item=mask_rcnn_r50_fpn_1x_coco
-bs_item=2
-fp_item=fp32
-run_mode=DP
-device_num=N1C8
-max_iter=100
-num_workers=2
-
-# get data
-bash test_tipc/static/${model_item}/benchmark_common/prepare.sh
-# run
-bash test_tipc/static/${model_item}/benchmark_common/run_benchmark.sh ${model_item} ${bs_item} ${fp_item} ${run_mode} ${device_num} ${max_iter} ${num_workers} 2>&1;

test_tipc/static/mask_rcnn_r50_fpn_1x_coco/benchmark_common/prepare.sh

-#!/usr/bin/env bash
-# 执行路径在模型库的根目录下
-################################# 安装框架 如:
-echo "*******prepare benchmark start ***********"
-pip install -U pip
-echo `pip --version`
-pip install Cython
-pip install -r requirements.txt
-
-################################# 准备训练数据 如:
-wget -nc -P static/dataset/coco/ https://paddledet.bj.bcebos.com/data/coco_benchmark.tar
-cd ./static/dataset/coco/ && tar -xf coco_benchmark.tar && mv -u coco_benchmark/* .
-rm -rf coco_benchmark/ && cd ../../../
-echo "*******prepare benchmark end***********"

test_tipc/static/mask_rcnn_r50_fpn_1x_coco/benchmark_common/run_benchmark.sh

-#!/usr/bin/env bash
-# Test training benchmark for a model.
-# Usage：bash run_benchmark.sh ${model_item} ${bs_item} ${fp_item} ${run_mode} ${device_num}
-function _set_params(){
-    model_item=${1:-"model_item"}   # (必选) 模型 item
-    base_batch_size=${2:-"2"}       # (必选) 如果是静态图单进程，则表示每张卡上的BS，需在训练时*卡数
-    fp_item=${3:-"fp32"}            # (必选) fp32|fp16
-    run_mode=${4:-"DP"}             # (必选) MP模型并行|DP数据并行|PP流水线并行|混合并行DP1-MP1-PP1|DP1-MP4-PP1
-    device_num=${5:-"N1C1"}         # (必选) 使用的卡数量，N1C1|N1C8|N4C32 （4机32卡）
-    profiling=${PROFILING:-"false"}      # (必选) Profiling  开关，默认关闭，通过全局变量传递
-    model_repo="PaddleDetection"          # (必选) 模型套件的名字
-    speed_unit="samples/sec"         # (必选)速度指标单位
-    skip_steps=10                  # (必选)解析日志，跳过模型前几个性能不稳定的step
-    keyword="ips:"                 # (必选)解析日志，筛选出性能数据所在行的关键字
-    convergence_key="loss:"        # (可选)解析日志，筛选出收敛数据所在行的关键字 如：convergence_key="loss:"
-    max_iter=${6:-"100"}           # （可选）需保证模型执行时间在5分钟内，需要修改代码提前中断的直接提PR 合入套件；或使用max_epoch参数
-    num_workers=${7:-"8"}                  # (可选)
-#   以下为通用执行命令，无特殊可不用修改
-    model_name=${model_item}_bs${base_batch_size}_${fp_item}_${run_mode}  # (必填) 且格式不要改动,与竞品名称对齐
-    device=${CUDA_VISIBLE_DEVICES//,/ }
-    arr=(${device})
-    num_gpu_devices=${#arr[*]}
-    run_log_path=${TRAIN_LOG_DIR:-$(pwd)}  # （必填） TRAIN_LOG_DIR  benchmark框架设置该参数为全局变量
-    profiling_log_path=${PROFILING_LOG_DIR:-$(pwd)}  # （必填） PROFILING_LOG_DIR benchmark框架设置该参数为全局变量
-    speed_log_path=${LOG_PATH_INDEX_DIR:-$(pwd)}
-
-    train_log_file=${run_log_path}/${model_repo}_${model_name}_${device_num}_log
-    profiling_log_file=${profiling_log_path}/${model_repo}_${model_name}_${device_num}_profiling
-    speed_log_file=${speed_log_path}/${model_repo}_${model_name}_${device_num}_speed
-}
-function _train(){
-    export FLAGS_eager_delete_tensor_gb=0.0
-    export FLAGS_fraction_of_gpu_memory_to_use=0.98
-    export FLAGS_memory_fraction_of_eager_deletion=1.0
-    cd ./static
-    batch_size=${base_batch_size}  # 如果模型跑多卡单进程时,请在_train函数中计算出多卡需要的bs
-    echo "current CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES}, model_name=${model_name}, device_num=${device_num}, is profiling=${profiling}"
-
-    if [ ${profiling} = "true" ];then
-        log_file=${profiling_log_file}
-        is_profiler=1
-    else
-        log_file=${train_log_file}
-        is_profiler=0
-    fi
-    if [ ${fp_item} = "fp16" ]; then
-        use_fp16_cmd="--fp16"
-    else
-        use_fp16_cmd=""
-    fi
-
-    train_cmd="-c configs/mask_rcnn_r50_fpn_1x.yml -o LearningRate.base_lr=0.001 snapshot_iter=100000 \
-         TrainReader.batch_size=${batch_size} \
-         max_iters=${max_iter} log_iter=1 \
-         TrainReader.worker_num=${num_workers} ${use_fp16_cmd} \
-         --is_profiler=${is_profiler} "
-#   以下为通用执行命令，无特殊可不用修改
-    case ${run_mode} in
-    DP) if [[ ${device_num} = "N1C1" ]];then
-            echo "run ${run_mode} ${device_num}"
-            train_cmd="python -u tools/train.py ${train_cmd}"
-        else
-            rm -rf ./mylog
-            train_cmd="python -m paddle.distributed.launch --log_dir=./mylog --gpus=$CUDA_VISIBLE_DEVICES \
-                  tools/train.py ${train_cmd}"
-        fi
-        ;;
-    DP1-MP1-PP1)  echo "run run_mode: DP1-MP1-PP1" ;;
-    *) echo "choose run_mode "; exit 1;
-    esac
-
-    echo "train_cmd: ${train_cmd}  log_file: ${log_file}"
-    timeout 15m ${train_cmd} > ${log_file} 2>&1
-    if [ $? -ne 0 ];then
-        echo -e "${model_name}, FAIL"
-    else
-        echo -e "${model_name}, SUCCESS"
-    fi
-    # kill -9 `ps -ef|grep 'python'|awk '{print $2}'`
-    if [ ${device_num} != "N1C1" -a -d mylog ]; then
-        rm ${log_file}
-        cp mylog/workerlog.0 ${log_file}
-    fi
-    cd ../
-}
-source ${BENCHMARK_ROOT}/scripts/run_model.sh   # 在该脚本中会对符合benchmark规范的log使用analysis.py 脚本进行性能数据解析;如果不联调只想要产出训练log可以注掉本行,提交时需打开
-_set_params $@
-#_train       # 如果只产出训练log,不解析,可取消注释
-_run     # 该函数在run_model.sh中,执行时会调用_train; 如果不联调只产出训练log可以注掉本行,提交时需打开

test_tipc/static/yolov3_darknet53_270e_coco/N1C1/yolov3_darknet53_270e_coco_bs8_fp32_DP.sh

-model_item=yolov3_darknet53_270e_coco
-bs_item=8
-fp_item=fp32
-run_mode=DP
-device_num=N1C1
-max_iter=100
-num_workers=8
-
-# get data
-bash test_tipc/static/${model_item}/benchmark_common/prepare.sh
-# run
-bash test_tipc/static/${model_item}/benchmark_common/run_benchmark.sh ${model_item} ${bs_item} ${fp_item} ${run_mode} ${device_num} ${max_iter} ${num_workers} 2>&1;
-# run profiling
-sleep 10;
-export PROFILING=true
-bash test_tipc/static/${model_item}/benchmark_common/run_benchmark.sh ${model_item} ${bs_item} ${fp_item} ${run_mode} ${device_num} 11 ${num_workers} 2>&1;

test_tipc/static/yolov3_darknet53_270e_coco/N1C8/yolov3_darknet53_270e_coco_bs8_fp32_DP.sh

-model_item=yolov3_darknet53_270e_coco
-bs_item=8
-fp_item=fp32
-run_mode=DP
-device_num=N1C8
-max_iter=100
-num_workers=8
-
-# get data
-bash test_tipc/static/${model_item}/benchmark_common/prepare.sh
-# run
-bash test_tipc/static/${model_item}/benchmark_common/run_benchmark.sh ${model_item} ${bs_item} ${fp_item} ${run_mode} ${device_num} ${max_iter} ${num_workers} 2>&1;

test_tipc/static/yolov3_darknet53_270e_coco/benchmark_common/prepare.sh

-#!/usr/bin/env bash
-# 执行路径在模型库的根目录下
-################################# 安装框架 如:
-echo "*******prepare benchmark start ***********"
-pip install -U pip
-echo `pip --version`
-pip install Cython
-pip install -r requirements.txt
-
-################################# 准备训练数据 如:
-wget -nc -P static/dataset/coco/ https://paddledet.bj.bcebos.com/data/coco_benchmark.tar
-cd ./static/dataset/coco/ && tar -xf coco_benchmark.tar && mv -u coco_benchmark/* .
-rm -rf coco_benchmark/ && cd ../../../
-echo "*******prepare benchmark end***********"

test_tipc/static/yolov3_darknet53_270e_coco/benchmark_common/run_benchmark.sh

-#!/usr/bin/env bash
-# Test training benchmark for a model.
-# Usage：bash run_benchmark.sh ${model_item} ${bs_item} ${fp_item}  ${run_mode} ${device_num}
-function _set_params(){
-    model_item=${1:-"model_item"}   # (必选) 模型 item
-    base_batch_size=${2:-"2"}       # (必选) 如果是静态图单进程，则表示每张卡上的BS，需在训练时*卡数
-    fp_item=${3:-"fp32"}            # (必选) fp32|fp16
-    run_mode=${4:-"DP"}             # (必选) MP模型并行|DP数据并行|PP流水线并行|混合并行DP1-MP1-PP1|DP1-MP4-PP1
-    device_num=${5:-"N1C1"}         # (必选) 使用的卡数量，N1C1|N1C8|N4C32 （4机32卡）
-    profiling=${PROFILING:-"false"}      # (必选) Profiling  开关，默认关闭，通过全局变量传递
-    model_repo="PaddleDetection"          # (必选) 模型套件的名字
-    speed_unit="samples/sec"         # (必选)速度指标单位
-    skip_steps=10                  # (必选)解析日志，跳过模型前几个性能不稳定的step
-    keyword="ips:"                 # (必选)解析日志，筛选出性能数据所在行的关键字
-    convergence_key="loss:"        # (可选)解析日志，筛选出收敛数据所在行的关键字 如：convergence_key="loss:"
-    max_iter=${6:-"100"}           # （可选）需保证模型执行时间在5分钟内，需要修改代码提前中断的直接提PR 合入套件；或使用max_epoch参数
-    num_workers=${7:-"8"}                  # (可选)
-#   以下为通用执行命令，无特殊可不用修改
-    model_name=${model_item}_bs${base_batch_size}_${fp_item}_${run_mode}  # (必填) 且格式不要改动,与竞品名称对齐
-    device=${CUDA_VISIBLE_DEVICES//,/ }
-    arr=(${device})
-    num_gpu_devices=${#arr[*]}
-    run_log_path=${TRAIN_LOG_DIR:-$(pwd)}  # （必填） TRAIN_LOG_DIR  benchmark框架设置该参数为全局变量
-    profiling_log_path=${PROFILING_LOG_DIR:-$(pwd)}  # （必填） PROFILING_LOG_DIR benchmark框架设置该参数为全局变量
-    speed_log_path=${LOG_PATH_INDEX_DIR:-$(pwd)}
-
-    train_log_file=${run_log_path}/${model_repo}_${model_name}_${device_num}_log
-    profiling_log_file=${profiling_log_path}/${model_repo}_${model_name}_${device_num}_profiling
-    speed_log_file=${speed_log_path}/${model_repo}_${model_name}_${device_num}_speed
-}
-function _train(){
-    export FLAGS_eager_delete_tensor_gb=0.0
-    export FLAGS_fraction_of_gpu_memory_to_use=0.98
-    export FLAGS_memory_fraction_of_eager_deletion=1.0
-    cd ./static
-    batch_size=${base_batch_size}  # 如果模型跑多卡单进程时,请在_train函数中计算出多卡需要的bs
-    echo "current CUDA_VISIBLE_DEVICES=${CUDA_VISIBLE_DEVICES}, model_name=${model_name}, device_num=${device_num}, is profiling=${profiling}"
-
-    if [ ${profiling} = "true" ];then
-        log_file=${profiling_log_file}
-        is_profiler=1
-    else
-        log_file=${train_log_file}
-        is_profiler=0
-    fi
-    if [ ${fp_item} = "fp16" ]; then
-        use_fp16_cmd="--fp16"
-    else
-        use_fp16_cmd=""
-    fi
-
-    train_cmd="-c configs/yolov3_darknet.yml -o LearningRate.base_lr=0.002 snapshot_iter=100000 \
-         TrainReader.batch_size=${batch_size} \
-         max_iters=${max_iter} log_iter=1 \
-         TrainReader.worker_num=${num_workers} ${use_fp16_cmd} \
-         --is_profiler=${is_profiler} "
-#   以下为通用执行命令，无特殊可不用修改
-    case ${run_mode} in
-    DP) if [[ ${device_num} = "N1C1" ]];then
-            echo "run ${run_mode} ${device_num}"
-            train_cmd="python -u tools/train.py ${train_cmd}"
-        else                
-            rm -rf ./mylog
-            train_cmd="python -m paddle.distributed.launch --log_dir=./mylog --gpus=$CUDA_VISIBLE_DEVICES \
-                  tools/train.py ${train_cmd}"
-        fi
-        ;;
-    DP1-MP1-PP1)  echo "run run_mode: DP1-MP1-PP1" ;;
-    *) echo "choose run_mode "; exit 1;
-    esac
-    
-    echo "train_cmd: ${train_cmd}  log_file: ${log_file}"
-    timeout 15m ${train_cmd} > ${log_file} 2>&1
-    if [ $? -ne 0 ];then
-        echo -e "${model_name}, FAIL"
-    else
-        echo -e "${model_name}, SUCCESS"
-    fi
-    # kill -9 `ps -ef|grep 'python'|awk '{print $2}'`
-    if [ ${device_num} != "N1C1" -a -d mylog ]; then
-        rm ${log_file}
-        cp mylog/workerlog.0 ${log_file}
-    fi
-    cd ../
-}
-source ${BENCHMARK_ROOT}/scripts/run_model.sh   # 在该脚本中会对符合benchmark规范的log使用analysis.py 脚本进行性能数据解析;如果不联调只想要产出训练log可以注掉本行,提交时需打开
-_set_params $@
-#_train       # 如果只产出训练log,不解析,可取消注释
-_run     # 该函数在run_model.sh中,执行时会调用_train; 如果不联调只产出训练log可以注掉本行,提交时需打开