recommend-emr-dataprocess/ad/02_ad_model_dnn_v11_update.sh

#!/bin/sh
set -x

export PATH=$SPARK_HOME/bin:$PATH
export HADOOP_CONF_DIR=/etc/taihao-apps/hadoop-conf
export JAVA_HOME=/usr/lib/jvm/java-1.8.0
export PREDICT_CACHE_PATH=/root/fengzhoutian/xgboost-dev/predict_cache/
export SEGMENT_BASE_PATH=/dw/recommend/model/36_model_attachment/score_calibration_file

sh_path=$(cd $(dirname $0); pwd)
source ${sh_path}/00_common.sh

source /root/anaconda3/bin/activate py37


# 全局常量
LOG_PREFIX=广告模型训练任务
HADOOP=/opt/apps/HADOOP-COMMON/hadoop-common-current/bin/hadoop
TRAIN_PATH=/dw/recommend/model/31_ad_sample_data_v5
BUCKET_FEATURE_PATH=/dw/recommend/model/33_ad_train_data_v5
TABLE=alg_recsys_ad_sample_all
# 特征文件名
feature_file=20240703_ad_feature_name.txt
# 模型本地临时保存路径
model_local_home=/root/fengzhoutian/xgboost-dev/

# 模型HDFS保存路径，测试时修改为其他路径，避免影响线上
MODEL_PATH=/dw/recommend/model/35_ad_model
# 预测结果保存路径，测试时修改为其他路径，避免影响线上
PREDICT_RESULT_SAVE_PATH=/dw/recommend/model/34_ad_predict_data
# 模型OSS保存路径，测试时修改为其他路径，避免影响线上
MODEL_OSS_PATH=oss://art-recommend.oss-cn-hangzhou-internal.aliyuncs.com/fengzhoutian/
# 线上模型名，测试时修改为其他模型名，避免影响线上
model_name=model_xgb
model_ver=351_1000_14d_v1
model_name=${model_name}_${model_ver}
model_local_home=${model_local_home}/${model_name}
# 线上校准文件名
OSS_CALIBRATION_FILE_NAME=${model_name}_calibration
# 用于存放一些临时的文件
PREDICT_CACHE_PATH=/root/fengzhoutian/xgboost-dev/predict_cache/


# 本地保存HDFS模型路径文件，测试时修改为其他模型名，避免影响线上
model_path_file=${model_local_home}/online_model_path.txt
# 获取当前是星期几，1表示星期一
current_day_of_week="$(date +"%u")"

# 任务开始时间
start_time=$(date +%s)
# 前一天
today_early_1="$(date -d '1 days ago' +%Y%m%d)"
# 线上模型在HDFS中的路径
# online_model_path=`cat ${model_path_file}`
# 训练用的数据路径
train_data_path=""
train_data_days=14
# 评估用的数据路径
predict_date_path=""
#评估结果保存路径
new_model_predict_result_path=""
# 模型保存路径
model_save_path=""
# 评测结果保存路径，后续需要根据此文件评估是否要更新模型
predict_analyse_file_path=""
# 校准文件保存路径
calibration_file_path=""

# 保存模型评估的分析结果
old_incr_rate_avg=0
new_incr_rate_avg=0
# Top10的详情
top10_msg=""
# AUC值
old_auc=0
new_auc=0

declare -A real_score_map
declare -A old_score_map
declare -A new_score_map

# 校验命令的退出码
check_run_status() {
    local status=$1
    local step_start_time=$2
    local step_name=$3
    local msg=$4

    local step_end_time=$(date +%s)
    local step_elapsed=$(($step_end_time - $step_start_time))

    if [[ -n "${old_auc}" && "${old_auc}" != "0" ]]; then
      msg+="\n\t - 老模型AUC: ${old_auc}"
    fi
    if [[ -n "${new_auc}" && "${new_auc}" != "0" ]]; then
      msg+="\n\t - 新模型AUC: ${new_auc}"
    fi


    if [ ${status} -ne 0 ]; then
        echo "${LOG_PREFIX} -- ${step_name}失败: 耗时 ${step_elapsed}"
        local elapsed=$(($step_end_time - $start_time))
        /root/anaconda3/bin/python ${sh_path}/ad_monitor_util.py --level error --msg "${msg}" --start "${start_time}" --elapsed "${elapsed}" --top10 "${top10_msg}"
        exit 1
    else
        echo "${LOG_PREFIX} -- ${step_name}成功: 耗时 ${step_elapsed}"
    fi
}

send_success_upload_msg(){ 
  # 发送更新成功通知
  local msg=" 广告模型文件更新完成"
  msg+="\n\t - 老模型AUC: ${old_auc}"
  msg+="\n\t - 新模型AUC: ${new_auc}"
  msg+="\n\t - 老模型Top10差异平均值: ${old_incr_rate_avg}"
  msg+="\n\t - 新模型Top10差异平均值: ${new_incr_rate_avg}"
  msg+="\n\t - 模型在HDFS中的路径: ${model_save_path}"
  msg+="\n\t - 模型上传OSS中的路径: ${MODEL_OSS_PATH}/${model_name}.tar.gz"

  local step_end_time=$(date +%s)
  local elapsed=$((${step_end_time} - ${start_time}))

  /root/anaconda3/bin/python ${sh_path}/ad_monitor_util.py --level info --msg "${msg}" --start "${start_time}" --elapsed "${elapsed}" --top10 "${top10_msg}"
}

init() {
  set +x
  declare -a date_keys=()
  local count=1
  local current_data="$(date -d "${today_early_1} -1 day" +%Y%m%d)"
  # 循环获取前 n 天的非节日日期
  while [[ ${count} -le $train_data_days ]]; do
    date_key=$(date -d "${current_data}" +%Y%m%d)
    # 判断是否是节日，并拼接训练数据路径
    if [ $(is_not_holidays ${date_key}) -eq 1 ]; then

      # 将 date_key 放入数组
      date_keys+=("${date_key}")

      if [[ -z ${train_data_path} ]]; then
        train_data_path="${BUCKET_FEATURE_PATH}/${date_key}"
      else
        train_data_path="${BUCKET_FEATURE_PATH}/${date_key},${train_data_path}"
      fi 
      count=$((count + 1))
    else
      echo "日期: ${date_key}是节日，跳过"
    fi
    current_data=$(date -d "${current_data} -1 day" +%Y%m%d)
  done

  last_index=$((${#date_keys[@]} - 1))
  train_first_day=${date_keys[$last_index]}
  train_last_day=${date_keys[0]}

  model_save_path=${MODEL_PATH}/${model_name}_${train_first_day: -4}_${train_last_day: -4}
  predict_date_path=${BUCKET_FEATURE_PATH}/${today_early_1}
  new_model_predict_result_path=${PREDICT_RESULT_SAVE_PATH}/${today_early_1}_${model_ver}_${train_first_day: -4}_${train_last_day: -4}
  online_model_predict_result_path=${PREDICT_RESULT_SAVE_PATH}/${today_early_1}_${model_ver}_${online_model_path: -9}
  predict_analyse_file_path=${model_local_home}/predict_analyse_file/${today_early_1}_${model_ver}_analyse.txt
  calibration_file_path=${model_local_home}/${OSS_CALIBRATION_FILE_NAME}.txt

  echo "init param train_data_path: ${train_data_path}"
  echo "init param predict_date_path: ${predict_date_path}"
  echo "init param new_model_predict_result_path: ${new_model_predict_result_path}"
  echo "init param online_model_predict_result_path: ${online_model_predict_result_path}"
  echo "init param model_save_path: ${model_save_path}"
  echo "init param online_model_path: ${online_model_path}"
  echo "init param feature_file: ${feature_file}"
  echo "init param model_name: ${model_name}"
  echo "init param model_local_home: ${model_local_home}"
  echo "init param model_oss_path: ${MODEL_OSS_PATH}"
  echo "init param predict_analyse_file_path: ${predict_analyse_file_path}"
  echo "init param calibration_file_path: ${calibration_file_path}"
  echo "init param current_day_of_week: ${current_day_of_week}"

  echo "当前Python环境安装的Python版本: $(python --version)"
  echo "当前Python环境安装的三方包: $(python -m pip list)"
  set -x
  mkdir -p ${model_local_home}
}

# 校验大数据任务是否执行完成
check_ad_hive() {
  local step_start_time=$(date +%s)
  local max_hour=05
  local max_minute=30
  local elapsed=0
  while true; do
      local python_return_code=$(python ${sh_path}/ad_utils.py --excute_program check_ad_origin_hive --partition ${today_early_1} --hh 23)

      elapsed=$(($(date +%s) - ${step_start_time}))
      if [ "${python_return_code}" -eq 0 ]; then
          break
      fi
      echo "Python程序返回非0值，等待五分钟后再次调用。"
      sleep 300
      local current_hour=$(date +%H)
      local current_minute=$(date +%M)
      if (( ${current_hour} > ${max_hour} || ( ${current_hour} == ${max_hour} && ${current_minute} >= ${max_minute} ) )); then
          local msg="大数据数据生产校验失败, 分区: ${today_early_1}"
          echo -e "${LOG_PREFIX} -- 大数据数据生产校验 -- ${msg}: 耗时 ${elapsed}"
          /root/anaconda3/bin/python ${sh_path}/ad_monitor_util.py --level error --msg "${msg}" --start "${start_time}" --elapsed "${elapsed}"
          exit 1
      fi
  done
  echo "${LOG_PREFIX} -- 大数据数据生产校验 -- 大数据数据生产校验通过: 耗时 ${elapsed}"
}

bucket_feature_from_origin_to_hive() {
  (
    export outputTable=ad_easyrec_train_data_v3_sampled
    source ${sh_path}/25_xgb_make_data_origin_bucket.sh
    make_bucket_feature_from_origin_to_hive
  )
}

run_pai_flow() {
  local step_start_time=$(date +%s)

  python ad/pai_flow_operator2.py
  
  local return_code=$?
  check_run_status ${return_code} ${step_start_time} "PAI工作流任务" "PAI工作流执行失败"
}

calc_model_predict() {
  local count=0
  local max_line=10
  local old_total_diff=0
  local new_total_diff=0
  top10_msg="| CID  | 老模型相对真实CTCVR的变化 | 新模型相对真实CTCVR的变化 |"
  top10_msg+=" \n| ---- | --------- | -------- |"
  while read -r line && [ ${count} -lt ${max_line} ]; do

      # 使用 ! 取反判断，只有当行中不包含 "cid" 时才执行继续的逻辑
      if [[ "${line}" == *"cid"* ]]; then
          continue
      fi

      read -a numbers <<< "${line}"

      # 分数分别保存
      real_score_map[${numbers[0]}]=${numbers[3]}
      old_score_map[${numbers[0]}]=${numbers[6]}
      new_score_map[${numbers[0]}]=${numbers[7]}

      # 拼接Top10详情的飞书消息
      top10_msg="${top10_msg} \n| ${numbers[0]} | ${numbers[6]} | ${numbers[7]} | "

      # 计算top10相对误差绝对值的均值
      old_abs_score=$( echo "${numbers[6]} * ((${numbers[6]} >= 0) * 2 - 1)" | bc -l )
      new_abs_score=$( echo "${numbers[7]} * ((${numbers[7]} >= 0) * 2 - 1)" | bc -l )

      old_total_diff=$( echo "${old_total_diff} + ${old_abs_score}" | bc -l )
      new_total_diff=$( echo "${new_total_diff} + ${new_abs_score}" | bc -l )

      count=$((${count} + 1))

  done < "${predict_analyse_file_path}"

  local return_code=$?
  check_run_status ${return_code} ${step_start_time} "计算Top10差异" "计算Top10差异异常"

  old_incr_rate_avg=$( echo "scale=6; ${old_total_diff} / ${count}" | bc -l )
  check_run_status $? ${step_start_time} "计算老模型Top10差异" "计算老模型Top10差异异常"


  new_incr_rate_avg=$( echo "scale=6; ${new_total_diff} / ${count}" | bc -l )
  check_run_status $? ${step_start_time} "计算新模型Top10差异" "计算新模型Top10差异异常"

  echo "老模型Top10差异平均值: ${old_incr_rate_avg}"
  echo "新模型Top10差异平均值: ${new_incr_rate_avg}"
  echo "新老模型分数对比: "
  for cid in "${!new_score_map[@]}"; do
    echo "\t CID: $cid, 老模型分数: ${old_score_map[$cid]}, 新模型分数: ${new_score_map[$cid]}"
  done
}

calc_auc() {
  old_auc=`cat ${PREDICT_CACHE_PATH}/old_1.txt | /root/sunmingze/AUC/AUC`
  new_auc=`cat ${PREDICT_CACHE_PATH}/new_1.txt | /root/sunmingze/AUC/AUC`
}

model_predict() {
  # 线上模型评估最新的数据
  local step_start_time=$(date +%s)
  /opt/apps/SPARK3/spark-3.3.1-hadoop3.2-1.0.5/bin/spark-class org.apache.spark.deploy.SparkSubmit \
  --class com.tzld.piaoquan.recommend.model.pred_01_xgb_ad_hdfsfile_20240813 \
  --master yarn --driver-memory 1G --executor-memory 3G --executor-cores 1 --num-executors 30 \
  --conf spark.yarn.executor.memoryoverhead=1024 \
  --conf spark.shuffle.service.enabled=true \
  --conf spark.shuffle.service.port=7337 \
  --conf spark.shuffle.consolidateFiles=true \
  --conf spark.shuffle.manager=sort \
  --conf spark.storage.memoryFraction=0.4 \
  --conf spark.shuffle.memoryFraction=0.5 \
  --conf spark.default.parallelism=200 \
  /root/fengzhoutian/recommend-model/recommend-model-produce/target/recommend-model-produce-jar-with-dependencies.jar \
  featureFile:20240703_ad_feature_name.txt \
  testPath:${predict_date_path} \
  savePath:${online_model_predict_result_path} \
  negSampleRate:0.04 \
  modelPath:${online_model_path}

  local return_code=$?
  check_run_status ${return_code} ${step_start_time} "线上模型评估${predict_date_path: -8}的数据" "线上模型评估${predict_date_path: -8}的数据失败"
}

compare_predictions() {
  local step_start_time=$(date +%s)

  mkdir -p ${model_local_home}/predict_analyse_file
  # 结果分析
  python ${sh_path}/model_predict_analyse.py -op ${online_model_predict_result_path} -np ${new_model_predict_result_path} -af ${predict_analyse_file_path} -cf ${calibration_file_path}
  local python_return_code=$?
  check_run_status ${python_return_code} ${step_start_time} "分析线上模型评估${predict_date_path: -8}的数据" "分析线上模型评估${predict_date_path: -8}的数据失败"

  calc_model_predict

  calc_auc

  if (( $(echo "${new_incr_rate_avg} > 0.100000" | bc -l ) ));then 
    echo "线上模型评估${predict_date_path: -8}的数据，绝对误差大于0.1，请检查"
    check_run_status 1 ${step_start_time} "${predict_date_path: -8}的数据，绝对误差大于0.1" "线上模型评估${predict_date_path: -8}的数据，绝对误差大于0.1，请检查"
    exit 1
  fi 


  # 对比两个模型的差异
  score_diff=$( echo "${new_incr_rate_avg} - ${old_incr_rate_avg}" | bc -l )
  if (( $(echo "${score_diff} > 0.050000" | bc -l ) ));then 
    echo "两个模型评估${predict_date_path: -8}的数据，两个模型分数差异为: ${score_diff}, 大于0.05, 请检查"
    check_run_status 1 ${step_start_time} "两个模型评估${predict_date_path: -8}的数据" "两个模型评估${predict_date_path: -8}的数据，两个模型分数差异为: ${score_diff}, 大于0.05"
    exit 1
  fi 
}

draw_q_distribution() {
  local step_start_time=$(date +%s)

  python ${sh_path}/draw_predict_distribution.py -op ${online_model_predict_result_path} -np ${new_model_predict_result_path} --output ${today_early_1}_${model_ver}_${train_first_day: -4}_${train_last_day: -4}.png
  python_return_code=$?
}

make_train_node_conf() {
  train_data_path=''
  TABLE_PART_PREFIX="odps://loghubods/tables/ad_easyrec_train_data_v3_sampled/dt"
  declare -a date_keys=()
  local count=1
  local current_data="$(date -d "${today_early_1} -1 day" +%Y%m%d)"
  # 循环获取前 n 天的非节日日期
  while [[ ${count} -le $train_data_days ]]; do
    date_key=$(date -d "${current_data}" +%Y%m%d)
    # 判断是否是节日，并拼接训练数据路径
    if [ $(is_not_holidays ${date_key}) -eq 1 ]; then

      # 将 date_key 放入数组
      date_keys+=("${date_key}")

      if [[ -z ${train_data_path} ]]; then
        train_data_path="${TABLE_PART_PREFIX}=${date_key}"
      else
        train_data_path="${TABLE_PART_PREFIX}=${date_key},${train_data_path}"
      fi 
      count=$((count + 1))
    else
      echo "日期: ${date_key}是节日，跳过"
    fi
    current_data=$(date -d "${current_data} -1 day" +%Y%m%d)
  done

  last_index=$((${#date_keys[@]} - 1))
  train_first_day=${date_keys[$last_index]}
  train_last_day=${date_keys[0]}

  predict_data_path=${TABLE_PART_PREFIX}=${today_early_1}
}

# 主方法
main() {
  init

  check_ad_hive
  bucket_feature_from_origin_to_hive
  make_train_node_conf

  run_pai_flow
}


main