rov-offline/rov_train.py

import os
import random
import time

import lightgbm as lgb
import pandas as pd

from sklearn.model_selection import train_test_split
from sklearn.metrics import mean_absolute_error, r2_score, mean_absolute_percentage_error

from config import set_config
from utils import read_from_pickle, write_to_pickle, data_normalization, \
    request_post, filter_video_status, update_video_w_h_rate
from log import Log
from db_helper import RedisHelper, MysqlHelper

config_, env = set_config()
log_ = Log()


def process_data(filename):
    """
    数据清洗、预处理
    :param filename: type-DataFrame
    :return: x, y, video_ids, features
    """
    # 获取数据
    data = read_from_pickle(filename)
    # 获取y,并将 y <= 0 的值更新为1
    data['futre7dayreturn'].loc[data['futre7dayreturn'] <= 0] = 1
    y = data['futre7dayreturn']

    # 获取视频id列
    video_ids = data['videoid']

    # 获取x
    drop_columns = ['videoid', 'dt', 'futre7dayreturn', 'videotags', 'words_without_tags']
    x = data.drop(columns=drop_columns)
    # 计算后一天的回流比前一天的回流差值
    x['stage_four_return_added'] = x['stage_four_retrn'] - x['stage_three_retrn']
    x['stage_three_return_added'] = x['stage_three_retrn'] - x['stage_two_retrn']
    x['stage_two_return_added'] = x['stage_two_retrn'] - x['stage_one_retrn']
    # 计算后一天回流比前一天回流的增长率
    x['stage_four_return_ratio'] = x['stage_four_return_added'] / x['stage_four_retrn']
    x['stage_three_return_ratio'] = x['stage_three_return_added'] / x['stage_three_retrn']
    x['stage_two_return_ratio'] = x['stage_two_return_added'] / x['stage_two_retrn']

    # 缺失值填充为0
    x.fillna(0)

    # 获取当前所使用的特征列表
    features = list(x)

    return x, y, video_ids, features


def process_predict_data(filename):
    """
    预测数据清洗、预处理
    :param filename: type-DataFrame
    :return: x, y, video_ids, features
    """
    # 获取数据
    data = read_from_pickle(filename)

    # 获取视频id列
    video_ids = data['videoid']
    # 视频状态过滤
    video_id_list = [int(video_id) for video_id in video_ids]
    filtered_videos = [str(item) for item in filter_video_status(video_ids=video_id_list)]
    data = data.loc[data['videoid'].isin(filtered_videos)]

    video_id_final = data['videoid']

    # 获取x
    drop_columns = ['videoid', 'dt', 'futre7dayreturn', 'videotags', 'words_without_tags']
    x = data.drop(columns=drop_columns)
    # 计算后一天的回流比前一天的回流差值
    x['stage_four_return_added'] = x['stage_four_retrn'] - x['stage_three_retrn']
    x['stage_three_return_added'] = x['stage_three_retrn'] - x['stage_two_retrn']
    x['stage_two_return_added'] = x['stage_two_retrn'] - x['stage_one_retrn']
    # 计算后一天回流比前一天回流的增长率
    x['stage_four_return_ratio'] = x['stage_four_return_added'] / x['stage_four_retrn']
    x['stage_three_return_ratio'] = x['stage_three_return_added'] / x['stage_three_retrn']
    x['stage_two_return_ratio'] = x['stage_two_return_added'] / x['stage_two_retrn']

    # 缺失值填充为0
    x.fillna(0)

    return x, video_id_final


def train(x, y, features):
    """
    训练模型
    :param x: X
    :param y: Y
    :param features: 特征列表
    :return: None
    """
    # 训练集、测试集分割
    x_train, x_test, y_train, y_test = train_test_split(x, y, test_size=0.33)
    log_.info('x_train shape: {}, y_train shape: {}'.format(x_train.shape, y_train.shape))
    log_.info('x_test shape: {}, y_test shape: {}'.format(x_test.shape, y_test.shape))

    # 训练参数设置
    params = {
        "objective": "regression",
        "reg_sqrt": True,
        "metric": "mape",
        "max_depth": -1,
        "num_leaves": 50,
        "learning_rate": 0.1,
        "bagging_fraction": 0.7,
        "feature_fraction": 0.7,
        "bagging_freq": 8,
        "bagging_seed": 2018,
        "lambda_l1": 0.11,
        "boosting": "dart",
        "nthread": 4,
        "verbosity": -1
    }
    # 初始化数据集
    train_set = lgb.Dataset(data=x_train, label=y_train)
    test_set = lgb.Dataset(data=x_test, label=y_test)
    # 模型训练
    evals_result = {}

    model = lgb.train(params=params, train_set=train_set, num_boost_round=5000,
                      valid_sets=[test_set], early_stopping_rounds=100,
                      verbose_eval=100, evals_result=evals_result)

    # 将模型特征重要度存入csv
    feature_importance_data = {'feature': features, 'feature_importance': model.feature_importance()}
    feature_importance_filename = 'model_feature_importance.csv'
    pack_result_to_csv(filename=feature_importance_filename, sort_columns=['feature_importance'],
                       ascending=False, **feature_importance_data)

    # 测试集预测
    pre_y_test = model.predict(data=x_test, num_iteration=model.best_iteration)
    y_test = y_test.values

    err_mape = mean_absolute_percentage_error(y_test, pre_y_test)
    r2 = r2_score(y_test, pre_y_test)

    # 将测试集结果存入csv
    test_data = {'pre_y_test': pre_y_test, 'y_test': y_test}
    test_result_filename = 'test_result.csv'
    pack_result_to_csv(filename=test_result_filename, sort_columns=['pre_y_test'], ascending=False, **test_data)

    log_.info('err_mape={}, r2={}'.format(err_mape, r2))

    # 保存模型
    write_to_pickle(data=model, filename=config_.MODEL_FILENAME)


def pack_result_to_csv(filename, sort_columns=None, filepath=config_.DATA_DIR_PATH, ascending=True, **data):
    """
    打包数据并存入csv
    :param filename: csv文件名
    :param sort_columns: 指定排序列名列名，type-list, 默认为None
    :param filepath: csv文件存放路径，默认为config_.DATA_DIR_PATH
    :param ascending: 是否按指定列的数组升序排列，默认为True，即升序排列
    :param data: 数据, type-dict
    :return: None
    """
    if not os.path.exists(filepath):
        os.makedirs(filepath)
    file = os.path.join(filepath, filename)
    df = pd.DataFrame(data=data)
    if sort_columns:
        df = df.sort_values(by=sort_columns, ascending=ascending)
    df.to_csv(file, index=False)


def pack_list_result_to_csv(filename, data, columns=None, sort_columns=None, filepath=config_.DATA_DIR_PATH, ascending=True):
    """
    打包数据并存入csv, 数据为字典列表
    :param filename: csv文件名
    :param data: 数据，type-list [{}, {},...]
    :param columns: 列名顺序
    :param sort_columns: 指定排序列名列名，type-list, 默认为None
    :param filepath: csv文件存放路径，默认为config_.DATA_DIR_PATH
    :param ascending: 是否按指定列的数组升序排列，默认为True，即升序排列
    :return: None
    """
    if not os.path.exists(filepath):
        os.makedirs(filepath)
    file = os.path.join(filepath, filename)
    df = pd.DataFrame(data=data)
    if sort_columns:
        df = df.sort_values(by=sort_columns, ascending=ascending)
    df.to_csv(file, index=False, columns=columns)


def predict():
    """预测"""
    # 读取预测数据并进行清洗
    x, video_ids = process_predict_data(config_.PREDICT_DATA_FILENAME)
    log_.info('predict data shape: x={}'.format(x.shape))
    # 获取训练好的模型
    model = read_from_pickle(filename=config_.MODEL_FILENAME)
    # 预测
    y_ = model.predict(x)
    log_.info('predict finished!')

    # 将结果进行归一化到[0, 100]
    normal_y_ = data_normalization(list(y_))
    log_.info('normalization finished!')

    # 按照normal_y_降序排序
    predict_data = []
    for i, video_id in enumerate(video_ids):
        data = {'video_id': video_id, 'normal_y_': normal_y_[i], 'y_': y_[i]}
        predict_data.append(data)
    predict_data_sorted = sorted(predict_data, key=lambda temp: temp['normal_y_'], reverse=True)

    # 按照排序，从100以固定差值做等差递减，以该值作为rovScore
    predict_result = []
    redis_data = {}
    json_data = []
    video_id_list = []
    for j, item in enumerate(predict_data_sorted):
        video_id = int(item['video_id'])
        rov_score = 100 - j * config_.ROV_SCORE_D
        item['rov_score'] = rov_score
        predict_result.append(item)
        redis_data[video_id] = rov_score
        json_data.append({'videoId': video_id, 'rovScore': rov_score})
        video_id_list.append(video_id)

    # 打包预测结果存入csv
    predict_result_filename = 'predict.csv'
    pack_list_result_to_csv(filename=predict_result_filename,
                            data=predict_result,
                            columns=['video_id', 'rov_score', 'normal_y_', 'y_'],
                            sort_columns=['rov_score'],
                            ascending=False)

    # 上传redis
    key_name = config_.RECALL_KEY_NAME_PREFIX + time.strftime('%Y%m%d')
    redis_helper = RedisHelper()
    redis_helper.add_data_with_zset(key_name=key_name, data=redis_data)
    log_.info('data to redis finished!')

    # 清空修改ROV的视频数据
    redis_helper.del_keys(key_name=config_.UPDATE_ROV_KEY_NAME)

    # 通知后端更新数据
    log_.info('json_data count = {}'.format(len(json_data)))
    result = request_post(request_url=config_.NOTIFY_BACKEND_UPDATE_ROV_SCORE_URL, request_data={'videos': json_data})
    if result['code'] == 0:
        log_.info('notify backend success!')
    else:
        log_.error('notify backend fail!')

    # ##### 下线
    # # 更新视频的宽高比数据
    # if video_id_list:
    #     update_video_w_h_rate(video_ids=video_id_list,
    #                           key_name=config_.W_H_RATE_UP_1_VIDEO_LIST_KEY_NAME['rov_recall'])
    #     log_.info('update video w_h_rate to redis finished!')


def predict_test():
    """测试环境数据生成"""
    # 获取测试环境中最近发布的40000条视频
    sql = "SELECT id FROM wx_video ORDER BY id DESC LIMIT 40000;"
    mysql_helper = MysqlHelper()
    data = mysql_helper.get_data(sql=sql)
    video_ids = [video[0] for video in data]
    # 视频状态过滤
    filtered_videos = filter_video_status(video_ids)
    log_.info('filtered_videos count = {}'.format(len(filtered_videos)))
    # 随机生成 0-100 数作为分数
    redis_data = {}
    json_data = []
    for video_id in filtered_videos:
        score = random.uniform(0, 100)
        redis_data[video_id] = score
        json_data.append({'videoId': video_id, 'rovScore': score})
    log_.info('json_data count = {}'.format(len(json_data)))
    # 上传Redis
    redis_helper = RedisHelper()
    key_name = config_.RECALL_KEY_NAME_PREFIX + time.strftime('%Y%m%d')
    redis_helper.add_data_with_zset(key_name=key_name, data=redis_data)
    log_.info('test data to redis finished!')
    # 清空修改ROV的视频数据
    redis_helper.del_keys(key_name=config_.UPDATE_ROV_KEY_NAME)
    # 通知后端更新数据
    result = request_post(request_url=config_.NOTIFY_BACKEND_UPDATE_ROV_SCORE_URL, request_data={'videos': json_data})
    if result['code'] == 0:
        log_.info('notify backend success!')
    else:
        log_.error('notify backend fail!')
    # ##### 下线
    # # 更新视频的宽高比数据
    # if filtered_videos:
    #     update_video_w_h_rate(video_ids=filtered_videos,
    #                           key_name=config_.W_H_RATE_UP_1_VIDEO_LIST_KEY_NAME['rov_recall'])
    #     log_.info('update video w_h_rate to redis finished!')


if __name__ == '__main__':
    log_.info('rov model train start...')
    train_start = time.time()
    train_filename = config_.TRAIN_DATA_FILENAME
    X, Y, videos, fea = process_data(filename=train_filename)
    log_.info('X_shape = {}, Y_sahpe = {}'.format(X.shape, Y.shape))
    train(X, Y, features=fea)
    train_end = time.time()
    log_.info('rov model train end, execute time = {}ms'.format((train_end - train_start)*1000))

    log_.info('rov model predict start...')
    predict_start = time.time()
    if env in ['dev', 'test']:
        predict_test()
    elif env in ['pre', 'pro']:
        predict()
    else:
        log_.error('env error')
    predict_end = time.time()
    log_.info('rov model predict end, execute time = {}ms'.format((predict_end - predict_start)*1000))