jiangcang_vj/main_pe.py

import os
import torch
import joblib
import pandas as pd
import numpy as np
import pickle
import time
from datetime import datetime, timedelta
from config import mongodb_config, vj_flight_route_list_hot, vj_flight_route_list_nothot, CLEAN_VJ_HOT_NEAR_INFO_TAB, CLEAN_VJ_NOTHOT_NEAR_INFO_TAB
from data_loader import mongo_con_parse, load_train_data
from data_preprocess import preprocess_data, standardization
from utils import chunk_list_with_index, create_fixed_length_sequences
from model import PriceDropClassifiTransModel
from predict import predict_future_distribute
from main_tr import features, categorical_features, target_vars


output_dir = "./data_shards"
photo_dir = "./photo"


def initialize_model():
    input_size = len(features)
    model = PriceDropClassifiTransModel(input_size, num_periods=2, hidden_size=64, num_layers=3, output_size=1, dropout=0.2)
    device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
    model.to(device)

    print(f"模型已初始化，输入尺寸：{input_size}")
    return model, device

def convert_date_format(date_str):
    """将 '2025-09-19 19:35:00' 转换为 '20250919193500' 格式"""
    dt = datetime.strptime(date_str, '%Y-%m-%d %H:%M:%S')
    return dt
    # return dt.strftime('%Y%m%d%H%M00')

def start_predict():

    # 确保目录存在
    os.makedirs(output_dir, exist_ok=True) 
    os.makedirs(photo_dir, exist_ok=True)

    # 清空上一次预测结果
    csv_file_list = ['future_predictions.csv']

    for csv_file in csv_file_list:
        try:
            csv_path = os.path.join(output_dir, csv_file)
            os.remove(csv_path)
        except Exception as e:
            print(f"remove {csv_path} error: {str(e)}")

    model, _ = initialize_model()

    date_end = (datetime.today() + timedelta(days=2)).strftime("%Y-%m-%d")
    date_begin = (datetime.today() + timedelta(days=2)).strftime("%Y-%m-%d")

    # 加载 scaler 列表
    feature_scaler_path = os.path.join(output_dir, 'feature_scalers.joblib')
    # target_scaler_path = os.path.join(output_dir, 'target_scalers.joblib')
    feature_scaler_list = joblib.load(feature_scaler_path)
    # target_scaler_list = joblib.load(target_scaler_path) 

    # 加载训练时保存的航班列表顺序
    with open(os.path.join(output_dir, f'order.pkl'), "rb") as f:
        flight_route_list = pickle.load(f)   

    flight_route_list_len = len(flight_route_list)
    route_len_hot = len(vj_flight_route_list_hot)
    route_len_nothot = len(vj_flight_route_list_nothot)
    
    assemble_size = 1           # 几个batch作为一个集群assemble
    current_assembled = -1      # 当前已加载的assemble索引
    group_size = 1              # 每几组作为一个批次

    chunks = chunk_list_with_index(flight_route_list, group_size)

    # 如果从中途某个批次预测, 修改起始索引
    resume_chunk_idx = 0
    chunks = chunks[resume_chunk_idx:]

    batch_starts = [start_idx for start_idx, _ in chunks]
    print(f"预测阶段起始索引顺序：{batch_starts}")

    # 测试阶段
    for i, (_, group_route_list) in enumerate(chunks, start=resume_chunk_idx):
        # 特殊处理，跳过不好的批次
        # client, db = mongo_con_parse()
        print(f"第 {i} 组 :", group_route_list)
        # batch_flight_routes = group_route_list

        # 根据索引位置决定是 热门 还是 冷门
        if 0 <= i < route_len_hot:
            is_hot = 1
            table_name = CLEAN_VJ_HOT_NEAR_INFO_TAB
        elif route_len_hot <= i < route_len_hot + route_len_nothot:
            is_hot = 0
            table_name = CLEAN_VJ_NOTHOT_NEAR_INFO_TAB
        else:
            print(f"无法确定热门还是冷门, 跳过此批次。")
            continue
        
        # 加载测试数据 (仅仅是时间段取到后天)
        start_time = time.time()
        df_test = load_train_data(mongodb_config, group_route_list, table_name, date_begin, date_end, output_dir, is_hot)
        end_time = time.time()
        run_time = round(end_time - start_time, 3)
        print(f"用时: {run_time} 秒")

        # client.close()

        if df_test.empty:
            print(f"测试数据为空，跳过此批次。")
            continue

        # 数据预处理
        df_test_inputs = preprocess_data(df_test, features, categorical_features, is_training=False)
        
        total_rows = df_test_inputs.shape[0]
        print(f"行数: {total_rows}")
        if total_rows == 0:
            print(f"预处理后的测试数据为空，跳过此批次。")
            continue

        # 找对应的特征缩放文件 
        batch_idx = i
        print("batch_idx:", batch_idx)
        feature_scaler = feature_scaler_list[batch_idx]
        if feature_scaler is None:
            print(f"批次{batch_idx}没有找到特征标准化缩放文件")
            continue

        # 标准化与归一化处理
        df_test_inputs, feature_scaler, _ = standardization(df_test_inputs, feature_scaler, is_training=False)

        print("标准化后数据样本:\n", df_test_inputs.head())

        # 生成序列
        sequences, _, group_ids = create_fixed_length_sequences(df_test_inputs, features, target_vars, is_train=False)
        print(f"序列数量：{len(sequences)}")

        #----- 新增：智能模型加载 -----#
        assemble_idx = batch_idx // assemble_size  # 计算当前集群索引
        print("assemble_idx:", assemble_idx)
        if assemble_idx != current_assembled:
            # 从文件加载并缓存
            model_path = os.path.join(output_dir, f'best_model_as_{assemble_idx}.pth')
            if os.path.exists(model_path):
                state_dict = torch.load(model_path)
                model.load_state_dict(state_dict)
                current_assembled = assemble_idx
                print(f"从文件加载并缓存 assemble {assemble_idx} 的模型参数")
            else:
                print(f"未找到 assemble {assemble_idx} 的模型文件，跳过")
                continue
        else:
            # 同一assemble直接使用已加载参数
            print(f"复用 assemble {assemble_idx} 的已加载模型参数")

        target_scaler = None
        # 预测未来数据
        predict_future_distribute(model, sequences, group_ids, target_scaler=target_scaler, output_dir=output_dir)

    print("所有批次的预测结束")
    # 所有批次的预测结束后, 统一过滤处理
    # csv_file = 'future_predictions.csv'
    # csv_path = os.path.join(output_dir, csv_file)

    # # 汇总预测结果
    # try:
    #     df_predict = pd.read_csv(csv_path)
    # except Exception as e:
    #     print(f"read {csv_path} error: {str(e)}")
    #     df_predict = None

    # 后续的处理
    pass


if __name__ == "__main__":
    start_predict()