ai-station-code/guangfufadian/model_base.py

from guangfufadian.cross_exp.exp_crossformer import Exp_crossformer  # 根据实际路径导入
from guangfufadian.utils.tools import StandardScaler
import pickle,os,sys
import torch
import pandas as pd
import torch
import numpy as np
# 获取当前工作目录
current_directory = os.getcwd()
current_directory = os.path.join(current_directory,'guangfufadian')

# 设置参数
class guangfufadian_Args:
    # 在这里定义模型参数
    data_dim = 14
    in_len = 192
    out_len = 96
    seg_len = 6
    win_size = 2
    factor = 10
    d_model = 256
    d_ff = 512
    n_heads = 4
    e_layers = 3
    dropout = 0.2
    use_multi_gpu = False
    use_gpu = True
    device_ids = [0]  # 如果使用多 GPU，设置设备 ID
    batch_size = 32
    train_epochs = 100
    patience = 10
    num_workers = 4
    # root_path = '/home/xiazj/project_test/Crossformer-master/datasets/station08_utf8.csv'  # 数据路径
    data_path = os.path.join(current_directory,'datasets/station08_utf8.csv') # 数据文件名
    checkpoints = os.path.join(current_directory,'checkpoints')               # 模型保存路径
    learning_rate = 0.001  # 学习率
    data_split = 0.8  # 数据划分比例
    scale_statistic = pickle.load(open(os.path.join(checkpoints,'Crossformer_station08_il192_ol96_sl6_win2_fa10_dm256_nh4_el3_itr0/scale_statistic.pkl'), 'rb'))
    baseline = False
    gpu =  0

# 实例化参数
guangfufadian_args = guangfufadian_Args()


class ModelInference:
    def __init__(self, model_path, args):
        # 初始化模型
        self.model_experiment = Exp_crossformer(args)
        self.model_experiment.model.load_state_dict(torch.load(model_path))  # 加载模型参数
        self.model_experiment.model.eval()  # 设置模型为评估模式
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.model_experiment.model.to(self.device)
        self.scaler = StandardScaler(mean = args.scale_statistic['mean'], std = args.scale_statistic['std'])


    """
    加载推理数据
    如果数据列发生变化，或者列不匹配，报错
    数据长度满足192，截取前192
    数据长度不满足192，前拼接满足数量级
    """
    def load_pred_data(self,data):
        raw_data = data.values
        length = raw_data.shape[0]
        # 顺序，必须一致才能进行下一步
        check_list = ['date_time', 'nwp_globalirrad', 'nwp_directirrad', 'nwp_temperature',
        'nwp_humidity', 'nwp_windspeed', 'nwp_winddirection', 'nwp_pressure',
        'lmd_totalirrad', 'lmd_diffuseirrad', 'lmd_temperature', 'lmd_pressure',
        'lmd_winddirection', 'lmd_windspeed', 'power']
        columns_match = list(data.columns) == check_list
        if columns_match:
            if length > guangfufadian_args.in_len:
                input_data = raw_data[:guangfufadian_args.in_len, 1:].astype(np.float32)
                input_tensor = torch.tensor(input_data, dtype=torch.float32).unsqueeze(0)  # 增加batch维度
            else:
                rows_to_add = guangfufadian_args.in_len - length
                first_row = raw_data[0]
                new_rows = [first_row] * rows_to_add  # 复制第一行的值
                # 拼接新的行和原始数据
                input_data = new_rows + list(raw_data)
                input_data = np.array(input_data)
                input_data = input_data[:guangfufadian_args.in_len, 1:].astype(np.float32)
                input_tensor = torch.tensor(input_data, dtype=torch.float32).unsqueeze(0)  # 增加batch维度
            return {'status':True, 'reason':input_tensor }
        else:
            print("文件不匹配，请检查上传文件与模版是否一致")
            return {'status':False, 'reason':'文件不匹配，请检查上传文件与模版是否一致'}

    # 数据归一化
    def preprocess_data(self, input_tensor):
        input_tensor =  self.scaler.transform(input_tensor)
        return input_tensor.to(self.device)
    # 数据预测
    def predict(self, input_tensor):
        with torch.no_grad():  # 不计算梯度
            predictions = self.model_experiment._predict_batch(input_tensor)
        # return predictions.cpu().numpy()  # 返回 NumPy 数组
        return predictions
    # 推理过程
    def run_inference(self, data):
        raw_data = self.load_pred_data(data)
        if raw_data['status'] == False:
            return {'status':False, 'reason':'文件不匹配，请检查上传文件与模版是否一致'}
        else:
            input_tensor = self.preprocess_data(raw_data['reason'])
            predictions = self.predict(input_tensor)
            predictions = self.scaler.inverse_transform(predictions)
            predictions = predictions.squeeze(0)
            predictions = predictions.cpu().numpy()
            return {'status':True, 'reason':predictions[:,-1:]}