畅销作品！LSTM专栏！ | 一次订阅，永久更新

import numpy as npimport pandas as pdimport matplotlib.pyplot as pltimport seaborn as snsfrom sklearn.preprocessing import MinMaxScalerfrom sklearn.model_selection import train_test_splitimport torchimport torch.nn as nnimport timeimport math, timefrom sklearn.metrics import mean_squared_error,r2_scoreimport plotly.express as pximport plotly.graph_objects as gofrom utils import split_dataimport argparseimport osfrom model import *from utils import *from pathlib import Pathparser = argparse.ArgumentParser()

# dataparser.add_argument('--path', type=str, default='../data/rlData.csv')parser.add_argument('--vision', type=bool, default=True)parser.add_argument('--input_features', type=list, default=['Open', 'High', 'Low', 'Volume', 'Close'])parser.add_argument('--output_features', type=list, default=['Close','Open','High','Low'])parser.add_argument('--window_size', type=int, default=20)parser.add_argument('--train_test_ratio', type=float, default=0.2)parser.add_argument('--random_state', type=int, default=34)# modelparser.add_argument('--model_name', type=str, default='LSTM-ekan',help='selection from LSTM,LSTM-ekan')parser.add_argument('--dropout', type=float, default=0.2)parser.add_argument('--hidden_dim', type=int, default=32)
##开始时间统计和结果保存start_time = time.time()result = []
##循环训练for t in range(args.num_epochs):    y_train_pred = model(x_train)    ##损失计算    loss = criterion(y_train_pred, y_train)    ##R2计算    R2 = r2_score(y_train_pred.detach().numpy(), y_train.detach().numpy())    print("Epoch ", t, "MSE: ", loss.item(), 'R2', R2.item())
    # 统计每个时间步的损失和R2指标    MSE_hist[t] = loss.item()    if R2 < 0:        R2 = 0    R2_hist[t] = R2    optimiser.zero_grad()    loss.backward()    optimiser.step()
training_time = time.time() - start_timeprint("Training time: {}".format(training_time))## 反缩放-恢复正常值predict = pd.DataFrame(target_scaler.inverse_transform(y_train_pred.detach().numpy()))original = pd.DataFrame(target_scaler.inverse_transform(y_train.detach().numpy()))
##对每个预测单元进行绘图for i in range(len(args.output_features)):    sns.set_style("white")    # 解决画图中文显示问题    plt.rcParams['font.sans-serif'] = ['SimHei']    plt.rcParams['axes.unicode_minus'] = False    fig = plt.figure()    fig.subplots_adjust(hspace=0.2, wspace=0.2)    # 计算当前组的行数和列数
    ax = sns.lineplot(x=range(len(original.index)), y=original.loc[:, i], label=f"{param_index_dict[i]}",                      color='royalblue')    ax = sns.lineplot(x=range(len(predict.index)), y=predict.loc[:, i], label=f"{args.model_name} ",                      color='tomato')    # print(predict.index)    # print(predict[0])
    ax.set_title(f'{param_index_dict[i]}', size=14, fontweight='bold')    ax.set_xlabel("时间", size=14)    ax.set_ylabel("价格", size=14)    ax.set_xticklabels('', size=10)
    Fitting_path = f'../{args.model_name}/train'    if not os.path.exists(Fitting_path):        # 使用 os.makedirs() 创建目录        os.makedirs(Fitting_path)        print(f"Directory '{Fitting_path}' was created.")    else:        print(f"Directory '{Fitting_path}' already exists.")
    plt.savefig(os.path.join(Fitting_path, f'{param_index_dict[i]}'))##绘制loss曲线和R2决定系数indication_path=f'../{args.model_name}/train'plt.plot()ax = sns.lineplot(data=MSE_hist, color='royalblue')ax.set_xlabel("Epoch", size = 14)ax.set_ylabel("Loss", size = 14)ax.set_title("训练损失", size = 14, fontweight='bold')fig.set_figheight(6)fig.set_figwidth(16)plt.savefig(os.path.join(indication_path,'loss'))plt.show()
plt.plot()ax = sns.lineplot(data=R2_hist, color='red')ax.set_xlabel("Epoch", size = 14)ax.set_ylabel("R2", size = 14)ax.set_title("R2", size = 14, fontweight='bold')fig.set_figheight(6)fig.set_figwidth(16)plt.savefig(os.path.join(indication_path,'R2'))plt.show()y_test_pred = model(x_test)trainScore = mean_squared_error(y_train.detach().numpy(), y_train_pred.detach().numpy(),squared=False)r2_train = r2_score(y_train.detach().numpy(), y_train_pred.detach().numpy())print('Train Score: %.2f RMSE' % (trainScore))print('Train R^2: %.2f' % (r2_train))testScore = math.sqrt(mean_squared_error(y_test.detach().numpy(), y_test_pred.detach().numpy()))r2_test = r2_score(y_test.detach().numpy(), y_test_pred.detach().numpy())print('Test Score: %.2f RMSE' % (testScore))print('Test R^2: %.2f' % (r2_test))result.append(trainScore)result.append(testScore)result.append(training_time)## 反缩放-恢复正常值y_train_pred = target_scaler.inverse_transform(y_train_pred.detach().numpy())y_train = target_scaler.inverse_transform(y_train.detach().numpy())y_test_pred = target_scaler.inverse_transform(y_test_pred.detach().numpy())y_test = target_scaler.inverse_transform(y_test.detach().numpy())file_path = 'model_scores.csv'


# 检查文件是否存在if Path(file_path).exists():    # 文件存在，追加数据    df.to_csv(file_path, mode='a', header=False)else:    # 文件不存在，创建文件并写入数据    df.to_csv(file_path)for i in range(len(args.output_features)):    trainPredictPlot = np.empty_like(target_scaled[:,i]).reshape(-1,1)    trainPredictPlot[:, :] = np.nan    trainPredictPlot[args.window_size:len(y_train_pred)+args.window_size, :] = y_train_pred[:,i].reshape(-1,1)