源码或数据集请点赞关注收藏后评论区留言或者私信博主要
由于独特的设计结构 LSTM适合于处理和预测时间序列中间隔和延迟非常长的重要事件。
LSTM是一种含有LSTM区块（blocks）或其他的一种类神经网络，文献或其他资料中LSTM区块可能被描述成智能网络单元，因为它可以记忆不定时间长度的数值，区块中有一个gate能够决定input是否重要到能被记住及能不能被输出output
LSTM有很多个版本，其中一个重要的版本是GRU（Gated Recurrent Unit），根据谷歌的测试表明，LSTM中最重要的是Forget gate，其次是Input gate，最次是Output gate。
介绍完LSTM的基本内容 接下来实战通过LSTM来预测股市收盘价格
先上结果
1：随着训练次数增加损失函数的图像如下 可以看出基本符合肘部方法 但是局部会产生突变
2：预测结果如下 红色的是预测值 蓝色的是真实值 可以看出除了某几个极值点正确率较高
代码如下
import pandas as pdimport numpy as npimport matplotlib.pyplot as pltimport torchimport torch.nn as nnfrom torch.utils.data import DataLoader, Datasetimport osos.environ["KMP_DUPLICATE_LIB_OK"]="TRUE"def read_dataset(dataset_type):assert dataset_type == "train" or dataset_type == "test"df = pd.read_csv(stock_market_price_{}.csv'.format(dataset_type))# 读入股票数据data = np.array(df['close'])# 获取收盘价序列data = data[::-1]# 反转，使数据按照日期先后顺序排列normalize_data = (data - np.mean(data)) / np.std(data)# 标准化normalize_data = normalize_data[:, np.newaxis]# 增加维度X, y = [], []for i in range(len(normalize_data) - time_step):_x = normalize_data[i:i + time_step]_y = normalize_data[i + time_step]X.append(_x.tolist())y.append(_y.tolist())# plt.figure()# plt.plot(data)# plt.show() # 以折线图展示datareturn X, y# 实验参数设置time_step = 7# 用前七天的数据预测第八天hidden_size = 4# 隐藏层维度lstm_layers = 1# 网络层数batch_size = 64# 每一批次训练多少个样例input_size = 1 # 输入层维度output_size = 1# 输出层维度lr = 0.05# 学习率class myDataset(Dataset):def __init__(self, x, y):self.x = xself.y = ydef __getitem__(self, index):return torch.Tensor(self.x[index]), torch.Tensor(self.y[index])def __len__(self):return len(self.x)class LSTM(nn.Module):def __init__(self, input_size, output_size, hidden_size, device):super(LSTM, self).__init__()self.input_size=input_sizeself.output_size=output_sizeself.hidden_size=hidden_sizeself.device=devicedef _one(a,b):return nn.Parameter(torch.FloatTensor(a,b).to(self.device))def _three():return(_one(input_size,hidden_size), _one(hidden_size,hidden_size), nn.Parameter(torch.zeros(hidden_size).to(self.device)))self.W_xi,self.W_hi,self.b_i=_three()self.W_xf, self.W_hf, self.b_f = _three()self.W_xo, self.W_ho, self.b_o = _three()self.W_xc, self.W_hc, self.b_c = _three()self.W_hq=_one(hidden_size,output_size)self.b_q=nn.Parameter(torch.zeros(output_size).to(self.device))self.params=[self.W_xi,self.W_hi,self.b_i,self.W_xf, self.W_hf, self.b_f, self.W_xo, self.W_ho, self.b_o,self.W_xc, self.W_hc, self.b_c, self.W_hq,self.b_q]for param in self.params:if param.dim()==2:nn.init.xavier_normal_(param)def init_lstm_state(self, batch_size):return (torch.zeros((batch_size, self.hidden_size), device=self.device),torch.zeros((batch_size, self.hidden_size), device=self.device))def forward(self, seq):(H,C)=self.init_lstm_state(seq.shape[0])for step in range(seq.shape[1]):X=seq[:,step,:]I=torch.sigmoid((X@self.W_xi)+(H@self.W_hi)+self.b_i)F = torch.sigmoid((X @ self.W_xf) + (H @ self.W_hf) + self.b_f)O = torch.sigmoid((X @ self.W_xo) + (H @ self.W_ho) + self.b_o)C_tilda=torch.tanh(torch.matmul(X.float(),self.W_xc)+torch.matmul(H.float(),self.W_hc)+self.b_c)C=F*C+I*C_tildaH=O*torch.tanh(C)Y=(H@self.W_hq)+self.b_qreturn Y,(H,C)X_train, y_train = read_dataset('train')X_test, y_test = read_dataset('test')train_dataset = myDataset(X_train, y_train)test_dataset = myDataset(X_test, y_test)train_loader = DataLoader(train_dataset, batch_size, shuffle=True)test_loader = DataLoader(test_dataset, 1)# 设定训练轮数num_epochs = 50device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu')hist = np.zeros(num_epochs)model = LSTM(input_size, output_size, hidden_size, device)# 定义优化器和损失函数optimiser = torch.optim.Adam(model.parameters(), lr=lr)# 使用Adam优化算法loss_func = torch.nn.MSELoss(reduction='mean')# 使用均方差作为损失函数for epoch in range(num_epochs):epoch_loss = 0for i, data in enumerate(train_loader):X, y = datapred_y, _ = model(X.to(device))loss = loss_func(pred_y, y.to(device))optimiser.zero_grad()loss.backward()optimiser.step()epoch_loss += loss.item()print("Epoch ", epoch, "MSE: ", epoch_loss)hist[epoch] = epoch_lossplt.plot(hist)plt.show()# 测试model.eval()result = []for i, data in enumerate(test_loader):X, y = datapred_y, _ = model(X.to(device))result.append(pred_y.item())plt.plot(range(len(y_test)), y_test, label="true_y", color="blue")plt.plot(range(len(result)), result, label="pred_y", color="red")plt.legend(loc='best')plt.show()