# 7.1.1 使用Pytorch搭建

```python
import torch
import torch.nn as nn

from sklearn.preprocessing import MinMaxScaler
import time
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
%matplotlib inline

from pandas.plotting import register_matplotlib_converters
register_matplotlib_converters()
```

**数据导入和准备**

```python
# 导入酒精销售数据
df = pd.read_csv('data\Alcohol_Sales.csv',index_col=0,parse_dates=True)
len(df)
```

```python
df.head()  # 观察数据集，这是一个单变量时间序列
```

```python
plt.figure(figsize=(12,4))
plt.grid(True)
plt.plot(df['S4248SM144NCEN'])
plt.show()
```

![](https://3993849477-files.gitbook.io/~/files/v0/b/gitbook-legacy-files/o/assets%2F-Mhv1ams1lvf_fMUn_is%2F-MiFh_8zxcPT2WmQwTZR%2F-MiFhcenBHjzyE4BA0Ia%2F8_2.png?alt=media\&token=efe70fd3-7129-41ee-bb3f-02fa23ce75c9)

```python
y = df['S4248SM144NCEN'].values.astype(float)

test_size = 12

# 划分训练和测试集，最后12个值作为测试集
train_set = y[:-test_size]
test_set = y[-test_size:]
```

```python
# 归一化至[-1,1]区间，为了获得更好的训练效果
scaler = MinMaxScaler(feature_range=(-1, 1))
train_norm = scaler.fit_transform(train_set.reshape(-1, 1))
```

**创建时间序列训练集**

```python
# 转换成 tensor
train_norm = torch.FloatTensor(train_norm).view(-1)

# 定义时间窗口，注意和前面的test size不是一个概念
window_size = 12

# 这个函数的目的是为了从原时间序列中抽取出训练样本，也就是用第一个值到第十二个值作为X输入，预测第十三个值作为y输出，这是一个用于训练的数据点，时间窗口向后滑动以此类推
def input_data(seq,ws):  
    out = []
    L = len(seq)
    for i in range(L-ws):
        window = seq[i:i+ws]
        label = seq[i+ws:i+ws+1]
        out.append((window,label))
    return out


train_data = input_data(train_norm,window_size)
len(train_data)  # 等于325（原始数据集长度）-12（测试集长度）-12（时间窗口）
```

```python
# 观察某一个数据点
train_data[0]
```

```
(tensor([-0.9268, -0.9270, -0.8340, -0.7379, -0.7966, -0.7439, -0.7547, -0.8109,
         -0.8128, -0.7901, -0.7933, -0.6743]),
 tensor([-1.]))
```

**定义模型**

```python
class LSTMnetwork(nn.Module):
    def __init__(self,input_size=1,hidden_size=100,output_size=1):
        super().__init__()
        self.hidden_size = hidden_size

        # 定义LSTM层
        self.lstm = nn.LSTM(input_size,hidden_size)

        # 定义全连接层
        self.linear = nn.Linear(hidden_size,output_size)

        # 初始化h0，c0
        self.hidden = (torch.zeros(1,1,self.hidden_size),
                       torch.zeros(1,1,self.hidden_size))

    def forward(self,seq):
        # 前向传播的过程是输入->LSTM层->全连接层->输出

        # https://pytorch.org/docs/stable/generated/torch.nn.LSTM.html?highlight=lstm#torch.nn.LSTM
        # 在观察查看LSTM输入的维度，LSTM的第一个输入input_size维度是(L, N, H_in), L是序列长度，N是batch size，H_in是输入尺寸，也就是变量个数
        # LSTM的第二个输入是一个元组，包含了h0,c0两个元素，这两个元素的维度都是（D∗num_layers,N,H_out)，D=1表示单向网络，num_layers表示多少个LSTM层叠加，N是batch size，H_out表示隐层神经元个数
        lstm_out, self.hidden = self.lstm(seq.view(len(seq),1,-1), self.hidden)

        pred = self.linear(lstm_out.view(len(seq),-1)) 
        return pred[-1]  # 输出只用取最后一个值
```

**定义损失函数和优化函数**

```python
torch.manual_seed(101)
model = LSTMnetwork()

criterion = nn.MSELoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)
```

```python
# 看下这个网络结构总共有多少个参数
def count_parameters(model):
    params = [p.numel() for p in model.parameters() if p.requires_grad]
    for item in params:
        print(f'{item:>6}')
    print(f'______\n{sum(params):>6}')

count_parameters(model)
```

```
   400
 40000
   400
   400
   100
     1
______
 41301
```

**训练模型**

```python
epochs = 100

start_time = time.time()

for epoch in range(epochs):

    for seq, y_train in train_data:

        # 每次更新参数前都梯度归零和初始化
        optimizer.zero_grad()
        model.hidden = (torch.zeros(1,1,model.hidden_size),
                        torch.zeros(1,1,model.hidden_size))

        y_pred = model(seq)

        loss = criterion(y_pred, y_train)
        loss.backward()
        optimizer.step()

    print(f'Epoch: {epoch+1:2} Loss: {loss.item():10.8f}')

print(f'\nDuration: {time.time() - start_time:.0f} seconds')
```

**模型预测和评估**

```python
future = 12

# 选取序列最后12个值开始预测
preds = train_norm[-window_size:].tolist()

# 设置成eval模式
model.eval()

# 循环的每一步表示向时间序列向后滑动一格
for i in range(future):
    seq = torch.FloatTensor(preds[-window_size:])
    with torch.no_grad():
        model.hidden = (torch.zeros(1,1,model.hidden_size),
                        torch.zeros(1,1,model.hidden_size))
        preds.append(model(seq).item())
```

```python
# 逆归一化还原真实值
true_predictions = scaler.inverse_transform(np.array(preds[window_size:]).reshape(-1, 1))
```

```python
# 对比真实值和预测值
plt.figure(figsize=(12,4))
plt.grid(True)
plt.plot(df['S4248SM144NCEN'])
x = np.arange('2018-02-01', '2019-02-01', dtype='datetime64[M]').astype('datetime64[D]')
plt.plot(x,true_predictions)
plt.show()
```

![](https://3993849477-files.gitbook.io/~/files/v0/b/gitbook-legacy-files/o/assets%2F-Mhv1ams1lvf_fMUn_is%2F-MiFh_8zxcPT2WmQwTZR%2F-MiFhdiMtoN1IqWjCoV7%2F8_3.png?alt=media\&token=99f0fe72-a545-4046-960f-3f7865cd9839)

```python
# 放大看
fig = plt.figure(figsize=(12,4))
plt.grid(True)
fig.autofmt_xdate()

plt.plot(df['S4248SM144NCEN']['2017-01-01':])
plt.plot(x,true_predictions)
plt.show()
```

![](https://3993849477-files.gitbook.io/~/files/v0/b/gitbook-legacy-files/o/assets%2F-Mhv1ams1lvf_fMUn_is%2F-MiFh_8zxcPT2WmQwTZR%2F-MiFheiL55wjOsCbRxQR%2F8_4.png?alt=media\&token=b048f29f-7aca-4956-877c-546623ffd943)

**预测未来的时间序列值**

```python
# 重新开始训练
epochs = 100

# 切回到训练模式
model.train()

y_norm = scaler.fit_transform(y.reshape(-1, 1))
y_norm = torch.FloatTensor(y_norm).view(-1)
all_data = input_data(y_norm,window_size)


start_time = time.time()

for epoch in range(epochs):

    for seq, y_train in all_data:  

        optimizer.zero_grad()
        model.hidden = (torch.zeros(1,1,model.hidden_size),
                        torch.zeros(1,1,model.hidden_size))

        y_pred = model(seq)

        loss = criterion(y_pred, y_train)
        loss.backward()
        optimizer.step()

    print(f'Epoch: {epoch+1:2} Loss: {loss.item():10.8f}')

print(f'\nDuration: {time.time() - start_time:.0f} seconds')
```

```python
# 重新预测
window_size = 12
future = 12
L = len(y)

preds = y_norm[-window_size:].tolist()

model.eval()
for i in range(future):  
    seq = torch.FloatTensor(preds[-window_size:])
    with torch.no_grad():

        model.hidden = (torch.zeros(1,1,model.hidden_size),
                        torch.zeros(1,1,model.hidden_size))  
        preds.append(model(seq).item())


true_predictions = scaler.inverse_transform(np.array(preds).reshape(-1, 1))


x = np.arange('2019-02-01', '2020-02-01', dtype='datetime64[M]').astype('datetime64[D]')

plt.figure(figsize=(12,4))
plt.grid(True)
plt.plot(df['S4248SM144NCEN'])
plt.plot(x,true_predictions[window_size:])
plt.show()
```

![](https://3993849477-files.gitbook.io/~/files/v0/b/gitbook-legacy-files/o/assets%2F-Mhv1ams1lvf_fMUn_is%2F-MiFh_8zxcPT2WmQwTZR%2F-MiFhfxoIs72dexzaZBH%2F8_5.png?alt=media\&token=ec9aa658-2ede-4272-9984-dbd25fab63ed)

```python
fig = plt.figure(figsize=(12,4))
plt.grid(True)
fig.autofmt_xdate()
plt.plot(df['S4248SM144NCEN']['2017-01-01':])
plt.plot(x,true_predictions[window_size:])
plt.show()
```

![](https://3993849477-files.gitbook.io/~/files/v0/b/gitbook-legacy-files/o/assets%2F-Mhv1ams1lvf_fMUn_is%2F-MiFh_8zxcPT2WmQwTZR%2F-MiFhh2nm7adrQlooNqt%2F8_6.png?alt=media\&token=2841aa18-b811-4cd7-ba84-31fe31de989b)