Python model training: LSTM time series sales forecast (training, saving, calling)

LSTM (long short-term memory) long short-term memory network, the specific theory will not be described one by one, just start directly

1. Data import

• Normal pandas reads the data and converts the time column into an index (see other tutorials to do this, it feels useless, just follow the chronological order)

# 获取数据
import pandas as pd
from datetime import datetime
dataset = pd.read_csv('../data.csv', index_col='时间', usecols=[0,2,3,5], date_parser=lambda x:datetime.strptime(x, '%Y年%m月'))
dataset

2. Data normalization

• Reduce the data to the range of 0-1. Here I group all the data into one column, so that the narrowing range is the same. You can use this to convert directly later

# 数据归一化
from sklearn.preprocessing import MinMaxScaler
values = dataset.values
# 转换成一列
values_res = values.reshape(values.shape[0] * values.shape[1], 1)
scaler = MinMaxScaler(feature_range=(0, 1))
# 训练 scaler
scaled = scaler.fit_transform(values_res)
# 再转换成原来的样子
scaled_dataset = scaled.reshape(values.shape)
scaled_dataset

3. Divide training set and test set

• The data needs to be in chronological order, so here we cut 20% before and after

# 切分训练集和测试集
split = round(len(scaled_dataset)*0.20)
train = scaled_dataset[:-split]
test = scaled_dataset[-split:]
test

4. Divide tags and attributes

• The first column of data is label data, and the second and third columns are attribute condition data

# 划分标签和属性
train_x, train_y = train[:, 1:], train[:, 0]
test_x, test_y = test[:, 1:], test[:, 0]
test_x

5. Convert to LSTM input format

• Convert to the input format of the LSTM model (samples, timesteps, features)

train_x_input = train_x.reshape((train_x.shape[0], 1, train_x.shape[1]))
test_x_input = test_x.reshape((test_x.shape[0], 1, test_x.shape[1]))
test_x_input

6. Design LSTM model

• There are two ways to design an LSTM model. The first is to know what the best parameters are, and the second is to input a few more parameters, and then find the best parameters

6.1 Direct Modeling

# 设计 LSTM 模型
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import LSTM, Dense
model = Sequential()
model.add(LSTM(50, input_shape=(1, 2)))
model.add(Dense(1))
model.compile(loss="mae", optimizer="adam")
model.fit(train_x_input, train_y, epochs=10, batch_size=1, validation_data=(test_x_input, test_y), verbose=2, shuffle=False)

6.2 Find the best

from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import LSTM, Dense, Dropout
from keras.wrappers.scikit_learn import KerasRegressor
from sklearn.model_selection import GridSearchCV

def build_model(optimizer):
    grid_model = Sequential()
    grid_model.add(LSTM(50,return_sequences=True,input_shape=(1,2)))
    grid_model.add(LSTM(50))
    grid_model.add(Dropout(0.2))
    grid_model.add(Dense(1))
    grid_model.compile(loss = 'mse',optimizer = optimizer)
    return grid_model

grid_model = KerasRegressor(build_fn=build_model,verbose=1,validation_data=(test_x_input,test_y))
# 把各种可能的参数都丢上去
parameters = {
    
    'batch_size' : [1],
            'epochs' : [10,11],
            'optimizer' : ['adam', 'rmsprop'] } 
grid_search = GridSearchCV(estimator = grid_model,
                          param_grid = parameters,
                          cv = 2)
# 训练
grid_search = grid_search.fit(train_x_input, train_y)
# 最好的参数
print(grid_search.best_params_)
# 最好参数对应的模型
model = grid_search.best_estimator_.model

7. Testing and Graphical Display

from matplotlib import pyplot as plt
from sklearn.metrics import mean_squared_error
import math

# 测试
pred = model.predict(test_x_input)
# 获取原始值
real = scaler.inverse_transform(test_y.reshape(1, -1)).reshape(-1, 1)
predicted = scaler.inverse_transform(pred)
plt.plot(real, color = 'red', label = 'Real')
plt.plot(predicted, color = 'blue', label = 'Predicted')
plt.title('Sale Prediction')
plt.xlabel('Time')
plt.ylabel('Sale')
plt.legend()
plt.show()
rmse = math.sqrt(mean_squared_error(real, predicted))
print("均方根误差：" + str(rmse))

Root mean square error: 2.1375958318221455

8. Save the model to pkl file

# 保存模型
import dill
with open('./sale_predict_model.pkl', 'wb') as outfile:
    dill.dump({
    
    
        'scaler': scaler,
        'model': model
    }, outfile)

9. Model call

• If the model needs to be deployed online for invocation, you can directly write a script for invocation. At the same time, considering that the model must be read once for each invocation, which wastes performance, the parameters are passed directly in the form of Socket, and a resident service is formed in the background.
  • Socket fixed incoming format "a,b"

9.1 Python model caller

import socket
import threading
import numpy as np
import pickle

# Socket 操作
sk = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
sk.bind(('127.0.0.1', 10001))
sk.listen(5)
count = 0
# 读取模型
file = 'sale_predict_model.pkl'
with open(file, 'rb') as f:
    model = pickle.load(f)


# 模型预测
def predict(a, b):
    data = np.array([[a, b]])
    # 转换格式，使用的是模型训练时训练出来的编译器
    data_scaled = model['scaler'].transform(data.reshape(data.shape[0] * data.shape[1], 1)).reshape(data.shape)
    # 直接导入模型，一样要进行转换格式
    data_res = model['model'].predict(data_scaled.reshape((data_scaled.shape[0], 1, data_scaled.shape[1])))
    # 返回最终结果
    return model['scaler'].inverse_transform(data_res)[0][0]


# 处理 Socket 连接
def tcp(sock, addr):
    try:
        print('Accept new connection from %s:%s...' % addr)
        print('Request count: %d' % count)
        # 读取参数
        data = sock.recv(1024)
        # 解码参数
        data_str = data.decode('utf-8')
        print("Param: %s" % data_str)
        # 切割参数
        data_list = data_str.split(',')
        # 判断参数合法性
        if len(data_list) == 2:
	        # 合法参数调用模型并返回数据
            sock.send(str(predict(data_list[0], data_list[1])).encode('utf-8'))
            print("Invoke success")
        else:
            sock.send(('Error param: %s' % data_str).encode('utf-8'))
            print('Error param: %s' % data_str)
    except Exception as e:
        print('Except:', e)
        sock.send('Invoke error'.encode('utf-8'))
    finally:
        sock.close()


if __name__ == '__main__':
    while True:
    	# 监听连接
        data, addr = sk.accept()
        count += 1
        # 交给线程处理
        thread = threading.Thread(target=tcp, args=(data, addr))
        # 启动线程
        thread.start()

9.2 Java program caller

package org.example.service;

import java.io.IOException;
import java.net.Socket;
import java.nio.charset.StandardCharsets;

public class InvokeModel {
    
    
	// service 测试
    public static void main(String[] args){
    
    
        System.out.println(invoke(54.4, 14.4));
    }
	// service 调用方法
    public static String invoke(Double sale1, Double sale2) {
    
    
    	// 拼装参数
        String req = sale1 + "," + sale2;
        Socket socket = null;
        try {
    
    
        	// 创建 Socket
            socket = new Socket("127.0.0.1", 10001);
            // 传输数据
            socket.getOutputStream().write(req.getBytes(StandardCharsets.UTF_8));
            System.out.println("Request param: " + req);
            byte[] buf = new byte[256];
            // 读取返回的数据
            int len = socket.getInputStream().read(buf);
            // 返回最终的结果（是一个 Double，方便操作直接用 String）
            return new String(buf, 0, len);
        } catch (IOException e) {
    
    
            throw new RuntimeException(e);
        } finally {
    
    
            try {
    
    
                if (socket != null)
                    socket.close();
            } catch (IOException e) {
    
    
                System.err.println("Invoke model error");
            }
        }
    }

}