很多算法比赛经常会遇到不同的物体产生同含义的时间序列信息,比如不同位置的时间序列信息,风力发电、充电桩用电。经常会遇到该如此场景,对所有数据做统一处理喂给模型,模型很难学到区分信息,因此设计如果对不同位置的装置做嵌入操作,这也是本文书写的主要目的之一,如果对不同位置装置的时序数据做模型呢?
RGU: 循环神经网络模块,经常用于处理时序数据。
Embedding
: 是 PyTorch 中的一个类,用于将离散的整数序列映射为连续的向量表示。
使用下面比赛的数据作为一个处理的DEMO:
2023中国华录杯数据湖算法大赛
import package
import numpy as np
import pandas as pd
import torch
import torch.nn as nn
#import tushare as ts
from sklearn.preprocessing import StandardScaler, MinMaxScaler
from sklearn.model_selection import train_test_split
from torch.utils.data import TensorDataset
from tqdm import tqdm
from torch.utils.data import Dataset, DataLoader
from sklearn.preprocessing import LabelEncoder
import matplotlib.pyplot as plt
import tqdm
import sys
import os
import gc
import argparse
import warnings
warnings.filterwarnings('ignore')
load data
class Config():
#data_path = '../data/data1/train/power.csv'
timestep = 14 # 时间步长,就是利用多少时间窗口
batch_size = 32 # 批次大小
feature_size = 1 # 每个步长对应的特征数量,这里只使用1维,每天的风速
hidden_size = 56 # 隐层大小
output_size = 1 # 由于是单输出任务,最终输出层大小为1,预测未来1天风速
num_layers = 1 # lstm的层数
epochs = 10 # 迭代轮数
best_loss = 0 # 记录损失
learning_rate = 0.00003 # 学习率
model_name = 'lstm' # 模型名称
save_path = './{}.pth'.format(model_name) # 最优模型保存路径
config = Config()
train_df = pd.read_csv('../初赛数据/phase1_train.csv')
test_df = pd.read_csv('../初赛数据/phase1_test.csv')
labelEncoder = LabelEncoder()
train_df['line_label'] = labelEncoder.fit_transform(train_df['line'])
#labelEncoder.transform(test_df['line'])
train_df = train_df.sort_values(["line",'date']).reset_index(drop=True)
train_df.line.unique()
array(['L01', 'L02', 'L03', 'L04', 'L05', 'L06', 'L08', 'L09', 'L10'], dtype=object)
使用前面14天预测未来第七天:
1,2,3,4,5,6,7,8,9,10,11,12,13,14 -》14+7
【1,2,3,4,5,6,7,8,9,10,11,12,13,14】+1 -》 14+7+1
。。。。。
#train_df.head()
his_pow_feats = []
for i in range(config.timestep):
train_df[f'shift_{7+i}'] = train_df.groupby("line_label")['passenger_flow'].shift(7+i)
his_pow_feats.append(f'shift_{7+i}')
train_df_drop_na = train_df[train_df[his_pow_feats].isna().sum(axis=1)==0]
class MyDataSet(Dataset):
def __init__(self,train_df_drop_na,his_pow_feats):
"""
train_df_drop_na
"""
self.train_df = train_df_drop_na.reset_index(drop=True)
def __len__(self):
return len(self.train_df)
def __getitem__(self,item):
label = self.train_df.loc[item,'passenger_flow']
id_encoder = self.train_df.loc[item,'line_label']
his_feats_list = self.train_df.loc[item,his_pow_feats].values.tolist()
return {
"input_ids":torch.tensor(id_encoder,dtype=torch.long),
"his_feats":torch.as_tensor(his_feats_list ,dtype=torch.float32).unsqueeze(-1),
"labels":torch.tensor(label,dtype=torch.float32)}
RANDOM_SEED = 1023
df_train, df_test = train_test_split(train_df_drop_na, test_size=0.2, random_state=RANDOM_SEED)
df_val, df_test = train_test_split(df_test, test_size=0.5, random_state=RANDOM_SEED)
df_train.shape, df_val.shape, df_test.shape
def create_data_loader(train_df_drop_na,his_pow_feats,batch_size=32):
ds = MyDataSet(train_df_drop_na,
his_pow_feats
)
return DataLoader(ds,batch_size=batch_size)
BATCH_SIZE = 32
train_data_loader = create_data_loader(df_train,his_pow_feats=his_pow_feats,batch_size=BATCH_SIZE)
val_data_loader = create_data_loader(df_val, his_pow_feats=his_pow_feats,batch_size=BATCH_SIZE)
test_data_loader = create_data_loader(df_test,his_pow_feats=his_pow_feats,batch_size=BATCH_SIZE)
#train_df[cols]
# 7.定义LSTM网络
class GRUModel(nn.Module):
def __init__(self, feature_size, hidden_size, num_layers, output_size):
super(GRUModel, self).__init__()
self.hidden_size = hidden_size # 隐层大小
self.num_layers = num_layers # lstm层数
# feature_size为特征维度,就是每个时间点对应的特征数量,这里为1
self.gru = nn.GRU(feature_size, hidden_size, num_layers, batch_first=True,bidirectional=True)
self.layer_norm = nn.LayerNorm(hidden_size*2)
self.fc = nn.Linear(hidden_size*2+2, output_size)
self.embedding = nn.Embedding(9, 2)
def forward(self, x,id_label, hidden=None):
#print(x.shape)
batch_size = x.shape[0] # 获取批次大小 batch, time_stamp , feat_size
# 初始化隐层状态
h_0 = x.data.new(2*self.num_layers, batch_size, self.hidden_size).fill_(0).float()
if hidden is not None:
h_0 = hidden
#print(h_0.size)
# GRU 运算
output, hidden = self.gru(x,h_0)
output = self.layer_norm(output)
last_output = output[:, -1, :]
#print('output',last_output.shape)
embed = self.embedding(id_label)
#print("embed",embed.shape)
#print('output',output.shape)
concatenated = torch.cat((embed, last_output), dim=1)
#print(concatenated.shape)
# 全连接层
output = self.fc(concatenated) # 形状为batch_size * timestep, 1
#print(output.shape)
# 我们只需要返回最后一个时间片的数据即可
return output
model = GRUModel(config.feature_size, config.hidden_size, config.num_layers, config.output_size) # 定义LSTM网络
loss_function = nn.L1Loss() # 定义损失函数
# class MAPELoss(nn.Module):
# def __init__(self):
# super(MAPELoss, self).__init__()
# def forward(self, y_pred, y_true):
# epsilon = 1e-8 # 用于避免除以零的小常数
# absolute_error = torch.abs(y_true - y_pred)
# relative_error = absolute_error / (torch.abs(y_true) + epsilon)
# mape = torch.mean(relative_error) * 100
# return mape
# loss_function = MAPELoss() # 定义损失函数
optimizer = torch.optim.AdamW(model.parameters(), lr=0.01) # 定义优化器
from tqdm import tqdm
# 8.模型训练
for epoch in range(500):
model.train()
running_loss = 0
train_bar = tqdm(train_data_loader) # 形成进度条
for data in train_bar:
x_train, y_train = data['his_feats'], data['labels'] # 解包迭代器中的X和Y
optimizer.zero_grad()
y_train_pred = model(x_train,data['input_ids'])
loss = loss_function(y_train_pred, y_train.reshape(-1, 1))
loss.backward()
optimizer.step()
running_loss += loss.item()
train_bar.desc = "train epoch[{}/{}] loss:{:.3f}".format(epoch + 1,
config.epochs,
loss)
# 模型验证
model.eval()
test_loss = 0
with torch.no_grad():
test_bar = tqdm(val_data_loader)
for data in test_bar:
x_test, y_test = data['his_feats'], data['labels']
y_test_pred = model(x_test, data['input_ids'])
test_loss = loss_function(y_test_pred, y_test.reshape(-1, 1))
if test_loss < config.best_loss:
config.best_loss = test_loss
torch.save(model.state_dict(), save_path)
print('Finished Training')