Pytorch实现多输入多输出LSTM神经网络(含代码)

由于做导师项目需要，学习了一段时间LSTM神经网络。主要根据的是官网进行学习的。

https://pytorch.org/docs/stable/generated/torch.nn.LSTM.html#torch.nn.LSTM

里面告诉了我们torch里面的LSTM具体输入参数是什么，这告诉了我们怎么去准备序列数据。

输入数据

LSTM属于是特殊的一种循环神经网络，准备的数据是一种序列数据。因此我们处理数据的参数需要序列长度，以及确认是否单输入单输出。处理数据的代码如下:

def data_2_seq(data, length, single=True, feature_index=[0,1], out_index = [2]):
    res = []
    if single:
        l = len(data)
        for i in range(l-length):
            seq = data[i: i+length]
            label = data[i+length: i+1+length]
            res.append((seq, label))
        return res
    else:
        l = data.size(1)
        for i in range(l-length):
            seq = data[feature_index,i: i+length]
            label = data[out_index,i+length: i+1+length]
            res.append((seq, label))
        return res

里面输入为序列，每一个序列对应一个标签，一般是序列的下一个数据。

模型定义

下面定义一个简单的网络，主要就是全连接层和LSTM层。

import torch
import torch.nn as nn 


class LSTM1(nn.Module):
    def __init__(self, inputsize = 1, hiddenlayer = 100, outputsize = 1):
        super().__init__() 
        self.hiddenlayer = hiddenlayer
        self.lstm = nn.LSTM(inputsize, hiddenlayer, batch_first=False) # batch_first=True，输入的数据为Batch*seq_len*inputsize;batch_first=False,输入的数据为seq__len*Batch*inputsize
        self.liner = nn.Linear(hiddenlayer, outputsize)
    
    def forward(self, seq):
        lstm_out, (hn, cn) = self.lstm(seq.contiguous().view(seq.size()[1], 1, -1))
        predictions = self.liner(lstm_out.view(seq.size()[1], -1))
        return predictions[-1]

定义好模型后，就是训练函数了。

def train(data, net, epochs, opt, loss_func):
    loss_hist = []
    for epoch in range(epochs):
        for seq, label in data:
            label = label.reshape(len(label))
            opt.zero_grad()
            y_pre = net(seq)
            loss = loss_func(y_pre, label)
            loss.backward()
            opt.step()
        print("epoch ",epoch+1,": ",loss.item())
        loss_hist.append(loss.item())
    return net, loss_hist

然后上述各个模块整定好了之后，进入main.py进行调用即可。

from DataProcess import data_2_seq
from model import LSTM1
from train import train
import pandas as pd
import numpy as np 
import torch
import torch.nn as nn
import matplotlib.pyplot as plt

all_data = pd.read_csv("flights.csv")
data = all_data['passengers'].values.astype(float)
train_data = data[:-12]
train_data = np.vstack((train_data, train_data))
train_data = train_data.reshape(2, -1)
train_data = torch.tensor(train_data, dtype=torch.float32)
train_seq = data_2_seq(train_data, 12, single=False, feature_index=[0,1], out_index=[0,1])

net = LSTM1(inputsize=2, hiddenlayer=100, outputsize=2)
lr = 1e-3
optim = torch.optim.Adam(net.parameters(),lr = lr)
epochs = 5
loss_func = nn.MSELoss()
trained_model, loss_hist = train(train_seq, net, epochs, optim, loss_func)
torch.save(trained_model, "/home/code/lstm-free/a.pth")

# plt.plot(loss_hist)
# plt.show()

以上就是我使用Pytorch实现多维输入输出的全过程，完整的项目可以在我的仓库下载使用。

https://github.com/Roy-fyq/LSTM-with-multi-dimensional-input-and-output