1 简要信息

循环神经网络的每一层不仅由上一层的输入决定，还由上一时刻当前层决定。故循环神经网络能够对时序信息有较好的拟合性。

可使用现成的循环神经网络pytorch.nn.RNN进行时序预测。
每一层的计算为： ${h_t} = \tanh ({x_t}W_{ih}^T + {b_{ih} } + {h_{t - 1} }W_{hh}^T + {b_{hh} })$ 。参考文档:pytorch.nn.RNN。

2 代码实现

2.1 导入

import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from matplotlib import pyplot as plt

2.2 神经网络

input_size = 1
hidden_size = 16
output_size = 1
forcast_length=30

class Net(nn.Module):

    def __init__(self, ):
        super(Net, self).__init__()

        self.rnn = nn.RNN(
            input_size=input_size,
            hidden_size=hidden_size,
            num_layers=1,
            batch_first=True,
        )
        for p in self.rnn.parameters():
          nn.init.normal_(p, mean=0.0, std=0.001)

        self.linear = nn.Linear(hidden_size, output_size)

    def forward(self, x, hidden_prev):

       out, hidden_prev = self.rnn(x, hidden_prev)
       # [b, seq, h]
       out = out.view(-1, hidden_size)
       out = self.linear(out)
       out = out.unsqueeze(dim=0)
       return out, hidden_prev

lr=0.0001
time_step=50

2.3 GPU训练

if(torch.cuda.is_available()):
    device='cuda'
    print("Cuda Apply!")
else:
    device='cpu'

2.4 学习

#net = nn.RNN(input_size=1, hidden_size=16,num_layers=1,batch_first=True)
net=Net()
net.to(device)
lossfunction = nn.MSELoss()
optimizer = optim.Adam(net.parameters(), lr)

hidden = torch.zeros(1,1, 16)
hidden=hidden.to(device)

for iter in range(10000):
    ranspot=np.random.randint((int)(forcast_length/5),size=1)[0]
    Xasix=np.linspace(ranspot,ranspot+forcast_length,time_step)
    data=np.sin(Xasix)
    data=data*np.exp(-Xasix/10)

    data.reshape(time_step,1)
    #数据集
    x=torch.tensor(data[:-1]).float().view(1,time_step-1,1)
    y=torch.tensor(data[1:]).float().view(1,time_step-1,1)
    x=x.to(device)
    y=y.to(device)

    output,hidden=net(x,hidden)

#    output = output.view(-1, 16)
#    aff=nn.Linear(16,1)
#    output = aff(output)
#    output = output.unsqueeze(dim=0)

    hidden=hidden.detach()
    loss=lossfunction(output,y)
    net.zero_grad()
    loss.backward()
    optimizer.step()

    if(iter%100 == 0):
        print("Step:{},loss:{}".format(iter,loss.item()))

2.5 验证

Xasix = np.linspace(0, 0 + forcast_length, time_step)
data=np.sin(Xasix)

data=data*np.exp(-Xasix/10)

data.reshape(time_step,1)
#数据集
x=torch.tensor(data[:-1]).float().view(1,time_step-1,1)
y=torch.tensor(data[1:]).float().view(1,time_step-1,1)
x=x.to(device)
y=y.to(device)

input=x
predict,hidden=net(input,hidden)


x=x.to('cpu')
y=y.to('cpu')
predict =predict.to('cpu')

x = x.data.numpy().ravel()
y = y.data.numpy().ravel()
predict = predict.data.numpy().ravel()

plt.plot(Xasix[:-1],x,label='Known')
plt.plot(Xasix[1:],y,label='Actual')
plt.plot(Xasix[1:],predict,label='Forecast')
plt.scatter(Xasix[time_step-2],x[time_step-2])
plt.scatter(Xasix[time_step-1],y[time_step-2])
plt.scatter(Xasix[time_step-1],predict[time_step-2])
plt.legend(loc=0)
#plt.plot(x.predict)
plt.show()

2.6 结果

最后一个预测数据与准确值相差不大。

3 思考与问题

3.1思考

example.一次预测可以使用0-49号的数据预测出2-50号的值，再将预测值输入神经网络预测3-51号数据，以此类推。

3.2 问题

RNN是否能实现概率预测？给出预测准确概率？

本文采用署名-非商业性使用-相同方式共享 4.0 国际许可协议，转载请注明出处。

【深度学习】循环神经网络笔记