Pytorch实现transformer并对时间序列预测_transformer 时间序列预测 pytorch

利用Pytorch框架实现用transformer通过输入sine预测cos，代码如下：

1. 网络结构

import math
import torch
from torch import nn


class PositionalEncoding(nn.Module):
    def __init__(self, d_model, dropout=0.1, max_len=5000):
        super(PositionalEncoding, self).__init__()
        self.dropout = nn.Dropout(p=dropout)

        pe = torch.zeros(max_len, d_model)
        position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)

        div_term = torch.exp(torch.arange(0, d_model, 2).float() * (-math.log(10000.0) / d_model))

        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)

        pe = pe.unsqueeze(0).transpose(0, 1)

        self.register_buffer('pe', pe)

    def forward(self, x):
        x = x + self.pe[:x.size(1), :].squeeze(1)
        return self.dropout(x)


class TransformerTS(nn.Module):
    def __init__(self,
                 input_dim,
                 dec_seq_len,
                 out_seq_len,
                 d_model=512,
                 nhead=8,
                 num_encoder_layers=6,
                 num_decoder_layers=6,
                 dim_feedforward=2048
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36