| import math |
|
|
| import torch |
| from torch import nn |
|
|
|
|
| class PositionalEncoding(nn.Module): |
| def __init__(self, d_model, max_seq_length): |
| super(PositionalEncoding, self).__init__() |
|
|
| pe = torch.zeros(max_seq_length, d_model) |
| position = torch.arange(0, max_seq_length, 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) |
|
|
| self.register_buffer('pe', pe.unsqueeze(0)) |
|
|
| def forward(self, x): |
| return x + self.pe[:, :x.size(1)] |
