| import math | |
| import torch | |
| import torch.nn as nn | |
| from torch import Tensor | |
| class PositionalEncoding(nn.Module): | |
| r""" | |
| Positional Encoding in "Attention Is All You Need" (section 3.5). | |
| "Attention Is All You Need" uses sine and cosine functions of different frequencies: | |
| PE_(pos, 2i) = sin(pos / power(10000, 2i / d_model)) | |
| PE_(pos, 2i+1) = cos(pos / power(10000, 2i / d_model)) | |
| only change is that calculations are done with -log(power(10000, 2i / d_model)) | |
| Uses OpenSpeech's PositionalEncoding, as I don't see the point in coding this from scratch. | |
| """ | |
| def __init__(self, d_model: int, dropout_p: float, max_length: int = 5000) -> None: | |
| super(PositionalEncoding, self).__init__() | |
| self.dropout = nn.Dropout(p=dropout_p) | |
| pe = torch.zeros(max_length, d_model, requires_grad=False) | |
| position = torch.arange(0, max_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) | |
| pe = pe.unsqueeze(0) | |
| self.register_buffer("pe", pe) | |
| def forward(self, x: Tensor) -> Tensor: | |
| x = x + (self.pe[:, :x.shape[1], :]).requires_grad_(False) | |
| return self.dropout(x) |