|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
import torch |
|
|
|
|
|
|
|
|
class PositionalEncoding(torch.nn.Module): |
|
|
""" |
|
|
Sinusoidal positional encoding module for Transformer models. |
|
|
|
|
|
This module injects information about the relative or absolute position of |
|
|
tokens in a sequence by adding fixed sinusoidal embeddings to the input |
|
|
embeddings. The positional encodings are non-learnable and follow the |
|
|
formulation introduced in the original Transformer architecture. |
|
|
""" |
|
|
def __init__(self, emb_dim: int, max_len: int = 5000, **kwargs): |
|
|
""" |
|
|
Initialize the positional encoding module. |
|
|
|
|
|
Parameters |
|
|
---------- |
|
|
emb_dim : int |
|
|
Dimensionality of the embedding space. |
|
|
max_len : int, optional |
|
|
Maximum supported sequence length for which positional encodings |
|
|
are precomputed. |
|
|
""" |
|
|
super().__init__(**kwargs) |
|
|
|
|
|
|
|
|
pe = torch.zeros(max_len, emb_dim) |
|
|
position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1) |
|
|
div_term = torch.exp(torch.arange(0, emb_dim, 2).float() * -(torch.log(torch.tensor(10000.0)) / emb_dim)) |
|
|
pe[:, 0::2] = torch.sin(position * div_term) |
|
|
pe[:, 1::2] = torch.cos(position * div_term) |
|
|
pe = pe.unsqueeze(0) |
|
|
|
|
|
|
|
|
self.register_buffer('positional_encoding', pe) |
|
|
|
|
|
def forward(self, x: torch.Tensor) -> torch.Tensor: |
|
|
""" |
|
|
Add positional encodings to the input embeddings. |
|
|
|
|
|
Parameters |
|
|
---------- |
|
|
x : torch.Tensor |
|
|
Input tensor of shape (batch_size, sequence_length, emb_dim). |
|
|
|
|
|
Returns |
|
|
------- |
|
|
torch.Tensor |
|
|
Tensor of the same shape as the input with positional encodings added. |
|
|
""" |
|
|
return x + self.positional_encoding[:, :x.size(-2), :] |
|
|
|
|
|
|
|
|
|
|
|
|
