|
|
|
import math |
|
|
|
import torch |
|
from torch import nn |
|
|
|
|
|
class TokenEmbedding(nn.Module): |
|
def __init__( |
|
self, |
|
embedding_dim: int, |
|
vocab_size: int, |
|
dropout: float = 0.0, |
|
): |
|
super().__init__() |
|
|
|
self.vocab_size = vocab_size |
|
self.embedding_dim = embedding_dim |
|
|
|
self.dropout = torch.nn.Dropout(p=dropout) |
|
self.word_embeddings = nn.Embedding(self.vocab_size, self.embedding_dim) |
|
|
|
@property |
|
def weight(self) -> torch.Tensor: |
|
return self.word_embeddings.weight |
|
|
|
def embedding(self, index: int) -> torch.Tensor: |
|
return self.word_embeddings.weight[index : index + 1] |
|
|
|
def forward(self, x: torch.Tensor): |
|
x = self.word_embeddings(x) |
|
x = self.dropout(x) |
|
return x |
|
|
|
|
|
class SinePositionalEmbedding(nn.Module): |
|
def __init__( |
|
self, |
|
embedding_dim: int, |
|
dropout: float = 0.0, |
|
scale: bool = False, |
|
alpha: bool = False, |
|
): |
|
super().__init__() |
|
self.embedding_dim = embedding_dim |
|
self.x_scale = math.sqrt(embedding_dim) if scale else 1.0 |
|
self.alpha = nn.Parameter(torch.ones(1), requires_grad=alpha) |
|
self.dropout = torch.nn.Dropout(p=dropout) |
|
|
|
self.reverse = False |
|
self.pe = None |
|
self.extend_pe(torch.tensor(0.0).expand(1, 4000)) |
|
|
|
def extend_pe(self, x): |
|
"""Reset the positional encodings.""" |
|
if self.pe is not None: |
|
if self.pe.size(1) >= x.size(1): |
|
if self.pe.dtype != x.dtype or self.pe.device != x.device: |
|
self.pe = self.pe.to(dtype=x.dtype, device=x.device) |
|
return |
|
pe = torch.zeros(x.size(1), self.embedding_dim) |
|
if self.reverse: |
|
position = torch.arange( |
|
x.size(1) - 1, -1, -1.0, dtype=torch.float32 |
|
).unsqueeze(1) |
|
else: |
|
position = torch.arange(0, x.size(1), dtype=torch.float32).unsqueeze(1) |
|
div_term = torch.exp( |
|
torch.arange(0, self.embedding_dim, 2, dtype=torch.float32) |
|
* -(math.log(10000.0) / self.embedding_dim) |
|
) |
|
pe[:, 0::2] = torch.sin(position * div_term) |
|
pe[:, 1::2] = torch.cos(position * div_term) |
|
pe = pe.unsqueeze(0) |
|
self.pe = pe.to(device=x.device, dtype=x.dtype).detach() |
|
|
|
def forward(self, x: torch.Tensor) -> torch.Tensor: |
|
self.extend_pe(x) |
|
output = x.unsqueeze(-1) if x.ndim == 2 else x |
|
output = output * self.x_scale + self.alpha * self.pe[:, : x.size(1)] |
|
return self.dropout(output) |
|
|
