Added embedder code for "Attention is all you need"
This commit is contained in:
parent
3b274ad807
commit
9a797a0485
32
Project_Model/Libs/Embedder/Classes/NanoSocratesEmbedder.py
Normal file
32
Project_Model/Libs/Embedder/Classes/NanoSocratesEmbedder.py
Normal file
@ -0,0 +1,32 @@
|
||||
import torch
|
||||
from ..Utils import fixed_positional_encoding
|
||||
class NanoSocratesEmbedder(torch.nn.Module):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vocabulary_size: int,
|
||||
embedding_size: int
|
||||
) -> None:
|
||||
super().__init__()
|
||||
self.__embedder = torch.nn.Embedding(
|
||||
vocabulary_size,
|
||||
embedding_size
|
||||
)
|
||||
|
||||
def forward(self, tokenized_sentence: list[int]) -> torch.Tensor:
|
||||
|
||||
TOKENIZED_TENSOR = torch.tensor(tokenized_sentence)
|
||||
|
||||
computed_embeddings: torch.Tensor = self.__embedder(TOKENIZED_TENSOR)
|
||||
|
||||
SENTENCE_LENGHT, EMBEDDING_SIZE = computed_embeddings.shape
|
||||
|
||||
POSITIONAL_ENCODINGS = fixed_positional_encoding(
|
||||
SENTENCE_LENGHT,
|
||||
EMBEDDING_SIZE
|
||||
)
|
||||
|
||||
computed_embeddings = computed_embeddings + POSITIONAL_ENCODINGS
|
||||
return computed_embeddings
|
||||
|
||||
|
||||
@ -0,0 +1,28 @@
|
||||
import torch
|
||||
|
||||
|
||||
def fixed_positional_encoding(
|
||||
sentence_dimension: int,
|
||||
embedding_dimension: int,
|
||||
) -> torch.Tensor:
|
||||
|
||||
BIG_CONST = int(1e4)
|
||||
INITIAL_ENCODING = torch.tensor([i for i in range(0, sentence_dimension)])
|
||||
|
||||
ENCODINGS: list[torch.Tensor] = []
|
||||
|
||||
for i in range(0, embedding_dimension):
|
||||
EMBEDDING_POSITION = i
|
||||
|
||||
# Note: The original paper did not specify
|
||||
# to compute: pos mod 2!!
|
||||
DIVISOR = BIG_CONST ** ((2 * (EMBEDDING_POSITION // 2)) / embedding_dimension)
|
||||
INTERMEDIATE_ENCODING = INITIAL_ENCODING / DIVISOR
|
||||
|
||||
if EMBEDDING_POSITION % 2 == 0:
|
||||
ENCODINGS.append(torch.sin(INTERMEDIATE_ENCODING))
|
||||
continue
|
||||
|
||||
ENCODINGS.append(torch.cos(INTERMEDIATE_ENCODING))
|
||||
|
||||
return torch.stack(ENCODINGS).transpose(0, 1)
|
||||
Loading…
x
Reference in New Issue
Block a user