Quick fix to architecture

Project_Model/Libs/Embedder/Classes/NanoSocratesEmbedder.py

@@ -9,11 +9,9 @@ class NanoSocratesEmbedder(torch.nn.Module):
         super().__init__()
         self.__embedder = torch.nn.Embedding(vocabulary_size, embedding_size)
 
-    def forward(self, tokenized_sentence: list[list[int]]) -> torch.Tensor:
+    def forward(self, tokenized_sentence: torch.Tensor) -> torch.Tensor:
 
-        TOKENIZED_TENSOR = torch.tensor(tokenized_sentence)
-
-        computed_embeddings: torch.Tensor = self.__embedder(TOKENIZED_TENSOR)
+        computed_embeddings: torch.Tensor = self.__embedder(tokenized_sentence)
 
         _, SENTENCE_LENGHT, EMBEDDING_SIZE = computed_embeddings.shape  # for batching
 

Project_Model/Libs/Transformer/Classes/Decoder.py

@@ -56,12 +56,12 @@ class Decoder(nn.Module):
         )
 
         # 2) Dropout
-        DROPPED_MASKED_ATTENTION = self.__dropout(MASKED_ATTENTION)
-        del MASKED_ATTENTION
+        # DROPPED_MASKED_ATTENTION = self.__dropout(MASKED_ATTENTION)
+        # del MASKED_ATTENTION
 
         # 3) Residual Connection
-        x = x + DROPPED_MASKED_ATTENTION
-        del DROPPED_MASKED_ATTENTION
+        x = x + MASKED_ATTENTION
+        del MASKED_ATTENTION
 
         # 4) Layer Normalization
         x = self.__layer_norm_1(x)
@@ -72,12 +72,12 @@ class Decoder(nn.Module):
         )
 
         # 6) Dropout
-        DROPPED_CROSS_ATTENTION = self.__dropout(CROSS_ATTENTION)
-        del CROSS_ATTENTION
+        # DROPPED_CROSS_ATTENTION = self.__dropout(CROSS_ATTENTION)
+        # del CROSS_ATTENTION
 
         # 7) Residual Connection
-        x = x + DROPPED_CROSS_ATTENTION
-        del DROPPED_CROSS_ATTENTION
+        x = x + CROSS_ATTENTION
+        del CROSS_ATTENTION
 
         # 8) Layer Normalization
         x = self.__layer_norm_2(x)
@@ -86,12 +86,12 @@ class Decoder(nn.Module):
         FEED_FORWARD = self.__feed_forward_network(x)
 
         # 10) Dropout
-        DROPPED_FEED_FORWARD = self.__dropout(FEED_FORWARD)
-        del FEED_FORWARD
+        # DROPPED_FEED_FORWARD = self.__dropout(FEED_FORWARD)
+        # del FEED_FORWARD
 
         # 11) Residual Connection
-        x = x + DROPPED_FEED_FORWARD
-        del DROPPED_FEED_FORWARD
+        x = x + FEED_FORWARD
+        del FEED_FORWARD
 
         # 12) Layer Normalization
         x = self.__layer_norm_3(x)

Project_Model/Libs/Transformer/Classes/Encoder.py

@@ -43,11 +43,12 @@ class Encoder(
         ATTENTION = self.__attention(x, x, x, key_padding_mask=padding_mask)
 
         # 2) Dropout
-        DROPPED_ATTENTION = self.__dropout(ATTENTION)
-        del ATTENTION
+        # DROPPED_ATTENTION = self.__dropout(ATTENTION)
+        # del ATTENTION
 
         # 3) Residual Connection
-        x = x + DROPPED_ATTENTION
+        x = x + ATTENTION
+        del ATTENTION
 
         # 4) Layer Normalization
         x = self.__layer_norm_1(x)
@@ -56,12 +57,12 @@ class Encoder(
         FEED_FORWARD = self.__feed_forward(x)
 
         # 6) Dropout
-        DROPPED_FEED_FORWARD = self.__dropout(FEED_FORWARD)
-        del FEED_FORWARD
+        # DROPPED_FEED_FORWARD = self.__dropout(FEED_FORWARD)
+        # del FEED_FORWARD
 
         # 7) Residual Connection
-        x = x + DROPPED_FEED_FORWARD
-        del DROPPED_FEED_FORWARD
+        x = x + FEED_FORWARD
+        del FEED_FORWARD
 
         # 8) Layer Normalization
         x = self.__layer_norm_2(x)

Project_Model/Libs/Transformer/Models/TrainingModel.py

@@ -37,12 +37,11 @@ class TrainingModel(torch.nn.Module):
 
         self.__detokener = DeToken(latent_space, vocabulary_size)
 
-    def forward(self, args: tuple[list[list[int]], list[list[bool]], list[list[int]]]):
+    def forward(self, args: tuple[torch.Tensor, torch.Tensor, torch.Tensor]):
         
-        encoder_embedder_input, padding_input, decoder_embedder_input = args
+        encoder_embedder_input, padding_tensor, decoder_embedder_input = args
 
         encoder_tensor = self.__encoder_embedder(encoder_embedder_input)
-        padding_tensor = torch.tensor(padding_input, dtype=torch.bool)
         decoder_tensor = self.__decoder_embedder(decoder_embedder_input)
 
         encoder_output, _ = self.__encoder((encoder_tensor, padding_tensor))

Project_Model/Libs/Transformer/Models/__init__.py

@@ -0,0 +1,5 @@
+from .TrainingModel import TrainingModel
+
+__all__ = [
+    "TrainingModel"
+]

Project_Model/Libs/Transformer/Utils/post_tokenization.py

@@ -2,6 +2,10 @@ def truncate_sequence(
     sequence: list[int], truncate_at: int, end_token: int
 ) -> list[int]:
 
+    if len(sequence) < truncate_at - 1:
+        sequence.append(end_token)
+        return sequence
+
     if len(sequence) < truncate_at:
         sequence[-1] = end_token
         return sequence

Project_Model/Libs/Transformer/__init__.py

@@ -1,5 +1,7 @@
 from .Classes import *
 from .Utils import *
+from .Models import *
 
 from . import Classes
 from . import Utils
+from . import Models