Merge branch 'dev.embedder' of https://repositories.communitynotfound.work/PoliBa-DeepLearning/NanoSocrates into dev.embedder

Project_Model/Libs/BPE/Classes/NanoSocratesBPE.py

@@ -189,7 +189,7 @@ class NanoSocratesBPE(Encoder):
             token_stack.appendleft(right_token)
             token_stack.appendleft(left_token)
 
-        return UTF_8_STRING_ARR.decode("utf-8")
+        return UTF_8_STRING_ARR.decode("utf-8", errors="ignore")
 
     def __token_decode(self, token_id: int) -> tuple[int, int]:
 

Project_Model/Libs/Transformer/Classes/Decoder.py

@@ -41,18 +41,19 @@ class Decoder(nn.Module):
             torch.Tensor,
             torch.Tensor,
             torch.Tensor,
+            torch.Tensor,
             torch.Tensor
         ]
     ):  # -> list[torch.Tensor]:  # k_x = v_x . While x_q = x
         # WARNING: args is needed to have sequential
-        x, k_x, v_x, padding_mask = args
+        x, k_x, v_x, src_padding_mask, tgt_padding_mask = args
 
         # build of attention mask
         attention_mask = get_causal_attention_mask(x.size(1))
 
         # 1) Masked Attention
         MASKED_ATTENTION = self.__masked_attention(
-            x, x, x, key_padding_mask=padding_mask, attention_mask=attention_mask
+            x, x, x, key_padding_mask=tgt_padding_mask, attention_mask=attention_mask
         )
 
         # 2) Dropout
@@ -68,7 +69,7 @@ class Decoder(nn.Module):
 
         # 5) Encoder–decoder (cross) attention
         CROSS_ATTENTION = self.__cross_attention(
-            x, k_x, v_x, key_padding_mask=padding_mask
+            x, k_x, v_x, key_padding_mask=src_padding_mask
         )
 
         # 6) Dropout
@@ -96,7 +97,7 @@ class Decoder(nn.Module):
         # 12) Layer Normalization
         x = self.__layer_norm_3(x)
 
-        return (x, k_x, v_x, padding_mask)
+        return (x, k_x, v_x, src_padding_mask, tgt_padding_mask)
 
 
 # use eval to disable dropout ecc

Project_Model/Libs/Transformer/Models/TrainingModel.py

@@ -37,17 +37,17 @@ class TrainingModel(torch.nn.Module):
 
         self.__detokener = DeToken(latent_space, vocabulary_size)
 
-    def forward(self, args: tuple[torch.Tensor, torch.Tensor, torch.Tensor]):
+    def forward(self, args: tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]):
 
-        encoder_embedder_input, padding_tensor, decoder_embedder_input = args
+        encoder_embedder_input, src_padding, decoder_embedder_input, tgt_padding = args
 
         encoder_tensor = self.__encoder_embedder(encoder_embedder_input)
         decoder_tensor = self.__decoder_embedder(decoder_embedder_input)
 
-        encoder_output, _ = self.__encoder((encoder_tensor, padding_tensor))
+        encoder_output, _ = self.__encoder((encoder_tensor, src_padding))
 
-        decoder_output, _, _, _ = self.__decoder(
-            (decoder_tensor, encoder_output, encoder_output, None)
+        decoder_output, _, _, _, _ = self.__decoder(
+            (decoder_tensor, encoder_output, encoder_output, src_padding, tgt_padding)
         )
 
         logits: torch.Tensor = self.__detokener(decoder_output)

Project_Model/Libs/Transformer/Utils/__init__.py

@@ -3,6 +3,7 @@ from .task_type import TaskType
 from .post_tokenization import truncate_sequence, pad_sequence, normalize_sequence, create_padding_mask
 from .inference_masking import inference_masking
 from .truncate_rdf_list import truncate_rdf_list
+from .decode_out import tensor2token
 
 __all__ = [
     "TaskType",
@@ -13,5 +14,6 @@ __all__ = [
     "create_padding_mask",
     "normalize_sequence",
     "inference_masking",
-    "truncate_rdf_list"
+    "truncate_rdf_list",
+    "tensor2token"
 ]

Project_Model/Libs/Transformer/Utils/decode_out.py

@@ -0,0 +1,27 @@
+from typing import Generator
+
+import torch
+
+
+def tensor2token(tensor: torch.Tensor, end_token: int) -> Generator[list[int]]:
+
+    if len(tensor.shape) < 1 or len(tensor.shape) > 2:
+        raise ValueError("Shape is not correct")
+
+    if len(tensor.shape) == 1:
+        token_list: list[int] = tensor.tolist()
+        token_list.append(end_token)
+        yield token_list
+        return
+
+    batch_len: int
+    batch_len, _ = tensor.shape
+
+    for i in range(batch_len):
+
+        smaller_tensor = tensor[i, :]
+        token_list: list[int] = smaller_tensor.tolist()
+        token_list.append(end_token)
+        yield token_list
+
+

Project_Model/Libs/Transformer/Utils/post_tokenization.py

@@ -1,17 +1,20 @@
 def truncate_sequence(
-    sequence: list[int], truncate_at: int, end_token: int
+    sequence: list[int], truncate_at: int, end_token: int, add_ending: bool
 ) -> list[int]:
 
     if len(sequence) < truncate_at - 1:
-        sequence.append(end_token)
+        if add_ending:
+            sequence.append(end_token)
         return sequence
 
     if len(sequence) < truncate_at:
-        sequence[-1] = end_token
+        if add_ending:
+            sequence[-1] = end_token
         return sequence
 
     TRUNCATED_SEQUENCE = sequence[:truncate_at]
-    TRUNCATED_SEQUENCE[-1] = end_token
+    if add_ending:
+        TRUNCATED_SEQUENCE[-1] = end_token
 
     return TRUNCATED_SEQUENCE
 
@@ -48,8 +51,9 @@ def normalize_sequence(
     max_length: int,
     pad_token: int,
     end_token: int,
+    add_ending: bool = True
 ) -> tuple[list[int], list[bool]]:
-    new_sequence = truncate_sequence(sequence, max_length, end_token)
+    new_sequence = truncate_sequence(sequence, max_length, end_token, add_ending)
     new_sequence = pad_sequence(new_sequence, max_length, pad_token)
     PADDING_MASK = create_padding_mask(new_sequence, pad_token)