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

Project_Model/Libs/Batch/Classes/Batcher.py

@@ -1,10 +1,13 @@
+import random
+from typing import Generator
 import pandas as pd
-from pathlib import Path
+
 import Project_Model.Libs.BPE as BPE
-#from BPE import TokeNanoCore as Tokenizer
 from Scripts.Libs.CleaningPipeline.special_token import SpecialToken
 from Project_Model.Libs.Transformer.Classes.SpannedMasker import SpannedMasker
-import random
+from TokenCompletation import TokenCompletationTransformer
+from Project_Model.Libs.BPE.Enums.SpecialToken import SpecialToken
+
 class Batcher:
 
     def __init__(self, dataset_path: str, batch_size:int, tokenizer: BPE.TokeNanoCore, masker: SpannedMasker) -> None:
@@ -14,34 +17,37 @@ class Batcher:
         #   text2rdf: X: ABSTRACT, X:TRIPLE 
         #   masking ( call masker): X: incomplete_triple Y: complete_triple (as exam) 
         #   completation: X: TRIPLE SUBSET, Y: related TRIPLE SUBSET
-        # self._DATASET = pd.read_csv(dataset_path)
         self._dataset_path = dataset_path
         self._batch_size = batch_size
         self._tokenizer = tokenizer
         self._masker = masker
 
-    def get_batch(self):
-        for batch in pd.read_csv(self._dataset_path, chunksize= int(self._batch_size/3)): #now we support 3 task
-            # each batch get 4 transformation for the 4 tasks and then shuffled
-            # now a batch is ["Abstract"], ["Triples"]
-            # tokenize the strings:
-            # batch = batch.drop(columns=['MovieID'])
-            tokenized_batch = pd.DataFrame()
-            # bho = batch.map(lambda x: self._tokenizer.encode(x))
-            tokenized_batch[["Abstract","RDFs"]] = batch[["Abstract","RDFs"]].map(
-                lambda t: self._tokenizer.encode(t))
+        sotl = self._tokenizer.encode(SpecialToken.START_TRIPLE_LIST.value)
+        eos = self._tokenizer.encode(SpecialToken.END_OF_SEQUENCE.value)
+        self._token_completation = TokenCompletationTransformer(sotl,eos)
 
 
+    def get_batch(self)-> Generator[pd.DataFrame]:
+        for batch in pd.read_csv(self._dataset_path, chunksize= int(self._batch_size/4)): #now we support 3 task
+
+            tokenized_batch = pd.DataFrame()
+            tokenized_batch[["Abstract","RDFs"]] = (
+                batch[["Abstract","RDFs"]]
+                .map(lambda t: self._tokenizer.encode(t))
+            )
+            
             rdf2txt_batch = self.__rdf2txt_transformation(tokenized_batch)
             txt2rdf_batch = self.__txt2rdf_transformation(tokenized_batch)
             mask_batch = self.__masking_trasformation(tokenized_batch)
+            completation_batch = self.__token_completation_task(tokenized_batch)
 
-            output = pd.concat([rdf2txt_batch,txt2rdf_batch,mask_batch],ignore_index=True)
-            output.sample(frac=1).reset_index(drop=True)
+            output = pd.concat([rdf2txt_batch,txt2rdf_batch,mask_batch,completation_batch],ignore_index=True)
+            output = output.sample(frac=1).reset_index(drop=True)
             yield output
             
 
     def __random_subset_rdfs(self, batch: pd.DataFrame, seed = 0):
+        # WIP 
         rng = random.Random(seed)
 
         def to_list(x):
@@ -52,16 +58,13 @@ class Batcher:
         )
 
     def __rdf2txt_transformation(self, batch: pd.DataFrame):
-        # rename ["Triples"] as ["X"]
-        # rename ["Abstract"] as ["Y"]
-        # return just them
         batch = batch.rename(columns={"RDFs": "X", "Abstract": "Y"})
-        return batch[["X", "Y"]] #.sample(frac=1).reset_index(drop=True)
+        return batch[["X", "Y"]]
 
 
     def __txt2rdf_transformation(self, batch: pd.DataFrame):
         batch = batch.rename(columns={ "Abstract": "X","RDFs": "Y"})
-        return batch[["X", "Y"]]# .sample(frac=1).reset_index(drop=True)
+        return batch[["X", "Y"]]
     
     def __masking_trasformation(self, batch: pd.DataFrame):
         # mask_sequence: List[int] -> Tuple[List[int], List[int]]
@@ -73,11 +76,19 @@ class Batcher:
         return output[["X", "Y"]] 
         
     
+    def __token_completation_task(self, batch: pd.DataFrame):
+        xy_tuples = batch["RDFs"].apply(self._token_completation.get_completation_tuple)
+        output = batch.copy()
+        output[["X", "Y"]] = pd.DataFrame(xy_tuples.tolist(), index=batch.index)
+        return output[["X", "Y"]] 
 
 
+
+"""
 DATASET_PATH = "Assets/Dataset/Tmp/rdf_text.csv"
 VOCABULARY_path = "Assets/Dataset/Tmp/trimmed.json"
 
+from pathlib import Path
 VOCABULARY = BPE.load_nanos_vocabulary(Path(VOCABULARY_path))
 SPECIAL_LIST = BPE.default_special_tokens()
 TOKENANO = BPE.TokeNanoCore(VOCABULARY, SPECIAL_LIST)
@@ -87,6 +98,7 @@ MASKER = SpannedMasker(TOKENANO.vocabulary_size,SPECIAL_TOKENS)
 
 prova = "<ABS>Cactus Flower is a 1969 American screwball comedy film directed by Gene Saks, and starring Walter Matthau, Ingrid Bergman and Goldie Hawn, who won an Academy Award for her performance.The screenplay was adapted by I. A. L. Diamond from the 1965 Broadway play of the same title written by Abe Burrows, which, in turn, is based on the French play Fleur de cactus by Pierre Barillet and Jean-Pierre Gredy. Cactus Flower was the ninth highest-grossing film of 1969."
 print(TOKENANO.encode(prova))
-batcher = Batcher(DATASET_PATH,9,TOKENANO,MASKER)
+batcher = Batcher(DATASET_PATH,8,TOKENANO,MASKER)
 for batch in batcher.get_batch():
     print(batch)
+"""

Scripts/Libs/CleaningPipeline/special_token.py

@@ -10,6 +10,7 @@ class SpecialToken(Enum):
     RELATIONSHIP = "<PRED>"
     OBJECT = "<OBJ>"
     ABSTRACT = "<ABS>"
+    END_OF_SENTENCE = "<EOS>"
     CORPUS_END = "<END>"
 
     ## Tasks' Token