fixed token-label alignment

iwslt2011.py

@@ -146,7 +146,7 @@ class IWSLT11(datasets.GeneratorBasedBuilder):
                 features=datasets.Features(
                     {
                         "ids": datasets.Sequence(datasets.Value("int32")),
-                        "tokens": datasets.Sequence(datasets.Value("string")),
+                        "tokens": datasets.Sequence(datasets.Value("int32")),
                         "labels": datasets.Sequence(
                            datasets.features.ClassLabel(
                                names=[p.name for p in self.config.punctuation]
@@ -194,12 +194,12 @@ class IWSLT11(datasets.GeneratorBasedBuilder):
             ]
         else:
             return [
-                datasets.SplitGenerator(
-                    name=datasets.Split.TRAIN,
-                    gen_kwargs={
-                        "filepath": files["train"]
-                    },
-                ),
+                # datasets.SplitGenerator(
+                #     name=datasets.Split.TRAIN,
+                #     gen_kwargs={
+                #         "filepath": files["train"]
+                #     },
+                # ),
                 datasets.SplitGenerator(
                     name=datasets.Split.VALIDATION,
                     gen_kwargs={
@@ -252,23 +252,24 @@ class IWSLT11(datasets.GeneratorBasedBuilder):
                 )
             ids = apply_window(np.arange(len(tokens)))
             tokens = apply_window(tokens)
+            tokens = self.config.tokenizer([t.tolist() for t in tokens], is_split_into_words=True, return_offsets_mapping=True, padding=True, truncation=True)
             labels = apply_window(labels)
-            for i, (ids, tokens, labels) in enumerate(zip(ids, tokens, labels)):
+            for i, (ids, labels) in enumerate(zip(ids, labels)):
                 if self.config.tokenizer is None:
                     raise ValueError('tokenizer argument has to be passed to load_dataset')
                 else:
-                    tokenized = self.config.tokenizer([tokens.tolist()], is_split_into_words=True, return_offsets_mapping=True, padding=True, truncation=True)
-                    offsets = np.array(tokenized['offset_mapping'][0])
+                    words = tokens[i].words
+                    offsets = np.array(tokens['offset_mapping'][i])
                     enc_labels = np.array([self.config.label_subword.name]*len(offsets), dtype=object)
-                    # todo: check if performance changes if in-word is set to NONE
-                    enc_labels[(offsets[:,0] == 0) & (offsets[:,1] != 0)] = [l.name for l in labels]
-                    #print(enc_labels)
-                    # not needed as long as the same tokenizer is used later?
-                    # tokens = {k:v[0] for k,v in tokenized if k != 'offset_mapping'}
+                    count = 0
+                    for j, word_id in enumerate(words):
+                        if word_id is not None and (j == 0 or words[j-1] != word_id):
+                            enc_labels[j] = labels[count].name
+                            count += 1
                     labels = enc_labels
                 yield i, {
                             "ids": ids,
-                            "tokens": tokens,
+                            "tokens": tokens['input_ids'][i],
                             "labels": labels,
                         }
             logging.info(f"Loaded number of tokens = {token_len}")