Create tokenizer_make.py

tokenizer_make.py

@@ -0,0 +1,53 @@
+from transformers import BertTokenizerFast
+import os
+
+class MiniSunTokenizer:
+    def __init__(self, vocab_file=None):
+        # You can use BERT's tokenizer or any custom vocabulary tokenizer
+        if vocab_file:
+            self.tokenizer = BertTokenizerFast(vocab_file=vocab_file, do_lower_case=False)
+        else:
+            # Default BERT tokenizer without a specific vocab file
+            self.tokenizer = BertTokenizerFast.from_pretrained('bert-base-uncased')
+        
+        # Define special tokens if needed (customizable)
+        self.pad_token = '[PAD]'
+        self.unk_token = '[UNK]'
+        self.cls_token = '[CLS]'
+        self.sep_token = '[SEP]'
+        self.mask_token = '[MASK]'
+
+    def tokenize(self, text):
+        # Tokenizes the input text
+        return self.tokenizer.tokenize(text)
+
+    def encode(self, text, max_length=512, padding=True, truncation=True):
+        # Converts the text into input IDs and attention mask
+        encoded = self.tokenizer.encode_plus(
+            text, 
+            add_special_tokens=True, 
+            max_length=max_length, 
+            padding='max_length' if padding else False, 
+            truncation=truncation,
+            return_attention_mask=True,
+            return_tensors='tf'
+        )
+        return encoded['input_ids'], encoded['attention_mask']
+
+    def decode(self, token_ids):
+        # Decodes token IDs back into text
+        return self.tokenizer.decode(token_ids, skip_special_tokens=True)
+
+    def save_pretrained(self, save_directory):
+        # Save the tokenizer in Hugging Face format
+        os.makedirs(save_directory, exist_ok=True)
+        self.tokenizer.save_pretrained(save_directory)
+
+# Example usage of the tokenizer
+tokenizer = MiniSunTokenizer()
+
+text = "Hello, this is a test sentence for MiniSun model."
+input_ids, attention_mask = tokenizer.encode(text, max_length=20)
+
+print("Input IDs:", input_ids)
+print("Attention Mask:", attention_mask)