Do not use pipeline to achieve faster generation of Chinese mask repl…

textattack/transformations/word_swaps/chn_transformations/chinese_word_swap_masked.py

@@ -13,23 +13,24 @@ class ChineseWordSwapMaskedLM(WordSwap):
     model."""
 
     def __init__(self, task="fill-mask", model="xlm-roberta-base", **kwargs):
-        self.unmasker = pipeline(task, model)
+        from transformers import BertTokenizer, BertForMaskedLM
+        import torch
+        self.tt = BertTokenizer.from_pretrained(model)
+        self.mm = BertForMaskedLM.from_pretrained(model)
+        self.mm.to("cuda")
         super().__init__(**kwargs)
 
     def get_replacement_words(self, current_text, indice_to_modify):
-        masked_text = current_text.replace_word_at_index(indice_to_modify, "<mask>")
-        outputs = self.unmasker(masked_text.text)
-        words = []
-        for dict in outputs:
-            take = True
-            for char in dict["token_str"]:
-                # accept only Chinese characters for potential substitutions
-                if not is_cjk(char):
-                    take = False
-            if take:
-                words.append(dict["token_str"])
-
-        return words
+        masked_text = current_text.replace_word_at_index(indice_to_modify, "[MASK]")  # 修改前<mask>，xlmrberta的模型
+        tokens = self.tt.tokenize(masked_text.text)
+        input_ids = self.tt.convert_tokens_to_ids(tokens)
+        input_tensor = torch.tensor([input_ids]).to("cuda")
+        with torch.no_grad():
+            outputs = self.mm(input_tensor)
+            predictions = outputs.logits
+        predicted_token_ids = torch.argsort(predictions[0, indice_to_modify], descending=True)[:50]
+        predicted_tokens = self.tt.convert_ids_to_tokens(predicted_token_ids.tolist()[1:])
+        return predicted_tokens
 
     def _get_transformations(self, current_text, indices_to_modify):
         words = current_text.words