BERT等语言模型的BertForMaskedLM避的坑

在用transformers中的BertForMaskedLM来预测被mask掉的单词时一定要加特殊字符$[CLS]和[SEP]$。不然效果很差很差！！！

from transformers import AlbertTokenizer, AlbertForMaskedLM
import torch

tokenizer = AlbertTokenizer.from_pretrained('albert-base-v2', cache_dir='E:/Projects/albert/')
model = AlbertForMaskedLM.from_pretrained('E:/Projects/albert')

sentence = "It is a very beautiful book."
tokens = ['[CLS]'] + tokenizer.tokenize(sentence) + ['[SEP]']

# i就是被mask掉的id
for i in range(1, len(tokens)-1):
    tmp = tokens[:i] + ['[MASK]'] + tokens[i+1:]
    masked_ids = torch.tensor([tokenizer.convert_tokens_to_ids(tmp)])
    segment_ids = torch.tensor([[0]*len(tmp)])

    outputs = model(masked_ids, token_type_ids=segment_ids)
    prediction_scores = outputs[0]
    print(tmp)
    # 打印被预测的字符
    prediction_index = torch.argmax(prediction_scores[0, i]).item()
    predicted_token = tokenizer.convert_ids_to_tokens([prediction_index])[0]
    print(predicted_token)