0

我正在尝试从 simpletransformers 训练 QuestionAnsweringModel bert-base-multilingual-uncased 并面临下一个问题：

AttributeError                            Traceback (most recent call last)
<ipython-input-10-40e9356ccee6> in <module>()
----> 1 model.train(traindata, output_dir='/content/drive/MyDrive')

1 frames
/usr/local/lib/python3.7/dist-packages/simpletransformers/question_answering/question_answering_model.py in train(self, train_dataset, output_dir, show_running_loss, eval_data, verbose, **kwargs)
    578                     steps_trained_in_current_epoch -= 1
    579                     continue
--> 580                 batch = tuple(t.to(device) for t in batch)
    581 
    582                 inputs = self._get_inputs_dict(batch)

/usr/local/lib/python3.7/dist-packages/simpletransformers/question_answering/question_answering_model.py in <genexpr>(.0)
    578                     steps_trained_in_current_epoch -= 1
    579                     continue
--> 580                 batch = tuple(t.to(device) for t in batch)
    581 
    582                 inputs = self._get_inputs_dict(batch)

AttributeError: 'str' object has no attribute 'to'

我的数据准备：

!wget https://onti2020.ai-academy.ru/task/rucos_test.jsonl
!wget https://onti2020.ai-academy.ru/task/rucos_val.jsonl
!wget https://onti2020.ai-academy.ru/task/rucos_train.jsonl.zip
!unzip rucos_train.jsonl.zip

!pip install nltk
import nltk
nltk.download('all')
from nltk.tokenize import word_tokenize

def get_train_data(jsonfile):
    res=[]
    with open(jsonfile, 'r') as data:
        trainlist=list(data)
        for item in tqdm(trainlist):
            item=json.loads(item)
            dictt={}
            dictt['context']=word_tokenize(item['passage']['text'])
            qas=[]
            qlist=item['qas']
            for q in qlist:
                qdict={}
                qdict['id']=str(q['idx']).rjust(6, '0')
                answers=[]
                qdict['is_impossible']=True
                qdict['question']=q['query']
                alist=q['answers']
                for a in alist:
                    adict={}
                    adict['text']=a['text']
                    adict['answer_start']=a['start']
                    answers.append(adict)
                qdict['answers']=answers
                qas.append(qdict)
            dictt['qas']=qas
            res.append(dictt)
    return res

traindata, evaldata=get_train_data('rucos_train.jsonl'), get_train_data('rucos_val.jsonl')

建筑模型：

!pip install simpletransformers
!pip install torch==1.5.0

from simpletransformers.question_answering import QuestionAnsweringModel, QuestionAnsweringArgs
model = QuestionAnsweringModel(
    "bert",
    "bert-base-multilingual-uncased",
    args=QuestionAnsweringArgs(n_best_size=2)
)

模型训练：

model.train(traindata, output_dir='/content/drive/MyDrive')

此代码在 Colab Pro 中执行并基于文档https://simpletransformers.ai/docs/qa-model/。

请帮我解决这个问题。

4

1 回答 1

1

您必须使用 train_model 函数而不是 train 函数。

model.train_model(traindata)

希望这个答案还为时不晚。

于 2021-04-19T06:47:34.360 回答