python - BertModel训练期间的Python“无法腌制本地对象”异常

Question

我正在使用 simpletransformers.classification 来训练一个 Bert 模型来对一些文本输入进行分类。这是我的代码。

from simpletransformers.classification import ClassificationModel

import torch

import numpy as np # linear algebra

import pandas as pd # data processing, CSV file I/O (e.g. pd.read_csv)

from simpletransformers.classification import ClassificationModel

from pytorch_pretrained_bert import BertTokenizer, BertModel, BertForMaskedLM, BertForSequenceClassification

import parallelTestModule

# Lets import the csv file in pandas dataframe first
train_df = pd.read_csv('D:\\7allV03Small.csv', encoding='utf-8', header=None, names=['cat', 'text'])

# Check the df
print(train_df.head())

# unique categories
print(train_df.cat.unique())
print("Total categories",len(train_df.cat.unique()))

# convert string labels to integers
train_df['labels'] = pd.factorize(train_df.cat)[0]

print(train_df.head())

# Let's create a train and test set
from sklearn.model_selection import train_test_split

train, test = train_test_split(train_df, test_size=0.2, random_state=42)

print('Eğitim veri seti boyutu : ' + str(train.shape), ' Test eğitim seti : ' + str(test.shape))

if __name__ == "__main__":
    from multiprocessing import freeze_support

    model = ClassificationModel('bert', 'bert-base-multilingual-uncased',  use_cuda=False,  num_labels=8, args={'reprocess_input_data': True, 'overwrite_output_dir': True, 'num_train_epochs': 1,'train_batch_size':1})
    
    freeze_support()
    
    # Now lets fine tune bert with the train set
    model.train_model(train)

一切看起来都很好，它开始训练。但在训练结束时，它会出现如下错误。

Traceback (most recent call last):
  File "c:/Users/arslanom/Desktop/text/try.py", line 45, in <module>
    model.train_model(train)
  File "C:\Users\arslanom\AppData\Roaming\Python\Python36\site-packages\simpletransformers\classification\classification_model.py", line 269, in train_model
    **kwargs,
  File "C:\Users\arslanom\AppData\Roaming\Python\Python36\site-packages\simpletransformers\classification\classification_model.py", line 544, in train
    self._save_model(output_dir_current, optimizer, scheduler, model=model)
  File "C:\Users\arslanom\AppData\Roaming\Python\Python36\site-packages\simpletransformers\classification\classification_model.py", line 1113, in _save_model
    torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
  File "C:\Users\arslanom\AppData\Roaming\Python\Python36\site-packages\torch\serialization.py", line 209, in save
    return _with_file_like(f, "wb", lambda f: _save(obj, f, pickle_module, pickle_protocol))
  File "C:\Users\arslanom\AppData\Roaming\Python\Python36\site-packages\torch\serialization.py", line 134, in _with_file_like
    return body(f)
  File "C:\Users\arslanom\AppData\Roaming\Python\Python36\site-packages\torch\serialization.py", line 209, in <lambda>
    return _with_file_like(f, "wb", lambda f: _save(obj, f, pickle_module, pickle_protocol))
  File "C:\Users\arslanom\AppData\Roaming\Python\Python36\site-packages\torch\serialization.py", line 282, in _save
    pickler.dump(obj)
AttributeError: Can't pickle local object 'get_linear_schedule_with_warmup.<locals>.lr_lambda'

听起来这个问题与 worker_count 有关，因为它使用多线程运行。但我找不到任何解决方案。

操作系统：Windows 10

内存：16 Gb

标签： pythonnlptext-classificationtransformerbert-language-model