PyTorch -OverflowError的Flair框架: int太大，无法转换

Question

我试图用Flair框架(https://github.com/flairNLP/flair)来训练一个命名的实体识别模型，其嵌入方式是：TransformerWordEmbeddings('emilyalsentzer/Bio_ClinicalBERT')。但是，它在OverflowError: int too big to convert中总是失败。在其他一些转换器字嵌入(如XLNet )中也会发生这种情况。然而，BERT和RoBERTa工作得很好。

以下是错误的完整跟踪：

2021-04-15 09:34:48,106 ----------------------------------------------------------------------------------------------------
2021-04-15 09:34:48,106 Corpus: "Corpus: 778 train + 259 dev + 260 test sentences"
2021-04-15 09:34:48,106 ----------------------------------------------------------------------------------------------------
2021-04-15 09:34:48,106 Parameters:
2021-04-15 09:34:48,106  - learning_rate: "0.1"
2021-04-15 09:34:48,106  - mini_batch_size: "32"
2021-04-15 09:34:48,106  - patience: "3"
2021-04-15 09:34:48,106  - anneal_factor: "0.5"
2021-04-15 09:34:48,106  - max_epochs: "200"
2021-04-15 09:34:48,106  - shuffle: "True"
2021-04-15 09:34:48,106  - train_with_dev: "False"
2021-04-15 09:34:48,106  - batch_growth_annealing: "False"
2021-04-15 09:34:48,107 ----------------------------------------------------------------------------------------------------
2021-04-15 09:34:48,107 Model training base path: "/home/xxx/data/xxx-clinical-bert"
2021-04-15 09:34:48,107 ----------------------------------------------------------------------------------------------------
2021-04-15 09:34:48,107 Device: cuda:0
2021-04-15 09:34:48,107 ----------------------------------------------------------------------------------------------------
2021-04-15 09:34:48,107 Embeddings storage mode: gpu
2021-04-15 09:34:48,116 ----------------------------------------------------------------------------------------------------
Traceback (most recent call last):
  File "train_medical_2.py", line 144, in <module>
    train_ner(d + '-base-ent',corpus_base)
  File "train_medical_2.py", line 136, in train_ner
    max_epochs=200)
  File "/home/d111199102201607101/flair/lib/python3.6/site-packages/flair/trainers/trainer.py", line 381, in train
    loss = self.model.forward_loss(batch_step)
  File "/home/d111199102201607101/flair/lib/python3.6/site-packages/flair/models/sequence_tagger_model.py", line 637, in forward_loss
    features = self.forward(data_points)
  File "/home/d111199102201607101/flair/lib/python3.6/site-packages/flair/models/sequence_tagger_model.py", line 642, in forward
    self.embeddings.embed(sentences)
  File "/home/d111199102201607101/flair/lib/python3.6/site-packages/flair/embeddings/token.py", line 81, in embed
    embedding.embed(sentences)
  File "/home/d111199102201607101/flair/lib/python3.6/site-packages/flair/embeddings/base.py", line 60, in embed
    self._add_embeddings_internal(sentences)
  File "/home/d111199102201607101/flair/lib/python3.6/site-packages/flair/embeddings/token.py", line 923, in _add_embeddings_internal
    self._add_embeddings_to_sentence(sentence)
  File "/home/d111199102201607101/flair/lib/python3.6/site-packages/flair/embeddings/token.py", line 999, in _add_embeddings_to_sentence
    truncation=True,
  File "/home/d111199102201607101/flair/lib/python3.6/site-packages/transformers/tokenization_utils_base.py", line 2438, in encode_plus
    **kwargs,
  File "/home/d111199102201607101/flair/lib/python3.6/site-packages/transformers/tokenization_utils_fast.py", line 472, in _encode_plus
    **kwargs,
  File "/home/d111199102201607101/flair/lib/python3.6/site-packages/transformers/tokenization_utils_fast.py", line 379, in _batch_encode_plus
    pad_to_multiple_of=pad_to_multiple_of,
  File "/home/d111199102201607101/flair/lib/python3.6/site-packages/transformers/tokenization_utils_fast.py", line 330, in set_truncation_and_padding
    self._tokenizer.enable_truncation(max_length, stride=stride, strategy=truncation_strategy.value)
OverflowError: int too big to convert

我试图改变embedding_storage_mode、hidden_size和mini_batch_size。这些都没有给我解决问题的办法。

有人有同样的问题吗？有什么办法解决这个问题吗？

谢谢

Answer 1

可以使用以下参数来限制令牌的长度

embedding = TransformerWordEmbeddings('emilyalsentzer/Bio_ClinicalBERT')
embedding.max_subtokens_sequence_length = 512
embedding.stride = 512//2