使用BERT模型时遇CaseFoldUTF8未注册Op错误求助

报错信息 #

RuntimeError: Op type not registered 'CaseFoldUTF8' in binary running on MacBook-Pro-21.local. Make sure the Op and Kernel are registered in the binary running in this process. Note that if you are loading a saved graph which used ops from tf.contrib (e.g. tf.contrib.resampler), accessing should be done before importing the graph, as contrib ops are lazily registered when the module is first accessed.

可行解决方案 #

• 对齐TensorFlow与BERT相关库的版本：CaseFoldUTF8是TensorFlow内置操作，版本不匹配会导致注册失败。建议搭配使用TensorFlow 2.15.0和transformers 4.35.2，这两个版本兼容性较好。
• 显式触发Op注册：在加载模型前添加以下代码，强制TensorFlow注册该操作：

  import tensorflow as tf
  from tensorflow.python.ops import string_ops
  # 触发CaseFoldUTF8的注册
  _ = string_ops.case_fold_utf8(tf.constant("dummy_text"))
  

• 重新初始化模型而非加载旧图：如果是使用他人导出的预训练模型文件，可能包含旧版TensorFlow的图结构。改用transformers库直接从官方源加载权重：

  from transformers import TFBertForSequenceClassification, BertTokenizer
  
  tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
  # 根据任务调整num_labels，这里是二分类（紧急/不紧急）
  model = TFBertForSequenceClassification.from_pretrained('bert-base-uncased', num_labels=2)
  

• 适配Apple M2芯片：确保安装的是针对Apple Silicon优化的TensorFlow版本，执行以下命令安装：

  pip uninstall -y tensorflow
  pip install tensorflow-macos tensorflow-metal
  

• 清理依赖缓存后重装：彻底移除旧依赖并重新安装指定版本，避免缓存冲突：

  pip uninstall -y tensorflow transformers
  pip install tensorflow==2.15.0 transformers==4.35.2
  

内容的提问来源于stack exchange，提问作者asin