使用TFHub的BERT预处理时出现NotFoundError问题求助

在2021款MacBook Pro（Apple Silicon）设备上，使用Python 3.9.13和TensorFlow v2.9.2调用TensorFlow Hub的预训练BERT预处理模型时，执行文本预处理操作返回NotFoundError，无法解决。预处理模型地址为：https://tfhub.dev/tensorflow/bert_en_uncased_preprocess/3

代码 #

import tensorflow_hub as hub

bert_preprocess = hub.KerasLayer("https://tfhub.dev/tensorflow/bert_en_uncased_preprocess/3")
bert_encoder = hub.KerasLayer("https://tfhub.dev/tensorflow/bert_en_uncased_L-12_H-768_A-12/4")
print(bert_preprocess(["test"]))

报错信息 #

Output exceeds the size limit. Open the full output data in a text editor
---------------------------------------------------------------------------
NotFoundError                             Traceback (most recent call last)
Cell In [42], line 3
      1 bert_preprocess = hub.KerasLayer("https://tfhub.dev/tensorflow/bert_en_uncased_preprocess/3")
      2 bert_encoder = hub.KerasLayer("https://tfhub.dev/tensorflow/bert_en_uncased_L-12_H-768_A-12/4")
----> 3 print(bert_preprocess(["test"]))

File ~/miniforge3/envs/tfenv/lib/python3.9/site-packages/keras/utils/traceback_utils.py:67, in filter_traceback.<locals>.error_handler(*args, **kwargs)
     65 except Exception as e:  # pylint: disable=broad-except
     66   filtered_tb = _process_traceback_frames(e.__traceback__)
---> 67   raise e.with_traceback(filtered_tb) from None
     68 finally:
     69   del filtered_tb

File ~/miniforge3/envs/tfenv/lib/python3.9/site-packages/tensorflow_hub/keras_layer.py:237, in KerasLayer.call(self, inputs, training)
    234   else:
    235     # Behave like BatchNormalization. (Dropout is different, b/181839368.)
    236     training = False
---> 237   result = smart_cond.smart_cond(training,
    238                                  lambda: f(training=True),
    239                                  lambda: f(training=False))
    241 # Unwrap dicts returned by signatures.
    242 if self._output_key:

File ~/miniforge3/envs/tfenv/lib/python3.9/site-packages/tensorflow_hub/keras_layer.py:239, in KerasLayer.call.<locals>.<lambda>()
...
     [[StatefulPartitionedCall/StatefulPartitionedCall/bert_pack_inputs/PartitionedCall/RaggedConcat/ArithmeticOptimizer/AddOpsRewrite_Leaf_0_add_2]] [Op:__inference_restored_function_body_209194]

Call arguments received by layer "keras_layer_6" (type KerasLayer):
  • inputs=["'test'"]
  • training=None

核心原因 #

Apple Silicon（M系列）芯片的TensorFlow环境中，旧版本TensorFlow（如v2.9.2）与部分TensorFlow Hub预训练模型存在兼容性问题，导致模型执行时出现资源定位失败的NotFoundError。

解决步骤 #

• 升级TensorFlow版本：将TensorFlow升级到v2.10及以上版本，新版本对M系列芯片的支持更完善，修复了大量兼容性bug。执行命令：

  pip install --upgrade tensorflow
  

• 确认环境架构：确保当前Python环境是通过Miniforge/Conda-forge创建的Apple Silicon专属环境，避免使用Rosetta模拟的x86环境，这类环境容易引发适配问题。
• 本地加载模型：如果升级后仍有问题，手动下载模型到本地目录，再通过本地路径加载，避免网络加载时的潜在问题：
  1. 下载预处理模型和编码器模型到本地文件夹（如./bert_preprocess和./bert_encoder）
  2. 修改代码加载本地模型：

     bert_preprocess = hub.KerasLayer("./bert_preprocess")
     bert_encoder = hub.KerasLayer("./bert_encoder")
     

• 检查输入格式：报错显示输入为["'test'"]，确认代码中输入是纯字符串列表["test"]，避免多余引号导致的格式异常。

内容的提问来源于stack exchange，提问作者Joseph Yu