TensorFlow导入报错：annotation_types模块无Float8e4m3fn属性

问题场景 #

开发GTA操控AI（先实现驾驶、再实现射击功能），基于Jupyter Notebook编写代码。首次运行导入TensorFlow正常，第二次运行触发AttributeError：module 'tensorflow.security.fuzzing.py.annotation_types' has no attribute 'Float8e4m3fn'，重装TensorFlow无效。

报错栈示例 #

AttributeError                            Traceback (most recent call last)
Cell In[2], line 1
----> 1 import tensorflow as tf

File ~/.local/lib/python3.10/site-packages/tensorflow/__init__.py:45
     42 from tensorflow.python import pywrap_tensorflow as _pywrap_tensorflow
     43 from tensorflow.python.platform import app
---> 45 from tensorflow.security.fuzzing.py import annotation_types
     46 from tensorflow.core.framework import function_pb2
     47 from tensorflow.core.framework import graph_pb2

File ~/.local/lib/python3.10/site-packages/tensorflow/security/fuzzing/py/annotation_types.py:28
     25 from tensorflow.python.framework import dtypes
     26 from tensorflow.python.platform import tf_logging as logging
---> 28 Float8e4m3fn = dtypes.float8_e4m3fn
     29 Float8e5m2 = dtypes.float8_e5m2
     31 _DTYPE_TO_ANNOTATION_TYPE = {
     32     dtypes.float16: annotation_pb2.AnnotationType.FLOAT16,
     33     dtypes.bfloat16: annotation_pb2.AnnotationType.BFLOAT16,
   (...)
     38     dtypes.int64: annotation_pb2.AnnotationType.INT64,
     39 }

AttributeError: module 'tensorflow.security.fuzzing.py.annotation_types' has no attribute 'Float8e4m3fn'

解决步骤 #

• 重启Jupyter内核：Jupyter首次运行后可能残留TensorFlow加载状态，导致二次运行冲突。操作：点击界面上方Kernel -> Restart，重新运行所有单元格。
• 清理TensorFlow缓存：缓存文件可能引发版本冲突，执行终端命令删除缓存：

  # Linux/macOS
  rm -rf ~/.keras/cache
  rm -rf ~/.tensorflow/cache
  
  # Windows
  Remove-Item -Recurse -Force $env:USERPROFILE\.keras\cache
  Remove-Item -Recurse -Force $env:USERPROFILE\.tensorflow\cache
  

• 指定稳定TensorFlow版本：当前版本可能存在Float8特性相关bug，切换到2.15.x等稳定版本：

  pip uninstall tensorflow -y
  pip install tensorflow==2.15.0
  

• 禁用IPython自动重载：若代码中使用%autoreload魔法命令，可能导致模块重复加载出错，注释相关代码：

  # %load_ext autoreload
  # %autoreload 2
  

• 排查依赖冲突：检查numpy、protobuf等依赖版本是否与TensorFlow兼容：

  pip list | grep -E "tensorflow|numpy|protobuf"
  

  若版本不匹配，重装对应依赖：

  pip install numpy==1.26.4 protobuf==4.25.3
  

内容的提问来源于stack exchange，提问作者Jul Go