TensorFlow首次启动报错：pywrap_tensorflow.py导入失败求助

Hey there, let's work through this TensorFlow import error you're running into. That ImportError originating from pywrap_tensorflow.py is one of the most common issues folks hit after installing TensorFlow, and it's almost always tied to mismatched versions, missing system libraries, or a corrupted installation. Here are the step-by-step fixes I recommend trying:

• Verify Python & TensorFlow version compatibility
  You're using Python 3.5 with TensorFlow 1.x (since your code uses tf.Session()—that's a TF1-only API). TensorFlow 1.15 is the last stable release that fully supports Python 3.5. If you installed a newer TF version (like 2.x), even with compatibility mode, it can cause this import issue. Reinstall the matching version with:

  pip uninstall tensorflow
  pip install tensorflow==1.15
  

• Install missing system dependencies
  The pywrap_tensorflow module relies on core system libraries like libstdc++. For CPU-only installations, fix this by installing the required libraries:

  • On Ubuntu/Debian:

    sudo apt-get update && sudo apt-get install libstdc++6
    

  • On CentOS/RHEL:

    sudo yum install libstdc++-devel
    

  If you're using the GPU version of TensorFlow, double-check that your CUDA and cuDNN versions exactly match the requirements for your TensorFlow release (TF 1.15 needs CUDA 10.0 and cuDNN 7.4).

• Reinstall TensorFlow to fix corrupted files
  Sometimes pip downloads can get corrupted mid-install. Force a fresh reinstall without using cached files:

  pip install --force-reinstall --no-cache-dir tensorflow==1.15
  

• Confirm you're using the right Python environment
  If you're using a virtual environment, make sure you've activated it before running your code. Run these commands to verify:

  which python  # Checks which Python interpreter is active
  pip list | grep tensorflow  # Confirms TensorFlow is installed in this environment
  

• Optional: Switch to TensorFlow 2.x compatible code
  If you're open to updating your code to work with modern TensorFlow, you can ditch tf.Session() entirely. Here's the TF2 version of your test code:

  import tensorflow as tf
  hello = tf.constant('Hello, TensorFlow!')
  print(hello.numpy().decode('utf-8'))
  

  This avoids the old TF1 API compatibility issues altogether—just make sure you install a TF2 version that supports Python 3.5 (TF 2.0 is the last one that does).

内容的提问来源于stack exchange，提问作者Rufledore