升级TensorFlow至1.4后导入时出现Segmentation fault问题求助

Hey there, sorry to hear you're hitting a segmentation fault right after upgrading TensorFlow from 1.2 to 1.4—let's work through some common fixes that should help resolve this issue.

• Verify underlying library compatibility
  TensorFlow 1.4 has stricter requirements for dependencies like CUDA and cuDNN compared to 1.2. For example:

  • The GPU version requires CUDA 8.0 (not 7.5 or 9.1) and cuDNN 6.0
  • Double-check your versions with these commands:
    • CUDA: nvcc --version
    • cuDNN: Look for the CUDNN_MAJOR macro in /usr/local/cuda/include/cudnn.h
      If your versions don't match, you'll need to update or roll back these libraries to align with TensorFlow 1.4's specs.

• Do a clean reinstall of TensorFlow 1.4
  Partial upgrades often leave conflicting files behind. Let's start fresh:

  1. Uninstall all existing TensorFlow versions:

     pip uninstall -y tensorflow tensorflow-gpu
     

     (Use pip3 if you're using Python 3)
  2. Clear pip's cache to avoid corrupted packages:

     pip cache purge
     

  3. Reinstall the exact version you need:
     • CPU-only: pip install tensorflow==1.4
     • GPU-enabled: pip install tensorflow-gpu==1.4

• Test with a minimal import script
  Rule out your project code as the cause by running a super simple script:
  Create test_tf.py with:

  import tensorflow as tf
  print(f"TensorFlow version: {tf.__version__}")
  

  Run it with python test_tf.py. If the segfault still happens, the problem is definitely in your environment, not your project code.

• Get detailed crash logs with GDB
  On Linux, you can use the GNU Debugger to pinpoint exactly where the crash is happening:

  1. Launch GDB with Python:

     gdb python
     

  2. Run your test script inside GDB:

     run test_tf.py
     

  3. When the segfault occurs, type bt to print the call stack. This will show you which library or TensorFlow module is causing the crash.

• Check Python version compatibility
  TensorFlow 1.4 only supports Python 2.7, 3.4, and 3.5. If you're using Python 3.6+ (which wasn't supported until later TensorFlow versions), that could be the root cause. Check your Python version with python --version and roll back if needed.

• Temporarily disable GPU acceleration
  If you're using the GPU version, test if the crash happens on CPU only:
  Modify your test script to:

  import os
  # Force TensorFlow to use CPU
  os.environ["CUDA_VISIBLE_DEVICES"] = "-1"
  import tensorflow as tf
  print(f"TensorFlow version: {tf.__version__}")
  

  If this runs without issues, your GPU setup (CUDA/cuDNN) is the problem. Reinstall those libraries following TensorFlow 1.4's official guidelines.

内容的提问来源于stack exchange，提问作者user808657