升级TensorFlow 1.13.1系列版本后运行异常，求助排查Issue#24748

Hey there, let's work through this DLL load failure you're hitting after upgrading TensorFlow to version 1.13.1, tensorflow-estimator 1.13.0, and tensorflow-gpu 1.13.1. I've dealt with similar Windows-specific TensorFlow issues before, so here are some targeted fixes to try:

• Verify CUDA and cuDNN version compatibility
  TensorFlow 1.13.1 requires exactly CUDA 10.0 and cuDNN 7.4.x (not newer or older versions). Double-check these details:

  • Confirm CUDA 10.0's bin and libnvvp directories are added to your system PATH (typically C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.0\bin)
  • Ensure cuDNN 7.4.x files are copied into the corresponding CUDA folders (move the bin, include, lib folders from the cuDNN zip to your CUDA 10.0 root directory)

• Reinstall Visual C++ Redistributables
  Windows builds of TensorFlow depend on the 64-bit version of the Visual C++ Redistributable for Visual Studio 2015-2019. Your upgrade might have corrupted this component, so download and reinstall it (make sure to pick the x64 installer).

• Clean reinstall TensorFlow packages
  Corrupted installation files are a common culprit. Uninstall all existing TensorFlow-related packages first, then reinstall with fresh copies:

  pip uninstall tensorflow tensorflow-estimator tensorflow-gpu -y
  pip install --no-cache-dir tensorflow==1.13.1 tensorflow-estimator==1.13.0 tensorflow-gpu==1.13.1
  

  The --no-cache-dir flag ensures you don't reuse any corrupted cached installation files.

• Don't worry about the imp.py Issue #24748
  That comment in imp.py refers to an internal Python core development issue (specific to Python 3.6's module loading logic), not a problem with TensorFlow. It's completely unrelated to your DLL load error, so you can ignore that line entirely—your root issue lies in missing or mismatched TensorFlow dependencies.

Give these steps a try in order, and you should be able to get TensorFlow running again.

内容的提问来源于stack exchange，提问作者Angelo Mascaro