Tensorflow Importerror: "dll Load Failed" And "no Module Named Pywrap_tensorflow_internal"

I am trying to install TensorFlow (not GPU version) on Windows 7. I have installed Python 3.5.2, which I can verify: $ python --version Python 3.5.2 I installed TensorFlow using t

Solution 1:

In my case, either cudnn v5 or v6 cannot work alone. I looked into the self check script, it seems that the proper installation of both cudnn64_5.dll and cudnn64_6.dll are checked:

cudnn5_found = Falsetry:
  cudnn5 = ctypes.WinDLL("cudnn64_5.dll")
  cudnn5_found = Trueexcept OSError:
  candidate_explanation = Trueprint("""
  - Could not load 'cudnn64_5.dll'. The GPU version of TensorFlow
  requires that this DLL be installed in a directory that is named in
  your %PATH% environment variable. Note that installing cuDNN is a
  separate step from installing CUDA, and it is often found in a
  different directory from the CUDA DLLs. You may install the
  necessary DLL by downloading cuDNN 5.1 from this URL:
  https://developer.nvidia.com/cudnn""")

cudnn6_found = Falsetry:
  cudnn = ctypes.WinDLL("cudnn64_6.dll")
  cudnn6_found = Trueexcept OSError:
  candidate_explanation = Trueifnot cudnn5_found ornot cudnn6_found:
  print()
  ifnot cudnn5_found andnot cudnn6_found:
    print("- Could not find cuDNN.")
  elifnot cudnn5_found:
   print("- Could not find cuDNN 5.1.")
else:
  print("- Could not find cuDNN 6.")
  print("""
  The GPU version of TensorFlow requires that the correct cuDNN DLL be 
  installed
  in a directory that is named in your %PATH% environment variable. Note 
  that
  installing cuDNN is a separate step from installing CUDA, and it is often
  found in a different directory from the CUDA DLLs. The correct version of
  cuDNN depends on your version of TensorFlow:

  * TensorFlow 1.2.1 or earlier requires cuDNN 5.1. ('cudnn64_5.dll')
  * TensorFlow 1.3 or later requires cuDNN 6. ('cudnn64_6.dll')

if either v5 or v6 is not found in Path, an OSError will occur. So I put both in my Path environment variable, and the check passed.

Solution 2:

This is a known issue while installing tensorflow-gpu.

If you are using conda the best way to solve this problem is by doing a conda installation of tensorflow-gpu. The steps are given below. (Tested in both Windows 10 and Ubuntu 16.04)

Uninstall existing installation of tensorflow-gpu

pip uninstall tensorflow-gpu

Then install the tensorflow-gpu using conda

conda install tensorflow-gpu

This should install tensorflow-gpu in your conda environment with all the dependencies.

If you are not using anaconda distribution of python, you may try using the correct versions of cudatoolkit, CuDNN and python. A list of common errors and their corresponding github threads with solutions can be found here.

https://www.tensorflow.org/install/errors

I would recommend you to use conda install than pip install if you don't want to spent much time on figuring out what went wrong and how to fix it.

Solution 3:

The script tensorflow_self_check.py works perfectly in my case. It points out that I miss the file cudnn64_6.dll of cuDNN v6. It is important to notice that in the official guide of TensorFlow for Window (https://www.tensorflow.org/install/install_windows), they insist that it must be cuDNN v5.1 with cuDNN64_5.dll ! They should update this guide by adding this tensorflow_self_check.py script.

EDIT: I should read carefully the release note of TensorFlow 1.3.0 https://github.com/tensorflow/tensorflow/blob/r1.3/RELEASE.md: "All our prebuilt binaries have been built with cuDNN 6. We anticipate releasing TensorFlow 1.4 with cuDNN 7."