How to get tensorflow-gpu v2 working on Windows with NVidia GPU

Question:

What are the steps to get tensorflow-gpu 2.x Python package working on Windows with an NVidia GPU?

I.e. how can I get rid of Could not find 'cudart64_101.dll' and then Could not find 'cudnn64_7.dll'?

Asked By: Danny Varod

||
Source

Answers:

Steps

  • Requires specific versions according to the error messages you see, not latest versions!

1. Download and install latest NVidia driver

https://www.nvidia.com/Download/index.aspx

2. Install Tensorflow Python package

pip uninstall tensorflow
pip install tensorflow-gpu

(*) For newer versions, tensorflow and tensorflow-gpu are the same package, so just update tensorflow using:

pip install --upgrade tensforflow

3. Test

At first the following test will fail, pay attention to version of missing file e.g. Could not find 'cudart64_101.dll'

import tensorflow

tensorflow.test.is_built_with_gpu_support() # Test install of pip package, should output True
tensorflow.test.is_gpu_available() # Should output True
tensorflow.test.gpu_device_name() # Should output something like /device:GPU:0
# update for last test:
tensorflow.config.list_physical_devices('GPU')

4. Download and install CUDA Toolkit 10.1 local setup

  • The version you need is the one not found e.g. cudart64_101.dll –> version 10.1.
  • You will need to unselect components because setup contains an older driver, in network setup this doesn’t work right

https://developer.nvidia.com/cuda-10.1-download-archive-base?target_os=Windows&target_arch=x86_64&target_version=10&target_type=exelocal

Select custom setup and:

  • Unselect CUDA / Visual Studio Integration
  • Unselect Driver components

5. Make sure these folders have been added to path:

  • And no other versions of CUDA
C:Program FilesNVIDIA GPU Computing ToolkitCUDAv10.1bin;
C:Program FilesNVIDIA GPU Computing ToolkitCUDAv10.1libnvvp;

6. Test again

  • Restart your IDE for it to receive new environment variables (including PATH)

Will fail, pay attention to version of missing file e.g. Could not find 'cudnn64_7.dll'

7. Download and manually install CUDNN 7.6.5 for CUDA 10.1

  • The version you need is the one not found e.g. cudnn64_7.dll –> version 7.x for CUDA 10.1 (or the version of CUDA you needed)

https://developer.nvidia.com/rdp/cudnn-archive

  • Unzip and copy content of cuda folder into: C:Program FilesNVIDIA GPU Computing ToolkitCUDAv10.1 (or the version you downloaded and installed)
  • If you have any conflicts, skip duplicate files

8. Test again

  • Should work this time, assuming you downloaded the correct versions
Answered By: Danny Varod