For the love of data and machines that can learn!
Installing Tensorflow-GPU on Windows
To use tensorflow library on GPU, NVIDIA CUDA Toolkit and cuDNN libraries need to be first installed. Installing
tensorflow-gpu is straight forward. Installing the NVIDIA CUDA Toolkit and cuDNN is slightly tricky, we need to ensure a couple of things so that CUDA toolkit works that tensorflow-gpu uses under the hood.
- Download CUDA Toolkit 8.0 and cuDNN v5.1 for CUDA Toolkit 8.0 from NVIDIA Developer portal
- Install CUDA Toolkit 8.0
- Sometimes if the GPU is a latest one, CUDA Toolkit may not have support for it. This shouldn’t stop you from using CUDA Tookit, but the CUDA Toolkit installation warns that it couldn’t find a compatible GPU Hardware. In that case, go for a custom installation and choose not to install the drivers. Should be easy to identify in the custom install screen.
- Copy the contents of the cuDNN downloaded archive into CUDA Toolkit installed folder.
- Add below two paths to the system ‘PATH’ environment variable (if CUDA Toolkit installation didn’t add it):
- C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0\lib\x64
- C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0\extras\CUPTI\libx64
- Ensure below three environment variables are available with value
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v8.0 or wherever the CUDA toolkit is installed. Create the missing variables.
CUDA_HOME - this needs to be added manually
CUDA_PATH - this is usually created after installing CUDA Toolkit 8.0
CUDA_PATH_V8_0 - this is usually created after installing CUDA Toolkit 8.0
- If tensorflow package without gpu support is installed, uninstall it.
- Install tensorflow-gpu using
pip install tensorflow-gpu OR
conda install tensorflow-gpu (if using anaconda). If already installed, uninstall it and then reinstall. It didn’t work for me until I reinstalled.
tflearn is required, install it with one of the below commands
pip install tflearn
conda install -c derickl tflearn
You can find the official tensorflow documentation guide for tensorflow-gpu setup here.
avannaldas [at] hotmail [dot] com