1. First of all, it is recommended to install with Anaconda
Because Anaconda itself has installed many common Python libraries by default, it can save a lot of library installation process and solve compatibility problems.
The installation of Anaconda itself is also very simple. Search the official website of Anaconda, download the latest exe file corresponding to Anaconda, about a few hundred megabytes, and double-click to install it.
Start Anaconda navigator, switch to the Environments tab, and select Open Terminal to start the DOS command window. As shown below:
2. Next, install the GPU version of Tensorflow
Enter this command at the DOS command line to start the installation:
pip install --upgrade tensorflow-gpu
If installing the specified version is:
pip install --upgrade tensorflow-gpu==1.3.0
The latest version is Tensorflow version 1.8.
3. Continue to install CUDA and cudnn dependent environment
1. To install CUDA, you need to install VISUAL STUDIO 2013 or above (preferably to install VS2013);
2. Install CUDA8.0 and cudnn v6.0;
All CUDA versions address: https://developer.nvidia.com/cuda-toolkit-archive
CUDA8 address: http://developer2.download.nvidia.com/compute/cuda/8.0/secure/prod/local_installers/cuda_8.0.44 _windows.exe
cudnn download address: https://developer.nvidia.com/rdp/cudnn-download, select the windows 7 version of cudnn v6.0 for CUDA 8.0 to run normally under windows 2012.
Unzip cudnn directly. For example, after unzipping it to C:\CUDNN, add the subfolder bin directory to the system variable path.
At this point, the GPU version of TensorFlow has been installed and configured.
Notes:
1. Before installing CUDA, you must close antivirus software and 360, etc., otherwise the installation may fail.
2. If a high-version graphics card driver has been installed before, you need to customize and cancel the bottom driver installation when installing CUDA, otherwise the installation will fail.
Problems and solutions:
ReadTimeoutError: HTTPSConnectionPool(host='pypi.python.org', port=443): Read timed out.
you encounter this download library supermarket error, you can set the timeout time by adding a parameter to the command
pip --default-timeout =100 install