EDITORIAL: Check GPU supports CUDA
first determine under its own graphics models (do not tell me you do not know how to view your video card model).
You can see if your video card supports the column from the following URL.
If your video card is relatively new, check here supports CUDA: https: //developer.nvidia.com/cuda-gpus
If your video card is very old, please go to the following link checker supports CUDA: https: //developer.nvidia .com / cuda-legacy-gpus
note the difference between a laptop and a desktop.
installation steps
cuda version and cudnn not choose the right version, it would encounter a lot of pits, but fortunately, I was able to choose a good version control tensorflow.
Here is something you want to install, you can turn to install, you can also press the order you like to install. But after installation to ensure that the VC ++ plug-ins and python, was installed tensorflow.
-
Visual C ++ Redistributable for Visual Studio 2015: tensorflow fact, VC ++ 2015 to develop, so you need to install on Visual C ++ Redistributable for Visual Studio 2015 . Once downloaded, the default installation directly on the line
-
python environment: I use anaconda3.x, created a new python3.5 environment, and then install the python3.5 in tensorflow-gpu. This is not the case, try searching for "anaconda installation, anaconda Tsinghua source installation" and "anaconda python new environment." .
-
cuda9.0: CUDA-V9.0 . Once downloaded, installed by default on the line, to choose the middle agree on the election of local consent.
First, install CUDA
1. Download the appropriate version of CUDA: https://developer.nvidia.com/cuda-downloads
If you need to download the historical version of CUDA, please go here: https://developer.nvidia.com/cuda-toolkit-archive
Once you have downloaded the installation package, open just click 'next' Next to install, the default installation path is C drive just fine.
2, after the end of the installation process, the environmental variable configuration
① right-click "My Computer" and select "Properties", click "Advanced System Settings"
② Select "Environment Variables" in the "System Variables" box to find the "path", click to select, and click the "Edit" button
③ clicking the "New" button, the next frame of the four paths added to it (if the path is already contained, may be repeated without addition), and then "OK." 
At this point, CUDA installed.
Second, the installation cuDNN:
1, cuDNN download, download need to register the class, is to facilitate the download, also uploaded to the cloud Baidu easy for everyone to download.
Official Download: https://developer.nvidia.com/rdp/cudnn-download
2, extracting archive, the archive in the bin, include, lib files are copied to the C: \ Program Files \ NVIDIA GPU Computing Toolkit \ corresponds to the CUDA \ v9.0 directory directory
3, the C: \ Program Files \ NVIDIA GPU Computing Toolkit \ CUDA \ v9.0 \ extras \ CUPTI \ libx64 \ cupti64_80.dll
copied into C: \ Program Files \ NVIDIA GPU Computing Toolkit \ CUDA \ v9.0 \ bin
At this point, cuDNN installed.
Third, the installation-GPU TensorFlow
the CPU version if installed before the TensorFlow can use the command pip uninstall tensorflow to uninstall.
Open a command prompt, use the command pip install tensorflow-gpu==1.12.0 for installation.
