The acceleration is divided into GPU and CPU acceleration, because my graphics card GT630 supports GPU acceleration, so both of these accelerations are used. The caffe interface has MATLAB and Python, which are not used here. It's best to take a look at the pits before loading! ! !
Read before installation:
1. Check whether the graphics card supports cuda and check the computing power of your graphics card. For example, the computing power of GeForceGT630 is 2.1.
URL:https://developer.nvidia.com/cuda-gpus. As long as the graphics card listed here supports cuda.
2. Download the latest driver for your graphics card. URL:http://www.nvidia.com/Download/index.aspx. (I did not do this step, because the driver was already installed when installing cuda. ​​Some people said that they must first install the latest driver for this graphics card. When installing cuda, then choose a custom installation to remove the driver for cuda, it is possible It ’s a different version.)
3. Secondly choose VS version: Visual Studio chooses 2013 is better. Anyway, they are all pits. It is best to follow the steps and find a successful case to install (install VS first, then install CUDA).
4. Select the cuda version. The choice of Cuda is related to your GPU model. Find "NVIDIA Control Panel" in the control panel to open-"System Information"-"Components", it is best to download the cuda version shown here. And remember the difference between windows and linux version when downloading cuda, don't make a mistake. (I directly streamlined the installation, some people said that the installation chooses custom installation, the driver is not selected, because it has already been installed before, why do n’t you know specifically, this is also the experience of previous people.)

5. Download the cuDNN library. link:https://developer.nvidia.com/rdp/cudnn-download. There are instructions for the cuda version (but it is recommended to download the version that has been tested successfully on the Internet).

6. One of the steps in the following installation is very important, the parameters are modified, as shown below:

1) Whether to use GPU acceleration
2) Whether to use cuDNN library and cuda version
3) Whether to use Python interface
4) Whether to use MATLAB interface
5) GPU computing power, (the existing computing power parameters are compute_20, sm_20; compute_30, sm_30; compute_35, sm_35; compute_50, sm_50; compute_52, sm_52). My graphics card GT630 has a computing power of 2.1, so the first choice was compute_21, sm_2, and the compilation was unsuccessful. I chose compute_20, sm_20 to succeed. I did not understand what happened.
6) Path after decompression of cuDNN compressed package
7. Some of the pits of the predecessors, the following is the link, please refer to it yourself.

GTX TITANX graphics card driver and cuda installation blood tear history
There is also a wrong installation of cuda version, uninstall and reinstall: