Introduction : The latest CUDA version is 10.1, but for some mysterious reason, tensorflow2.0 currently only supports CUDA10.0. This has been confirmed in my computer and some users feedback. tensorflow2.0 not only a bundled version of CUDA, CUDA installation configuration also has certain requirements. In order to achieve GPU accelerated, where the installation and configuration process of recording it.
Download and install CUDA10.0
CUDA offer a variety of download links: https://developer.nvidia.com/cuda-toolkit-archive
because we have to CUDA10.0, so choose the version when we must choose CUDA Toolkit 10.0 , Note 10.0, not 10.1 .
As shown above, the computer system according to their type and the number of bits to select the appropriate version can then download. A total of more than 2G, but Nvidia good server, download speed is quite fast.
Download down like this on the map, this is actually a decompression program. Advance is best to first build a folder to extract the complete installer put inside.
As shown above, I created folder cuda10.0 , run Setup introducing its path.
When the installation program will automatically run after decompression is finished, the map window will automatically pop up, click on Agree and Continue .
Then be sure to pay attention, otherwise easy to roll! As shown above, you must choose a custom installation .
Pay attention to the first step : try to components NVDIA GeForce Experience option checked off. Some people may know this, which does not adjust the parameters of the game screen thing. Ahem, this is available to the user NVDIA something similar to the control panel, but gave a general computer installed at present installed; and even if not, this thing with our graphics acceleration program is not helpful.
Step stress : the component CUDA willVisual Studio Intergration cancel. This is very interesting, because this component is specially configured to VS, but now use code compilers are basically vscode, pycharm and other lightweight compiler and his ilk, with relatively VS did not make sense; I saw on the Internet a statement that this must cancel, or not installed correctly. Anyway, canceled get away.
Pay attention to the third step : CUDA in the installation of the graphics driver installed by default in a NVDIA. The figure can be seen that the graphics driver version is 411.31. Plan focus on : "The current version of" This column will show you the model number on the computer video card driver present, such as shown in the figure, my computer drive is currently 430.86, a figure greater than 411.31, indicating that the update version number. If your current version is higher than the new version of the CUDA want to install, you must take the display driver cancel ! Otherwise, find your existing installation CUDA also advanced than I want security, I am also a safety hammer, improper installation failure occurs; if your current version is lower than the new version of the CUDA want to install, you must take the display check the driver , the driver version is not new, CUDA can not run. If the coincidence of the current version with the new version of the same, that hook is not checked at random .
OK installed, check nvcc.exe we go on the path shown in Fig exists, this program is very important.
Then check the path on the map in this dynamic library exists, this dynamic library is also very important.
Installation cudnn:
First offer URL: https://developer.nvidia.com/rdp/cudnn-archive
As shown above, to select for CUDA 10.0 versions
then select the appropriate version of their own system.
Downloaded is compressed, decompressed finished as shown above, called CUDA , there are three folders. We need to put its name from cuda changed cudnn , yes, that is renamed.
OK, then the renamed folder in the above figure copy and paste it onto the path like FIG.
Check here a little bit, cudnn64_7.dll this dynamic link inventory No, this library is also very important.
Environment variable configuration
Start configuration environment variable! Right-click on " My Computer " and select " Properties ", select " Advanced System Settings ", the next window as shown above. Next select " System Variables " inside the Path , click " Edit ."
As shown above, as CUDA has been laid, the path where the two have been added on the CUDA. But not enough, we also need to add two.
New environment variable method is very simple, click New , and then click Browse . Shown above according to the CUPTI \ libx64 path introduced. Similarly would then create a cudnn \ bin path introduced.
After importing We need to click on the move , the four path environment variable on top of CUDA, as these four sequential touches nothing. Click OK to exit.
Open a command-line tool, enter nvcc -V . As shown above, the configuration is successful.
After installing the FIG tensorflow2.0 can follow the input command statements, see the results shown above is completely OK.
