Installation of tensorflow and pytorch frames and CUDA pit records

pytorch

use conda Creating a virtual environment python=3.6 in the future , I found the use of conda install torchvision installed pytorch yes cpu Version of , I won't support it gpu, I made a mistake when running code .

import torch
torch.__version__
torch.cuda.is_available()

here cuda yes False,gpu Unavailable . Then you need to use conda uninstall torchvision uninstall torchvision.
Reuse pip install torchvision -i https://pypi.douban.com/simple install , verification gpu Is it available , Now it's True,gpu You can use .

Above test ,pip Installed torchvision The corresponding will be installed at the same time cuda toolkit package , and conda Installation only supports cpu.

tensorflow

After I tossed about for a long time, the graphics card driver 、cuda、cudnn in the future , Discover the use of conda install tensorflow, Will automatically install the corresponding cuda, verification gpu Is it available , Turned out to be True...

import tensorflow as tf
tf.test.is_gpu_available()

Personal test ：
Don't pretend cuda and cudnn, Direct adoption conda After creating the virtual environment , Use conda install tensorflow-gpu==1.15.0 Will automatically install the corresponding cuda toolkit package , And then verify gpu Is it available , yes True.

About cuda and cudnn

1、 The graphics card driver can be in ubuntu Of Software & Updates Medium additional drivers In the process of replacement .

2、nvidia-smi It shows the driver cuda Version is 11.2. And manually installed cuda toolkit It's different .

see cuda edition
Manual installation will be in usr/local Under the folder .

cat /usr/local/cuda/version.txt

No installation cuda Input nvcc -V command , Installation cuda Of sudo apt-get install command , Use this command to install cuda be not in local Under the folder , And in the usr/lib Under the folder . Entering the above command will show that there is no such directory . Manual access lib Under the folder , Also have version.txt Visible version . This method is not recommended .

see cudnn edition
1） Use tar.gz Installed cudnn

cat /usr/local/cuda/include/cudnn.h | grep CUDNN_MAJOR -A 2

2） Use deb Installed cudnn

dpkg -l | grep cudnn

3、 About 3 The version of the .

drive cuda Version down compatibility cuda toolkit edition . The popular explanation is ,nvidia-smi Displayed drive cuda yes 11.2, The highest support cuda11.2. here tensorflow1.15.0 Corresponding cuda by 10.0,cudnn by 7.4.2.