Detailed steps for tensorflow-gpu2.4.1 installation and configuration

I. Machine environment

Graphics card: RTX3060

System: windows10

Second, tensorflow-gpu configuration environment

CUDA:11.1.1

cuDNN:8.1.1

Language: Python3.8.8

Deep learning framework: Tensorflow-gpu2.4.1, Keras2.4.3

III. Installation process

1. Install Visual studio 2017.Compile dependencies.Check Desktop development using C++ when selecting the workload.Download address: https://visualstudio.microsoft.com/zh-hans/thank-you-downloading-visual-studio/?sku=Community&rel=15

2. Install CUDA.The CUDA11.1.1 version is installed here.File name: cuda_11.1.1_456.81_win10.exe.There are about 3.1G.After downloading, double-click to install it. It is installed to the C drive by default and takes up little space.Download address: https://developer.nvidia.com/cuda-toolkit-archive

3. Download cudnn.An account login is required, if you don't have one, register one.The download here is cuDNN8.1.1.File name: cudnn-11.2-windows-x64-v8.1.1.33.zip.There are about 661M.After decompression, there are three folders bin, include, lib.Copy the files in the three folders to the corresponding folder in the CUDA installation directory.My CUDA installation directory is: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.1.Take the bin folder as an example, copy all the files in the bin to C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.1\bin.Download address: https://developer.nvidia.com/zh-cn/cudnn

The above three are installed and restart the computer.

4. Install python.Installed here is python3.8.8.File name: python-3.8.8-amd64.exe.About 27M.Download address: https://www.python.org/getit/

5. Install tensorflow-gpu.Install using pip.Command line: pip install tensorflow-gpu==2.4.1 -i https://pypi.tuna.tsinghua.edu.cn/simple/ It is faster to install using Tsinghua image.

6. Install keras.Install using pip.Command line: pip install keras==2.4.3 -i https://pypi.tuna.tsinghua.edu.cn/simple/ It is faster to install using Tsinghua mirror.

Fourth, test.

import tensorflow as tfprint(tf.__version__)print(tf.test.is_built_with_cuda())print(tf.test.is_gpu_available())

First, bug handling.

When testing print(tf.test.is_gpu_available()) print is False, prompt: Not creating XLA devices, tf_xla_enable_xla_devices not set

Workaround: Find cusolver64_11.dll in the bin folder of the CUDA installation directory and rename it to cusolver64_10.dll.After renaming, test again and find that print(tf.test.is_gpu_available()) prints True, but there is still a Not creating XLA devices, tf_xla_enable_xla_devices not set prompt.This does not affect usage.

Second, test print information description:

Created TensorFlow device (/device:GPU:0 with 10491 MB memory) -> physical GPU (device: 0, name: NVIDIA GeForce RTX 3060, pci bus id: 0000:01:00.0, compute capability: 8.6)

True

Graphics card information: NVIDIA GeForce RTX 3060, graphics card computing power: 8.6, GPU availability: True

tips:

a. Enter nvcc -V in the terminal to view the cuda version.

b. Enter nvidia-smi in the terminal to check the graphics card driver version, the highest supported cuda version, and the graphics card usage.

c.python IDE tool, I personally recommend PyCharm.

d. The installation process of other graphics cards is similar, but pay attention to the cuda version supported by the graphics card and the versions of cuda, cudnn, python, tensorflow, and keras.