TensorFlow 2.10.0已于近日发布，但是目前网上鲜有该版本的安装教程，且官方测试的Python、CUDA、cuDNN版本配置没有更新（截至本文发表前，更新至2.6.0）。故本文对TensorFlow 2.10.0在Anaconda安装所需依赖库版本确定方法进行阐述，其完整安装教程可参考这里

1 依赖版本组合

先直接放出经过我或网友测试过的依赖版本组合，安装方法与其他版本大同小异。

2 系统环境

下面是测试过的系统环境

3 依赖版本确定方法

3.1推理法

3.1.1 TensorFlow依赖范围

假设目前已经安装完成TensorFlow本体，而未安装GPU依赖，运行任一调用Tensorflow的程序，如：

python											#进入python环境
import tensorflow as tf			#导入tensorflow

则打印的信息如下所示，可以看到其调用的cuda版本为11，cudnn版本为8，故由此判cuda 11.*和cudnn 8.*等所有版本为我们接下来考虑的范围。

...
... Could not load dynamic library 'libcudart.so.11.0' ...
...
... Could not load dynamic library 'libcudnn.so.8' ...
...

3.1.2 显卡驱动支持范围

打印显卡相关信息

nvidia-smi

打印如下图，可见本却东最高支持cuda 11.4

3.1.3 查阅官方测试表

由以上方法可以判断，CUDA版本只能在11.0~11.4之间选择，我们从官方测试过的版本组合得出最新版TensorFlow 2.6.0使用CUDA 11.2和cuDNN8.1。故CUDA版本可以11.2~11.4之间选择。

3.1.4 Anaconda自动确定

查询conda中cuDNN对CUDA版本的支持范围

conda search cudnn --info

从打印信息中可以找到如下信息，说明cudnn 8.1.0.77 h90431f1_0，支持- cudatoolkit 11.*

cudnn 8.1.0.77 h90431f1_0
-------------------------
file name   : cudnn-8.1.0.77-h90431f1_0.tar.bz2
name        : cudnn
version     : 8.1.0.77
build       : h90431f1_0
build number: 0
size        : 634.8 MB
license     : cuDNN Software License Agreement
subdir      : linux-64
url         : https://conda.anaconda.org/conda-forge/linux-64/cudnn-8.1.0.77-h90431f1_0.tar.bz2
md5         : 7b8da042080da30d6d1c00d2924d3f7c
timestamp   : 2021-02-24 23:51:26 UTC
dependencies: 
  - __glibc >=2.17,<3.0.a0
  - cudatoolkit 11.*
  - libgcc-ng >=3.0
  - libstdcxx-ng >=3.4

故暂定安装cuda 11.2

conda install cudatoolkit=11.2 -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/linux-64/

接下来直接进行安装cuDNN 8.1，conda会自动选择合适的小版本进行安装

conda install cudnn=8.1 -c https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/linux-64/

综上，我们完成Tensorlow的依赖环境库版本的确认。

3.2 Docker法

Docker 是在 Linux 上启用 TensorFlow GPU 支持的最简单方法，因为只需在主机上安装 NVIDIA® GPU 驱动程序，而不必安装 NVIDIA® CUDA® 工具包【引用自官网安装教程】。基于此我们可以先安装TensorFlow的Docker版本，在从该Docker容器中确定官方使用的依赖库版本。

由于Docker、NVIDIA® GPU 驱动程序和nvidia-container-toolkit之前已经安装（读者如果没有安装可以参考官网安装教程），故此处直接安装TensorFlow，安装之前需要对网络进行合理处理（可以尝试*墙、手机共享网络、切换有线网络、切换无线网络），或者换为国内源。

docker pull tensorflow/tensorflow:latest-gpu

测试

docker run --gpus all -it --rm tensorflow/tensorflow:latest-gpu    python -c "import tensorflow as tf; print(tf.reduce_sum(tf.random.normal([1000, 1000])))"

打印如下信息，证明成功

Created device /job:localhost/replica:0/task:0/device:GPU:0 with 6383 MB memory:  -> device: 0, name: NVIDIA GeForce RTX 2070 with Max-Q Design, pci bus id: 0000:01:00.0, compute capability: 7.5

此时进入tensorflow:latest-gpu的Docker环境

docker run --gpus all -it --rm tensorflow/tensorflow:latest-gpu

并在该Docker环境容器中运行

nvcc --version

可以看到打印信息如下，从中得出其使用Cuda 11.2.152

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2021 NVIDIA Corporation
Built on Sun_Feb_14_21:12:58_PST_2021
Cuda compilation tools, release 11.2, V11.2.152
Build cuda_11.2.r11.2/compiler.29618528_0

运行并打印，可得其 cuDNN版本为8.1.0.77

root@5b6303e8f6c7:~# dpkg -l | grep cudnn
ii  libcudnn8                     8.1.0.77-1+cuda11.2               amd64        cuDNN runtime libraries

进入Python，并打印，故，该Tensorflow Docker镜像中使用Python 3.8.10

python
Python 3.8.10 (default, Jun 22 2022, 20:18:18) 
[GCC 9.4.0] on linux

在python命令行运行如下指令并打印如下，说明此TensorFlow版本为2.10.0

>>> import tensorflow as tf
2022-10-03 06:52:40.225921: I tensorflow/core/platform/cpu_feature_guard.cc:193] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 FMA
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
2022-10-03 06:52:40.356008: E tensorflow/stream_executor/cuda/cuda_blas.cc:2981] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
>>> tf.__version__
'2.10.0'

至此，将所有官方Docker使用的依赖库版本确定完毕。

4 测试

然后测试TensorFlow可否正常使用GPU，运行

python			#进入python环境

打印

Python 3.8.13 | packaged by conda-forge | (default, Mar 25 2022, 06:04:18) 
[GCC 10.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.

运行

>>> import tensorflow as tf			#导入tensorflow

打印

2022-10-03 16:56:23.526783: I tensorflow/core/platform/cpu_feature_guard.cc:193] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 FMA
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
2022-10-03 16:56:23.651615: E tensorflow/stream_executor/cuda/cuda_blas.cc:2981] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
2022-10-03 16:56:24.194123: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer.so.7'; dlerror: libnvinfer.so.7: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /home/bit202/Programs/anaconda3/envs/tf2.x/lib:/home/bit202/ROS_WS/DGS_Paper/devel/lib:/home/bit202/Programs/Tools_RosBag2KITTI/catkin_ws/devel/lib:/home/bit202/ROS_WS/learning/devel/lib:/home/bit202/ROS_WS/Delta5BA_WS/devel/lib:/home/bit202/ROS_WS/Delta_WS/devel/lib:/home/bit202/catkin_ws/devel/lib:/opt/ros/melodic/lib
2022-10-03 16:56:24.194196: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer_plugin.so.7'; dlerror: libnvinfer_plugin.so.7: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /home/bit202/Programs/anaconda3/envs/tf2.x/lib:/home/bit202/ROS_WS/DGS_Paper/devel/lib:/home/bit202/Programs/Tools_RosBag2KITTI/catkin_ws/devel/lib:/home/bit202/ROS_WS/learning/devel/lib:/home/bit202/ROS_WS/Delta5BA_WS/devel/lib:/home/bit202/ROS_WS/Delta_WS/devel/lib:/home/bit202/catkin_ws/devel/lib:/opt/ros/melodic/lib
2022-10-03 16:56:24.194207: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Cannot dlopen some TensorRT libraries. If you would like to use Nvidia GPU with TensorRT, please make sure the missing libraries mentioned above are installed properly.

运行并打印，返回True则CUDA正常工作

>>> tf.test.is_built_with_cuda()
True

>>> tf.test.is_gpu_available()

打印，返回True则CUDA正常工作

WARNING:tensorflow:From <stdin>:1: is_gpu_available (from tensorflow.python.framework.test_util) is deprecated and will be removed in a future version.
Instructions for updating:
Use `tf.config.list_physical_devices('GPU')` instead.
2022-10-03 16:56:38.606309: I tensorflow/core/platform/cpu_feature_guard.cc:193] This TensorFlow binary is optimized with oneAPI Deep Neural Network Library (oneDNN) to use the following CPU instructions in performance-critical operations:  AVX2 FMA
To enable them in other operations, rebuild TensorFlow with the appropriate compiler flags.
2022-10-03 16:56:38.648103: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:980] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-10-03 16:56:38.654595: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:980] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-10-03 16:56:38.655430: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:980] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-10-03 16:56:39.071269: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:980] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-10-03 16:56:39.071660: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:980] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-10-03 16:56:39.071971: I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:980] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
2022-10-03 16:56:39.072273: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1616] Created device /device:GPU:0 with 6130 MB memory:  -> device: 0, name: NVIDIA GeForce RTX 2070 with Max-Q Design, pci bus id: 0000:01:00.0, compute capability: 7.5
True

4 报错处理

4.1 无法找到cuda或cudnn依赖库

若TensorFlow运行时依然打印如下报错信息，则说明conda未将依赖库添加到环境变量中，需要手动添加【参考自这里、这里和这里】。

...
... Could not load dynamic library 'libcudart.so.11.0' ...
...
... Could not load dynamic library 'libcudnn.so.8' ...
...

在你anaconda的对应环境路径下（如~/Programs/anaconda3/envs/tf2.x）打开终端运行

mkdir -p ./etc/conda/activate.d									#新建环境变量激活脚本文件夹
touch ./etc/conda/activate.d/activate.sh					#新建环境变量激活脚本
mkdir -p ./etc/conda/deactivate.d								#新建环境变量解除脚本文件夹
touch ./etc/conda/deactivate.d/deactivate.sh		#新建环境变量解除脚本

在activate.sh中添加如下代码，将其中~/Programs/anaconda3/envs/tf2.x/lib替换为你计算机上对应的路径

ORIGINAL_LD_LIBRARY_PATH=$LD_LIBRARY_PATH																			#备份原始环境变量
export LD_LIBRARY_PATH=~/Programs/anaconda3/envs/tf2.x/lib:$LD_LIBRARY_PATH	#添加自定义环境变量

在deactivate.sh添加如下代码

export LD_LIBRARY_PATH=$ORIGINAL_LD_LIBRARY_PATH	#恢复原始环境变量
unset ORIGINAL_LD_LIBRARY_PATH													#删除该定义的环境变量

重新激活虚拟环境，上述环境变量即可生效

4.2 cuBLAS和libnvinfer报错

anaconda环境中如果打印如下信息，则不用担心，这个并不会影响你的运行，这是因为libnvinfer.so.*和libnvinfer_plugin.so.*是TensorRT调用Nvidia GPU时的依赖库【参考这里和这里】，且在官网教程中也出现了同样的打印信息，但其仍然视为正常运行。

2022-10-03 12:12:49.208630: E tensorflow/stream_executor/cuda/cuda_blas.cc:2981] Unable to register cuBLAS factory: Attempting to register factory for plugin cuBLAS when one has already been registered
2022-10-03 12:12:49.745146: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer.so.7'; dlerror: libnvinfer.so.7: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /home/bit202/Programs/anaconda3/envs/tf2.x/lib:/home/bit202/catkin_ws/devel/lib:/opt/ros/melodic/lib
2022-10-03 12:12:49.745223: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'libnvinfer_plugin.so.7'; dlerror: libnvinfer_plugin.so.7: cannot open shared object file: No such file or directory; LD_LIBRARY_PATH: /home/bit202/Programs/anaconda3/envs/tf2.x/lib:/home/bit202/catkin_ws/devel/lib:/opt/ros/melodic/lib
2022-10-03 12:12:49.745234: W tensorflow/compiler/tf2tensorrt/utils/py_utils.cc:38] TF-TRT Warning: Cannot dlopen some TensorRT libraries. If you would like to use Nvidia GPU with TensorRT, please make sure the missing libraries mentioned above are installed properly.

注：cuBLAS 库提供基本线性代数子例程 (BLAS) 的 GPU 加速实现【引用自官网】。如果想要试图解决Unable to register cuBLAS factory ...问题可以参考这里和这里。