主要步骤

  1. 安装tensorflow, tensorflow 2.0之后已经不再区分CPU和GPU版本;
  2. 安装cuda;
  3. 添加cudnn库.

tensorflow安装

在正式安装前, 检查是否有msvcp140_1.dll, 如果有安装Visual Studio 2019的话是默认有的, 如果没有或不确定可以从微软官网下载安装, 选择Visual Studio 2015, 2017 and 2019版本即可.

tensorflow可以直接利用pip install tensorflow进行安装, 但为了便于管理, 建议在conda中安装.

conda可以选择anaconda, 从官网下载安装即可.

在最后可以选择是否添加到环境变量, 如果添加到环境变量中, 可以直接从终端(cmd或powershell)打开conda.




cudann与tensorflow版本对应_.net


然后利用conda创建一个tensorflow的环境. 激活后安装tensorflow


#创建名为`tf2`的环境, python版本为3.8(不指定会启用base中的版本)
conda create -n tf2 python=3.8
conda activate tf2
pip install --upgrade pip
pip install tensorflow


检查是否安装成功


import tensorflow as tf
print(tf.__version__)


2020-09-17 09:50:39.712664: W tensorflow/stream_executor/platform/default/ http:// dso_loader.cc:59 ] Could not load dynamic library 'cudart64_101.dll'; dlerror: cudart64_101.dll not found 2020-09-17 09:50:39.718061: I tensorflow/stream_executor/cuda/ http:// cudart_stub.cc:29 ] Ignore above cudart dlerror if you do not have a GPU set up on your machine. 2.3.0

在最后一行, 显示了当前的tensorflow版本. 需要留意的是第一行输出的warning: 缺少cudart64_101.dll. 这个信息在后面会使用到.

cuda安装

cuda可以直接从官网下载, 版本的选择可以参考上一个步骤中的warning信息, cudart64_101.dll表示当前电脑为64位, 需要cuda10.1. 下载对应版本即可.

安装过程没有什么复杂的, 直接下一步也可以. 但对于不希望安装在默认盘的同志, 有一些细节需要注意

  1. 第一次出现的路径选择是选择临时提取路径, 不要被骗了(你猜我被骗了吗?)


cudann与tensorflow版本对应_官网_02


  1. 在安装程序中选择自定义安装, 否则连路径选择的机会都没有


cudann与tensorflow版本对应_tensorflow_03


  1. 在安装选项中, CUDA是必选, 其他的看需求(可以不用)


cudann与tensorflow版本对应_安装vuecli3.0.3指定版本_04


  1. 最后才是安装路径的选择


cudann与tensorflow版本对应_官网_05


然后安装即可. 命令行中测试: 


#测试
nvcc -V


有输出即可: 

nvcc: NVIDIA (R) Cuda compiler driver Copyright (c) 2005-2019 NVIDIA Corporation Built on Sun_Jul_28_19:12:52_Pacific_Daylight_Time_2019 Cuda compilation tools, release 10.1, V10.1.243

添加cudnn库

首先进行测试: 


import tensorflow as tf

physical_devices = tf.config.list_physical_devices('GPU')
print(physical_devices)


这里会输出许多信息: 

2020-09-17 10:20:50.788436: I tensorflow/stream_executor/platform/default/ 
  http:// 
  dso_loader.cc:48 
  ] Successfully opened dynamic library cudart64_101.dll 2020-09-17 10:20:52.545573: I tensorflow/stream_executor/platform/default/ 
  http:// 
  dso_loader.cc:48 
  ] Successfully opened dynamic library nvcuda.dll 2020-09-17 10:20:52.569364: I tensorflow/core/common_runtime/gpu/ 
  http:// 
  gpu_device.cc:1716 
  ] Found device 0 with properties: pciBusID: 0000:01:00.0 name: Quadro K600 computeCapability: 3.0 coreClock: 0.8755GHz coreCount: 1 deviceMemorySize: 1.00GiB deviceMemoryBandwidth: 26.55GiB/s 2020-09-17 10:20:52.593774: I tensorflow/stream_executor/platform/default/ 
  http:// 
  dso_loader.cc:48 
  ] Successfully opened dynamic library cudart64_101.dll 2020-09-17 10:20:52.614110: I tensorflow/stream_executor/platform/default/ 
  http:// 
  dso_loader.cc:48 
  ] Successfully opened dynamic library cublas64_10.dll 2020-09-17 10:20:52.631976: I tensorflow/stream_executor/platform/default/ 
  http:// 
  dso_loader.cc:48 
  ] Successfully opened dynamic library cufft64_10.dll 2020-09-17 10:20:52.639591: I tensorflow/stream_executor/platform/default/ 
  http:// 
  dso_loader.cc:48 
  ] Successfully opened dynamic library curand64_10.dll 2020-09-17 10:20:52.648180: I tensorflow/stream_executor/platform/default/ 
  http:// 
  dso_loader.cc:48 
  ] Successfully opened dynamic library cusolver64_10.dll 2020-09-17 10:20:52.667679: I tensorflow/stream_executor/platform/default/ 
  http:// 
  dso_loader.cc:48 
  ] Successfully opened dynamic library cusparse64_10.dll 2020-09-17 10:20:52.677125: W tensorflow/stream_executor/platform/default/ 
  http:// 
  dso_loader.cc:59 
  ] Could not load dynamic library 'cudnn64_7.dll'; dlerror: cudnn64_7.dll not found 2020-09-17 10:20:52.692602: W tensorflow/core/common_runtime/gpu/ 
  http:// 
  gpu_device.cc:1753 
  ] Cannot dlopen some GPU libraries. Please make sure the missing libraries mentioned above are installed properly if you would like to use GPU. Follow the guide at 
  https://www. 
  tensorflow.org/install/ 
  gpu 
   for how to download and setup the required libraries for your platform. Skipping registering GPU devices...

其中info级别的信息可以忽略, 重点关注warning, 可以看到, 缺少cudnn64_7.dll动态库, 这表示我们需要下载cudnn库, 其版本为7.

然后从官网下载对应版本, 即对应cuda10.1版本的cudnn库, 其大版本应该为7.


cudann与tensorflow版本对应_.net_06


下载后解压, 可以直接把里面的文件拷贝到cuda安装目录下. 但更推荐直接添加环境变量, 指向<你的cudnn库路径>cudnn-10.1-windows10-x64-v7.6.5.32cudabin

再次运行


import tensorflow as tf

physical_devices = tf.config.list_physical_devices('GPU')
print(physical_devices)


查看可用的GPU设备