Windows 10 安装TensorFlow-GPU的踩坑&解决过程（新建虚拟环境）

install tf for win10

说明：
1. 此台电脑上装了anaconda，下列教程对比完整安装TensorFlow-GPU，缺少的部分是：安装anaconda，安装cuda+cudnn（win10下这两步挺简单的）
2. 我的cuda安装了10.2版本，tensorflow-gpu==2.1 支持的cuda是10.1，这个问题在后面解决了（至少表面上）……
3. 另外，由于anaconda的base环境里已经装好了 torch，我想把 torch 和 tf 分开，所以新建虚拟环境。如果直接安装 tf，请省略Step1,2。
4. 中间遇上的不少问题也是在网上找到了别人的方法……如果出现一些错误请指正~本人挺菜的……

总过程一共花了1个小时多点吧

Step1: 新建虚拟环境

打开cmd窗口
conda create -n tf_gpu python=3.7

conda activate tf_gpu
如果出现下述报错，(按提示)执行conda init cmd.exe
退出，打开新的cmd窗口即可

CommandNotFoundError: Your shell has not been properly configured to use 'conda activate'.
If using 'conda activate' from a batch script, change your invocation to 'CALL conda.bat activate'.

Step2: 出现问题pip 无法安装

(tf_gpu)  pip install nltk     
WARNING: pip is configured with locations that require TLS/SSL, however the ssl module in Python is not available.
中间省略报错
/simple/pip/ (Caused by SSLError("Can't connect to HTTPS URL because the SSL module is not available.")) - skipping

但我新建的虚拟环境里有openssl
所以……去对比了一下我的base环境，发现少了两个和ssl有关的包

决定先conda install pyopenssl 装完还是不行

conda install service_identity

ok pip正常

(tf_gpu) PS C:\Users\3\D\2019-master> 
pip install --upgrade pip
Requirement already up-to-date: pip in e:\anaconda\lib\site-packages (20.0.2)
(tf_gpu) PS C:\Users\30322\Desktop\BallClustering_ICCV2019-master>

Step3: 正式安装tf

官网：https://tensorflow.google.cn/install

感觉官方教程不太靠谱呢……pip install tensorflow这应该是默认cpu版的安装命令。

所以决定pip install tensorflow-gpu，不加版本号，试试会不会自动装最新的

嗯确实安装最新版了……就是这网速……

(tf_gpu) PS C:\Users\30322\Desktop\BallClustering_ICCV2019-master> 
pip install tensorflow-gpu
Collecting tensorflow-gpu
  Downloading tensorflow_gpu-2.1.0-cp37-cp37m-win_amd64.whl (356.5 MB)
     |                                | 245 kB 5.5 kB/s eta 17:55:29

然后我试了下使用已经添加好了国内源的conda: conda install tensorflow-gpu==2.1

发现又没有conda识别出我装好的的cuda……

找到了一个有2.1版本的源，速度很快~ pip install tensorflow-gpu==2.1 -i https://pypi.douban.com/simple

(tf_gpu) PS C:\Users\3\D\2019-master>
pip install tensorflow-gpu==2.1 -i https://pypi.douban.com/simple
Looking in indexes: https://pypi.douban.com/simple
Collecting tensorflow-gpu==2.1
  Downloading https://pypi.doubanio.com/…………………………………………
     |███                             | 32.4 MB 2.2 MB/s eta 0:02:29

但是：

ERROR: google-auth 1.11.2 has requirement setuptools>=40.3.0, but you'll have setuptools 40.2.0 which is incompatible.

哪些版本低了更新就好。pip install --upgrade setuptools

ERROR: Cannot uninstall 'wrapt'. It is a distutils installed project and thus we cannot accurately determine which files belong to it which would lead to only a partial uninstall.

查到的：pip install -U --ignore-installed wrapt enum34 simplejson netaddr

夜深了不愿意去查这句命令后几个包具体是啥以及是不是必须（其实就是菜）

结果：

Successfully installed cachetools-4.0.0 google-auth-1.11.2 google-auth-oauthlib-0.4.1 keras-preprocessing-1.1.0 markdown-3.2.1 oauthlib-3.1.0 opt-einsum-3.1.0 requests-2.22.0 requests-oauthlib-1.3.0 rsa-4.0 scipy-1.4.1 tensorboard-2.1.0 tensorflow-gpu-2.1.0

检验安装成功

进入python环境，import tensorflow as tf

Could not load dynamic library 'cudart64_101.dll'; dlerror: cudart64_101.dll not found

完蛋我装的是cuda10.2……torch也是官网说支持10.1但也能在10.2环境运行，为啥TF这么苛刻啊啊23333

再见，人间不值得

咳咳

上网一查发现神奇操作

把C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v10.2\bin中的 cudart64_102.dll 复制改名为 cudart64_101.dll 粘贴，然后就可以import了
检查是否成功的代码见下：

>>>import tensorflow as tf
>>> tf.config.list_physical_devices('GPU')
2020-02-21 00:52:26.690740: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1555] Found device 0 with properties:
pciBusID: 0000:01:00.0 name: GeForce GTX 1050 Ti computeCapability: 6.1
coreClock: 1.62GHz coreCount: 6 deviceMemorySize: 4.00GiB deviceMemoryBandwidth: 104.43GiB/s
2020-02-21 00:52:26.730078: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cudart64_101.dll
2020-02-21 00:52:26.749611: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cublas64_10.dll
2020-02-21 00:52:26.762755: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cufft64_10.dll
2020-02-21 00:52:26.767525: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library curand64_10.dll
2020-02-21 00:52:26.791410: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cusolver64_10.dll
2020-02-21 00:52:26.796559: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cusparse64_10.dll
2020-02-21 00:52:26.800775: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Successfully opened dynamic library cudnn64_7.dll
2020-02-21 00:52:26.818404: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1697] Adding visible gpu devices: 0
[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]

为了证明上面的改名操作对torch没有影响

>>> import torch
>>> print(torch.cuda.is_available())
True
>>>

大概装了这么多，估计以后一段时间内还是用CPU计算的，用GPU的话，我这1050Ti也撑不住呢

等待疫情结束回去用Ubuntu吧~