Could not load dynamic library cudart64_100.dll

使用环境：

• TensorFlow==1.15.0 Keras==2.3.1 cuda==10.1 cudnn==7.6.5
  **在玩bert的时，报：

2020-10-11 22:51:52.067448: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Could not load dynamic library 'cudart64_100.dll'; dlerror: cudart64_100.dll not found
2020-10-11 22:51:52.067648: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Could not load dynamic library 'cudart64_100.dll'; dlerror: cudart64_100.dll not found
2020-10-11 22:51:52.073182: W tensorflow/stream_executor/platform/default/dso_loader.cc:55] Could not load dynamic library 'cublas64_100.dll'; dlerror: cublas64_100.dll not found
2020-10-11 22:51:52.078544: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Could not load dynamic library 'cufft64_100.dll'; dlerror: cufft64_100.dll not found
2020-10-11 22:51:52.081144: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Could not load dynamic library 'curand64_100.dll'; dlerror: curand64_100.dll not found
2020-10-11 22:51:52.086796: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Could not load dynamic library 'cusolver64_100.dll'; dlerror: cusolver64_100.dll not found
2020-10-11 22:51:52.090414: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] Could not load dynamic library 'cusparse64_100.dll'; dlerror: cusparse64_100.dll not found
2020-10-11 22:51:52.100478: I tensorflow/stream_executor/platform/default/dso_loader.cc:44] 
Could not load dynamic library 'cudnn64_7.dll'; dlerror: cudnn64_7.dll not found

经过查找，发现其实cuda10.2中包含了cudart64_100.dll的向下兼容。网上说要将cudart64_100.dll的路径添加到环境变量中或更改cudart64_101.dll的名字。各种不行。最后在看到参考中提到可以用Anaconda，果断采用，最后确定方案是：

• 【1】win10 CUDA安装为CUDA10.1(因为TF2.x系列也就刚支持到CUDA10.1)
• 【2】pytorch1.6.0+cu101 和 TF2.3.1安装CUDA10.1版本的到纯净Python37(目录C:\Python37)环境下
• 【3】纯净Python37 python执行方式：C:\Python37\python
• 【4】纯净Python37 pip执行方式：C:\Python37\Scripts\pip list

【特别注意】Tensorflow版本关系可参考TensorFlow的一些版本约束, Pytorch版本关系可参考PyTorch一些trick

2 Anaconda 3.7

前文说道，为了解决tf1.15.0 cuda pytorch的不兼容问题，特别适用Anaconda来创建一个虚拟环境,conda方式安装cuda库(本机依然是cuda10.1)。

2.1 镜像下载

官网各版本Python对应关系https://docs.anaconda.com/anaconda/packages/oldpkglists/

官网下载仓库https://repo.anaconda.com/archive/
官网下载https://www.anaconda.com/distribution/
清华镜像https://mirrors.tuna.tsinghua.edu.cn/anaconda/archive

这里下载的是Anaconda 2020.02 for 64-bit Windows with Python 3.7

anaconda仓库 https://anaconda.org/anaconda/repo

2.2 安装cuda10.1和纯Python3.7

win10下安装cuda10.1和cudnn
win10下安装Python3.7

2.3 配置conda源

#添加清华的源【推荐使用】
conda config --add channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free
#中科大的源
conda config –add channels https://mirrors.ustc.edu.cn/anaconda/pkgs/free/ 
#阿里云的源
conda config --add channels http://mirrors.aliyun.com/pypi/simple/
#移除源
conda config --remove channels https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/

#显示当前源
conda config --show-sources
# 设置搜索时显示通道地址
conda config --set show_channel_urls yes

还有一种方式【在C:\Users\Anjos.condarc下直接 修改其值也可以,cmd下直接就能获取到】：

channels:
  - https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/free/
  - defaults
show_channel_urls: true

2.4 配置环境变量

将：C:\ProgramData\Anaconda3\Scripts、C:\ProgramData\Anaconda3\Library\bin、C:\ProgramData\Anaconda3添加到系统PATH环境变量中

2.5 虚拟环境

Warning:
This Python interpreter is in a conda environment, but the environment has
not been activated.  Libraries may fail to load.  To activate this environment
please see https://conda.io/activation

需要**当前环境即可。

# 搜索库和指定源安装 或 conda search cudatoolkit
anaconda search -t conda cudatoolkit
anaconda show mwojcikowski/cudatoolkit
conda install --channel https://conda.anaconda.org/mwojcikowski cudatoolkit=10.0

命令	含义
【环境检查和配置】
conda config --show-sources	显示当前源
conda config --set show_channel_urls yes	设置搜索时显示通道地址
conda env list	查看已安装的环境
conda info --envs
conda info -e
conda activate base	**名称为base的环境 也可以用绝对路径C:\ProgramData\Anaconda3
conda deactivate	退出环境
conda create -n tf115 python=X.X(2.7、3.6等)	创建环境tf115在C:\Users\Anjos\.conda\envs下可查【-n表示-name】
conda create -n tf115 --clone base	创建环境(基于已有环境py3.7)
conda remove -n tf115 --all	删除环境
【conda方式安装库】(若在当前环境下可不需要-n选项)
conda list	查看当前环境下已安装包
conda list -n tf115	查看环境名为tf115下已安装包
conda search numpy	搜索numpy库,在当前源下
conda install -n tf115 numpy	安装numpy库
conda remove/uninstall -n tf115 numpy	移除numpy库
conda install pip	给当前环境安装pip,以后就用pip安装库了
【pip方式安装库】【推荐这种安装,conda顶多帮安装cudnn和cuda】
pip install tensorflow-gpu==1.15.0　-i http://mirrors.aliyun.com/pypi/simple	安装TensorFlow库
pip可直接配置统一的源即可,配置方法见配置源

2.6 idea中使用

定为他即可。C:\ProgramData\Anaconda3\python.exe

也可以借助pycharn和idea自动识别Anaconda3并创建虚拟环境，借助base环境

2.7 安装TensorFlow1.15.0

conda create -n tf115 --clone base
conda activate tf115

conda install cudatoolkit=10.0
conda install cudnn=7.6.5
pip3 install tensorflow-gpu=1.15.0

# tf1.x 检查GPU是否可用
import tensorflow as tf
hello= tf.constant('Hello, TensorFlow!')
print(tf.Session().run(hello))
print(tf.test.is_gpu_available())

# tf2.x 检查GPU是否可用
import tensorflow as tf
print(tf.test.is_built_with_cuda())
print(tf.test.is_gpu_available(cuda_only=False,min_cuda_compute_capability=None))

# pytorch检查gpu是否可用【这里pytorch就采用纯净不安装了】
import torch
print(torch.cuda.is_available())

【纯净环境下cuda10.1+tf2.3.0】
2020-10-12 22:12:55.759800: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cudart64_101.dll
2020-10-12 22:12:55.770142: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cublas64_10.dll
2020-10-12 22:12:55.778848: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cufft64_10.dll
2020-10-12 22:12:55.785005: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library curand64_10.dll
2020-10-12 22:12:55.794067: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cusolver64_10.dll
2020-10-12 22:12:55.802166: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cusparse64_10.dll
2020-10-12 22:12:55.815182: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cudnn64_7.dll
2020-10-12 22:12:55.819302: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1858] Adding visible gpu devices: 0
2020-10-12 22:12:56.362918: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1257] Device interconnect StreamExecutor with strength 1 edge matrix:
2020-10-12 22:12:56.367031: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1263]      0
2020-10-12 22:12:56.368975: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1276] 0:   N

【conda tf115环境下cuda10.0+tf1.15.0】
2020-10-12 22:14:59.901777: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cudart64_101.dll
2020-10-12 22:15:12.576051: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cudart64_101.dll
2020-10-12 22:15:12.581228: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cublas64_10.dll
2020-10-12 22:15:12.584647: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cufft64_10.dll
2020-10-12 22:15:12.588276: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library curand64_10.dll
2020-10-12 22:15:12.592404: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cusolver64_10.dll
2020-10-12 22:15:12.596107: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cusparse64_10.dll
2020-10-12 22:15:12.600847: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library cudnn64_7.dll
2020-10-12 22:15:12.606824: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1858] Adding visible gpu devices: 0
2020-10-12 22:15:13.177512: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1257] Device interconnect StreamExecutor with strength 1 edge matrix:
2020-10-12 22:15:13.181737: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1263]      0
2020-10-12 22:15:13.184208: I tensorflow/core/common_runtime/gpu/gpu_device.cc:1276] 0:   N

Reference

在CUDA10.1版本下安装tensorflow-gpu=1.15版本
Conda命令 + 安装tensorflow