如何使用最新版本的 CUDA 和 cuDNN 安装 Tensorflow GPU

Question

安装 TensorFlow 进行对象检测有时很烦人，尤其是在通过微调预训练模型开始自己的对象检测项目后发生连线错误时。

如何安装最新的 Tensorflow GPU 支持和最新的 CUDA/CUDNN 而不会出现任何错误？

Answer 1

在 Ubuntu 16.04 上安装 tensorflow-gpu 对象检测 api

升级系统

sudo apt-get update

sudo apt-get upgrade

安装基本包

sudo apt-get install vim curl python-dev gnupg-curl python-tk git

curl https://bootstrap.pypa.io/get-pip.py -o get-pip.py

sudo -H python get-pip.py

安装 tensorflow-gpu（tensorflow 版本=1.14）

sudo -H pip install tensorflow-gpu

安装 CUDA（最终 CUDA 版本 = 10.1）

// Please turn off your secure boot from BIOS

sudo apt-get install gnupg-curl

// Here we install version 10.0 to avoid other issues. Later we can upgrade it to version 10.1
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/cuda-repo-ubuntu1604_10.0.130-1_amd64.deb

sudo dpkg -i cuda-repo-ubuntu1604_10.0.130-1_amd64.deb

sudo apt-key adv --fetch-keys https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1604/x86_64/7fa2af80.pub

sudo apt-get update

wget http://developer.download.nvidia.com/compute/machine-learning/repos/ubuntu1604/x86_64/nvidia-machine-learning-repo-ubuntu1604_1.0.0-1_amd64.deb

sudo apt install ./nvidia-machine-learning-repo-ubuntu1604_1.0.0-1_amd64.deb

sudo apt-get update 

安装 cuDNN（cuDNN 版本 = 7.6.3.30-1，兼容 CUDA 10.1）

sudo apt-get install --no-install-recommends cuda-10-0

// restart your computer here

nvidia-smi.

sudo apt-get install --no-install-recommends libcudnn7=7.6.3.30-1+cuda10.0 libcudnn7-dev=7.6.3.30-1+cuda10.0

sudo apt-get install -y --no-install-recommends libnvinfer5=5.1.5-1+cuda10.0 libnvinfer-dev=5.1.5-1+cuda10.0

sudo apt-get update

// this upgrade command will upgrade your CUDA to version 10.1
sudo apt-get upgrade

sudo apt-get autoremove

安装 TensorFlow 对象检测 api

sudo -H pip install Cython

sudo -H pip install contextlib2

sudo -H pip install pillow

sudo -H pip install lxml

sudo -H pip install jupyter

sudo -H pip install matplotlib


mkdir tensorflow
cd tensorflow
git clone https://github.com/tensorflow/models

// install protocbuf version 3.0.0
wget -O protobuf.zip https://github.com/google/protobuf/releases/download/v3.0.0/protoc-3.0.0-linux-x86_64.zip
unzip protobuf.zip
sudo cp ./bin/protoc /bin/
sudo cp -r ./include/google /usr/local/include/
cd tensorflow/models/research
protoc object_detection/protos/*.proto --python_out=.

git clone https://github.com/cocodataset/cocoapi.git
cd cocoapi/PythonAPI
make
cp -r pycocotools <path_to_tensorflow>/models/research/

导出环境

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-10.1/lib64
export PYTHONPATH=$PYTHONPATH:~/tensorflow/models/research:~/tensorflow/models/research/object_detection/slim

其他

If we install CUDA 10.1 it may have a compatibility issue that libxxxx.so.10.0 is not found when you start training your project.

To solve them:

a. sudo ln -s /usr/lib/x86_64-linux-gnu/libcublas.so.10.1 /usr/local/cuda-10.1/lib64/libcublas.so.10.0

b. sudo ln -s /usr/local/cuda-10.1/lib64/libcudart.so.10.1 /usr/local/cuda-10.1/lib64/libcudart.so.10.0

c. sudo ln -s /usr/local/cuda-10.1/lib64/libcufft.so.10 /usr/local/cuda-10.1/lib64/libcufft.so.10.0

d. sudo ln -s /usr/local/cuda-10.1/lib64/libcurand.so.10 /usr/local/cuda-10.1/lib64/libcurand.so.10.0

e. sudo ln -s /usr/local/cuda-10.1/lib64/libcusolver.so.10 /usr/local/cuda-10.1/lib64/libcusolver.so.10.0

f. sudo ln -s /usr/local/cuda-10.1/lib64/libcusparse.so.10 /usr/local/cuda-10.1/lib64/libcusparse.so.10.0

全部设置