Prerequisite:

Install Ubuntu 16.04 with nvidia graphics card

turn off BIOS Secure Boot and Fast Boot

Install CUDA 9.2

Download latest CUDA Toolkit runfile from here.

The file will be like cuda_9.2.148_396.37_linux.run

9.2 is the cuda version. 396.37 is the version of the bundled nvidia graphics driver.

Install using sudo sh cuda_9.2.148_396.37_linux.run

While installing, you can choose to install the bundled driver.

Or you can install the latest compatible graphics driver from the official nvidia ppa

sudo add-apt-repository ppa:graphics-drivers/ppa
sudo apt update
sudo apt install nvidia-396 # should be the same as the bundled driver of cuda.
nvidia-smi # check

Install cuDNN 7.2 for CUDA 9.2

Download cuDNN. Install according to NVIDIA’s documentation:

tar -xvzf cudnn-9.2-linux-x64-v7.2.1.38.tgz
sudo cp cuda/include/cudnn.h /usr/local/cuda/include
sudo cp cuda/lib64/libcudnn* /usr/local/cuda/lib64
sudo chmod a+r /usr/local/cuda/include/cudnn.h /usr/local/cuda/lib64/libcudnn*

Append following lines to /etc/environment:

LD_LIBRARY_PATH="/usr/local/cuda/lib64:/usr/local/cuda/extras/CUPTI/lib64"
CUDA_HOME="/usr/local/cuda"

Download or compile tensorlfow-gpu

We are going to use tensorflow 1.10.1 with cuda 9.2. There is no official released version of tensorflow that supports cuda 9.2, so you can choose to build one yourself or download others’ compiled one such as TensorFlow-wheels

wget https://github.com/evdcush/TensorFlow-wheels/releases/download/tf-1.10.0-gpu-9.2-tensorrt-mkl/tensorflow-1.10.0-cp36-cp36m-linux_x86_64.whl

Install Python (prefer conda, here using miniconda) and tensorlfow-gpu

wget https://repo.continuum.io/miniconda/Miniconda3-latest-Linux-x86_64.sh
sh Miniconda3-latest-Linux-x86_64.sh

remember to export conda path properly (modify ~/.zshrc if you’re using zsh)

create an environment that is compatible to your compiled or downloaded tensorlfow wheel.

conda create -n tensorflow-gpu
source activate tensorflow-gpu

Because tensorflow-1.10.0-cp36-cp36m-linux_x86_64.whl requires python 3.6, we need to install one.

conda install python=3.6 # 

Install tensorlfow-gpu

pip install tensorflow-1.10.0-cp36-cp36m-linux_x86_64.whl

Test if install

python
import tensorflow as tf
hello = tf.constant('Hello, TensorFlow!')
sess = tf.Session()
print(sess.run(hello))

Install OpenAI Gym/baselines

git clone https://github.com/openai/mujoco-py.git
cd mujoco-py
pip install -e .
cd ..
sudo apt install cmake libopenmpi-dev zlib1g-dev
git clone https://github.com/openai/baselines.git
cd baselines
pip install -e .
# If you experience too long time to setup `mujoco`, try using proxy to `pip install`, e.g. proxychains pip install -e .

*Notes

(Optional)Using thrid-party ubuntu repository

Replace the content of /etc/apt/sources.list to the below:

deb https://mirrors.tuna.tsinghua.edu.cn/ubuntu/ xenial main restricted universe multiverse
# deb-src https://mirrors.tuna.tsinghua.edu.cn/ubuntu/ xenial main restricted universe multiverse
deb https://mirrors.tuna.tsinghua.edu.cn/ubuntu/ xenial-updates main restricted universe multiverse
# deb-src https://mirrors.tuna.tsinghua.edu.cn/ubuntu/ xenial-updates main restricted universe multiverse
deb https://mirrors.tuna.tsinghua.edu.cn/ubuntu/ xenial-backports main restricted universe multiverse
# deb-src https://mirrors.tuna.tsinghua.edu.cn/ubuntu/ xenial-backports main restricted universe multiverse
deb https://mirrors.tuna.tsinghua.edu.cn/ubuntu/ xenial-security main restricted universe multiverse
# deb-src https://mirrors.tuna.tsinghua.edu.cn/ubuntu/ xenial-security main restricted universe multiverse

then

sudo apt update
sudo apt upgrade

(Optional)Using thrid-party pip repository

Create/Modify ~/.pip/pip.conf to use third party mirror:

[global]
index-url = https://pypi.tuna.tsinghua.edu.cn/simple

(Optional)Using Shadowsocksrr to run command through proxy

sudo apt install git
git clone https://github.com/shadowsocksrr/shadowsocksr.git
cd shadowsocksr
bash initcfg.sh
nano user-config.json # server address etc.
sudo nano /lib/systemd/system/shadowsocksr.service

Append following lines to /lib/systemd/system/shadowsocksr.service:

[Unit]
Description=Shadowsocks Client Service
After=network.target

[Service]
Type=simple
User=nobody
ExecStart=/usr/bin/python /home/[username]/shadowsocksr/shadowsocks/local.py -c /home/[username]/shadowsocksr/user-config.json

[Install]
WantedBy=multi-user.target

Change the [username] to your user name.

sudo systemctl enable shadowsocksr.service
sudo systemctl start shadowsocksr.service

Then install proxychains to proxy any command through socks5.

sudo apt-get install proxychains

Edit proxychains config file /etc/proxychains.conf

...
[ProxyList]
# add proxy here ...
# meanwile
# defaults set to "tor"
socks5  127.0.0.1 1080

Run like this:

proxychains git clone ...

Also you can install polipo to turn socks5 into a http proxy. Then export environment variable http_proxy,https_proxy and no_proxy to enable global http proxy.