ubuntu 20.04, cudn10.0, cudnn v7.6, support pytorch 1.2 gpu and tensorflow 1.15 GPU, config record.

0, install nvida driver
    if you had installed nvida driver before, please remove it using the following command:
    sudo apt-get remove --purge nvidia* 

    sudo add-apt-repository ppa:graphics-drivers/ppa
    then, open "software & updates" from "show application" and choose "additional drivers", and choose a fit driver:
    Using NVIDIA driver metapackage from nvidia-driver-440(proprietary)
    choose apply chaneges,
    then restart. 

1, downgrade gcc g++, for if you do not downgrade, cuda_xxx.run wuold not be executed.
    sudo apt-get install gcc-7 g++-7

    sudo ln -s /usr/bin/gcc-7 /usr/bin/gcc
    sudo ln -s /usr/bin/g++-7 /usr/bin/g++
    
    
output:
    gcc --version
    gcc (Ubuntu 7.5.0-6ubuntu2) 7.5.0
    Copyright (C) 2017 Free Software Foundation, Inc.
    This is free software; see the source for copying conditions.  There is NO
    warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

    g++ --version
    g++ (Ubuntu 7.5.0-6ubuntu2) 7.5.0
    Copyright (C) 2017 Free Software Foundation, Inc.
    This is free software; see the source for copying conditions.  There is NO
    warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
    
if system show the soft link exist, but version of gcc and g++ are wrong, please delete the soft link file.
    sudo rm /usr/bin/g++
    sudo rm /usr/bin/gcc
then rebuild the softlink.

2, install cuda and cudnn:

    Download site:

    nvidia CUDA history version CUDA Toolkit Archive | NVIDIA Developer

    cuDNN  history version cuDNN Archive | NVIDIA Developer

    Install commands:
    sudo chmod a+x cuda_10.0.130_410.48_linux.run
    sudo sh cuda_10.0.130_410.48_linux.run
    
    sudo gedit ~/.bashrc
    add follow text:
    "
    export CUDA_HOME=/usr/local/cuda 
    export PATH=$PATH:$CUDA_HOME/bin 
    export LD_LIBRARY_PATH=/usr/local/cuda-10.0/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}
    "
    source ~/.bashrc
    
    test cuda:
    cd /usr/local/cuda/samples/1_Utilities/deviceQuery 
    sudo make
    ./deviceQuery
    
    output:
    deviceQuery, CUDA Driver = CUDART, CUDA Driver Version = 10.2, CUDA Runtime Version = 10.0, NumDevs = 1
    Result = PASS
    
    test cuda version:
    nvcc --version
    
    output:
    nvcc: NVIDIA (R) Cuda compiler driver
    Copyright (c) 2005-2018 NVIDIA Corporation
    Built on Sat_Aug_25_21:08:01_CDT_2018
    Cuda compilation tools, release 10.0, V10.0.130
    
    install cudnn:
    tar -xzvf  cudnn-10.0-linux-x64-v7.6.5.32.tgz cuda/
    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*
    
3, install anaconda3:
    download 
    https://repo.anaconda.com/archive/Anaconda3-2020.07-Linux-x86_64.sh
    bash Anaconda3-2020.07-Linux-x86_64.sh 

 

4, install pytorch1.2, python3.7:
    conda create -n rltorch python=3.7
    conda activate rltorch
    pytorch with gpu support:
    conda install pytorch==1.2.0 torchvision==0.4.0 cudatoolkit=10.0 -c pytorch    #can not install by this command in china.

    get some help in follow site:
    https://pytorch.org/get-started/previous-versions/#via-pip
    https://download.pytorch.org/whl/torch_stable.html
    
    pip install torch==1.2.0 torchvision==0.4.0 -f https://download.pytorch.org/whl/torch_stable.html
    
    other options:
    pip install torch==1.4.0 -f https://download.pytorch.org/whl/cu100/torch-1.4.0%2Bcu100-cp37-cp37m-linux_x86_64.whl
    pip install torchvision==0.4.0 -f https://download.pytorch.org/whl/cu100/torchvision-0.4.0-cp37-cp37m-manylinux1_x86_64.whl --no-deps

5, test pytorch 
import torch as t
x = t.rand(5,3)
y = t.rand(5,3)
if t.cuda.is_available():
       x = x.cuda()
       y = y.cuda()
       print(x+y)

    output:
    ensor([[1.4095, 1.4061, 1.1705],
        [1.6440, 0.6937, 1.0405],
        [0.7109, 0.5343, 1.1778],
        [0.5223, 0.1559, 1.3047],
        [1.4479, 0.5002, 1.1370]], device='cuda:0')
    
    >>import torch    
    >>print(torch.cuda.get_device_name(0))
    GeForce RTX 2060

6, install gym, mujoco
gym:
    pip install -i https://pypi.tuna.tsinghua.edu.cn/simple gym 

7, structurenet

pip install -r requirements.txt

install PyTorch Scatter
pip install torch-scatter==latest+cu100 -f https://pytorch-geometric.com/whl/torch-1.4.0.html

8,install pytorch-geometric

    pip install torch-scatter==latest+${CUDA} -f https://pytorch-geometric.com/whl/torch-1.4.0.html
    pip install torch-sparse==latest+${CUDA} -f https://pytorch-geometric.com/whl/torch-1.4.0.html
    pip install torch-cluster==latest+${CUDA} -f https://pytorch-geometric.com/whl/torch-1.4.0.html
    pip install torch-spline-conv==latest+${CUDA} -f https://pytorch-geometric.com/whl/torch-1.4.0.html
    pip install torch-geometric    
    
    according to your own cudn version,replace ${CUDA}    
    pip install torch-scatter==latest+cu100 -f https://pytorch-geometric.com/whl/torch-1.4.0.html
    pip install torch-sparse==latest+cu100 -f https://pytorch-geometric.com/whl/torch-1.4.0.html
    pip install torch-cluster==latest+cu100 -f https://pytorch-geometric.com/whl/torch-1.4.0.html
    pip install torch-spline-conv==latest+cu100 -f https://pytorch-geometric.com/whl/torch-1.4.0.html
    pip install torch-geometric    

9, install tensorflow1.15 gpu, keras
    pip install tensorflow-gpu==1.15.0 -i https://pypi.tuna.tsinghua.edu.cn/simple/
    pip install keras==2.3.1  -i https://pypi.tuna.tsinghua.edu.cn/simple/
test:
    import tensorflow as tf
    a = tf.test.is_built_with_cuda()   #判断CUDA是否可用
    b = tf.test.is_gpu_available(cuda_only=False,min_cuda_compute_capability=None)     #判断GPU是否可用
    print(a)
    print(b)
    
10, install pybullet
    pip3 install -i https://pypi.tuna.tsinghua.edu.cn/simple pybullet --upgrade --user
before test the third example "python3 -m pybullet_envs.deep_mimic.testrl --arg_file run_humanoid3d_backflip_args.txt"    
please install:
    sudo apt install libopenmpi-dev
    pip install mpi4py  -i https://pypi.tuna.tsinghua.edu.cn/simple/
else it will give an error of "no module of mpi4py"

11, install mujoco, mujoco_py
    
1.1 mujoco证书秘钥获取
不同身份可获取不同证书 ：正常可获取30天免费试用 MuJoCo 权限，如果是学生，通过学校邮箱可获取更长使用期限的秘钥mjkey.txt和证书License.txt；
获取地址： https://www.roboti.us/license.html，需注意，非学生身份可立即获取邮件，学生身份则需要等待一段时间；
生成计算机id：点击右侧Linux，需要下载一个文件到本地，并执行，过程如下：
cd Downloads
chmod a+x getid_linux (给予执行权限)
./getid_linux  

1.2 mujoco下载
下载 MuJoCo version 2.0 Linux 版本 ：https://www.roboti.us/download/mujoco200_linux.zip
解压下载的 mujoco200 到目录 ~/.mujoco/mujoco200中, 替换license key (邮件附件中的 mjkey.txt) .
按下列流程操作
mkdir ~/.mujoco                                    # 创建文件夹
cp mujoco200_linux.zip的文件路径 ~/.mujoco         # 复制
cd .mujoco                                         # 进入隐藏文件夹
unzip mujoco200_linux.zip                          # 解压
mv mujoco200_linux mujoco200                       # 将mujoco200_linux重命名为mujoco200
cd Downloads                                       # 进入下载将mjkey.txt的文件夹
unzip MuJoCo\ Pro\ Trial.zip            # 

# 复制秘钥到路径
cp mjkey.txt ~/.mujoco/
cp mjkey.txt ~/.mujoco/mujoco200/bin

测试安装是否成功
cd ~/.mujoco/mujoco200/bin
./simulate ../model/humanoid.xml

2. mujoco的环境变量配置
配置环境变量
参考文献：https://www.cnblogs.com/qiuhong10/p/7815943.html

逐步操作

# 注意，路径名需要更改成你自己的路径名，替代“我的计算机名”
gedit ~/.bashrc              # 打开，在打开的窗口最后加入以下2行内容

export MUJOCO_KEY_PATH=~/.mujoco${MUJOCO_KEY_PATH}
export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/home/我的计算机名/.mujoco/mujoco200/bin

source ~/.bashrc             # 生效

3. mujoco_py的安装与配置
参考文献：https://blog.csdn.net/gsww404/article/details/80636676
https://github.com/deepmind/dm_control/commit/3b736812407d821ec5d48bdea08ac53e00d78d2e

这一步容易出错

3.1 mujoco_py下载
在桌面中另开终端，运行：

git clone https://github.com/openai/mujoco-py.git
1
3.2 mujoco_py安装
进入mujoco-py目录
cd mujoco-py
1
利用 pip3 安装requirements中的包
# When updating these, you may need to also update pyproject.toml
glfw>=1.4.0
numpy>=1.11
Cython>=0.27.2
imageio>=2.1.2
cffi>=1.10
fasteners~=0.15

安装mujoco_py
pip3 install -U 'mujoco-py<2.1,>=2.0'
1
遇见错误:
 patchelf 错误
No such file or directory: 'patchelf': 'patchelf'
1
解决办法：

sudo apt-get update -y
sudo apt-get install -y patchelf

 PEP 517错误
ubuntu18 Could not build wheels for mujoco-py which use PEP 517 and cannot be installed directly

解决办法：

pip3 install --no-use-pep517 'mujoco-py<2.1,>=2.0'.

测试
$ python3
>>> import mujoco_py                # 可能会出现一长串字符，等候一会儿
>>> import os
>>> mj_path, _ = mujoco_py.utils.discover_mujoco()
>>> xml_path = os.path.join(mj_path, 'model', 'humanoid.xml')
>>> model = mujoco_py.load_model_from_path(xml_path)
>>> sim = mujoco_py.MjSim(model)
>>> print(sim.data.qpos)

结果1：
测试结果1

>>> sim.step()
>>> print(sim.data.qpos)

结果2
测试结果2

运行mujoco_py中自带的例子

cd ~/mujoco-py/examples/
python3 setting_state.py 

GLEW 错误：（这个错误比较棘手）

 ERROR GLEW initalization error: Missing GL version

按如下操作进行：

sudo apt-get install libglfw3 libglew2.0

将运行程序与显卡进行动态链接

sudo gedit ~/.bashrc                     # 在.bashrc中添加以下内容

export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libGLEW.so:

source ~/.bashrc                         # 执行生效

进入解压好的目录～/mujoco-py/examples，再次运行：

python3 setting_state.py    
    
    
mujoco GLEW initalization error: Missing GL version

Step 1
Run update command to update package repositories and get the latest package information.
sudo apt-get update -y
Step 2
Run the install command with -y

flag to quickly install the packages and dependencies.
sudo apt-get install -y libglew-dev

Got these from : https://zoomadmin.com/HowToInstall/UbuntuPackage/libglew-dev
and then
Step 3:
export LD_PRELOAD=/usr/lib/x86_64-linux-gnu/libGLEW.so    

ERROR
1,ValueError: Object arrays cannot be loaded when allow_pickle=False
!pip install numpy==1.16.1