Quickly install torch spark, torch geometric and other packages in moment pool cloud

Rental machine , Choose a basic image according to your environment , What I choose here is Pytorch 1.10.

After the lease is successful, click... On the lease page Jupyterlab link .

Jupyterlab I'll build a new one Terminal Used to install the environment , First check Pytorch、CUDA edition , As shown in the figure below, you can see my environment Pytorch 1.10.0,CUDA 11.3.

Next install torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric These bags .

• Officially recommended installation method （ The latter method is recommended ）：

pip install torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric -f https://data.pyg.org/whl/torch-{version}+cu{version}.html

Among them, in the official source link torch and cuda The version needs to be changed to the corresponding one in your server environment .

https://data.pyg.org/whl/torch-{version}+cu{version}.html
 Such as  CUDA11.3 Pytorch1.10.0  Corresponding link ：https://data.pyg.org/whl/torch-1.10.0+cu113.html

There may be problems with the official installation method ： because pip If the package version is not specified during installation , The latest version is installed by default , With CUDA11.3 Pytorch1.10.0 install torch-sparse As an example , from Official source You can see torch-sparse The latest version is 0.6.13, But the latest version in the default source of the machine is ：0.6.14, So even if we specify -f Parameters ,pip You will also choose to install updates from the default source 0.6.14（ It's easy to make mistakes , Also need to compile , so much trouble ）.

• Recommended installation method ：

install torch-sparse Specify the version and official source address , In this way, the package will be downloaded and installed from the official source at all times , What is officially provided is that it has been compiled in the corresponding environment whl Installation package , You can avoid the compilation process , Install faster , There will be no unnecessary mistakes .

Get the following installation instructions ( Notice that I am CUDA11.3 Pytorch1.10.0, If you're in another version , Please add the official source torch and cuda Change the version to its own )：

pip install torch-scatter torch-sparse==0.6.13 torch-cluster torch-spline-conv -f https://data.pyg.org/whl/torch-1.10.0+cu113.html

After installation and bread filling , Install again torch-geometric,torch-geometric Not in official sources , So direct pip install Just fine ：

pip install torch-geometric

Be careful ： If there are version requirements , You can specify the version when installing , This makes it easier to install successfully .

After the above installation is completed, you can successfully import the relevant packages .

Refer to ： Official documents