Moment Pool Cloud quickly installs packages such as torch-sparse and torch-geometric

Rent a machine and choose an environment according to the environment you need. I chose Pytorch 1.10 here.

Click the Jupyterlab link on the rental page after the rental is successful.

Create a new Terminal in Jupyterlab to install the environment. First check the Pytorch and CUDA versions. You can see my environment Pytorch 1.10.0 and CUDA 11.3 as shown below.

Next install torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric packages.

• The official installation method provided by PyG (there will be errors in some torch versions):

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

The torch and cuda versions in the official source link need to be changed to the corresponding ones in their own server environment.

https://data.pyg.org/whl/torch-{version}+cu{version}.htmlSuch 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: If the version of the package is not specified during pip installation, the latest version will be installed by default. Take CUDA11.3 Pytorch1.10.0 as an example to install torch-sparse, fromOfficial sourceYou can see that the latest version of torch-sparse is 0.6.13, but the machine defaultsThe latest version in the source is: 0.6.14, so even if we specify the -f parameter during the installation process, pip will choose to install the updated 0.6.14 from the default source (error-prone, also need to compile, very troublesome).

• Our suggested installation method:

When installing torch-sparse, specify the version and the official source address, so that the installation will be installed from the official source download package. The official source is the whl installation package that has been compiled in the corresponding environment, which can avoid the compilation process., the installation is faster, and there are no unnecessary errors.

Get the following installation instructions (note that I am CUDA11.3 Pytorch1.10.0, if you have other versions, please change the torch and cuda versions in the official source to their corresponding ones):

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 the installation is complete, install torch-geometric, torch-geometric is not in the official source, so just pip install it directly:

pip install torch-geometric

Note: If there is a version requirement, you can specify the version during installation, which will make the installation easier.

After the above installation is complete, the relevant packages can be successfully imported.

Specific use reference: official documentation