Setup Instructions on a GPU Machine

February 14, 2022 · View on GitHub

1. Install Conda

Follow instructions on the Conda user guide. For example, to install Miniconda on an x86 Linux machine:

curl -Ls https://repo.anaconda.com/miniconda/Miniconda3-${CONDA_VER}-Linux-${conda_arch}.sh -o /tmp/Miniconda.sh &&\
bash Miniconda3-py38_4.10.3-Linux-x86_64.sh

SALIENT has been tested on Python 3.9.5.

It is highly recommended to create a new environment and do the subsequent steps therein. Example with a new environment called salient:

conda create -n salient python=3.9.5 -y
conda activate salient

You will also need to install pip for the subsequent steps:

conda install -y -c anaconda pip

2. Install PyTorch

Follow instructions on the PyTorch homepage. For example, to install on a linux machine with CUDA 11.3:

conda install pytorch torchvision torchaudio cudatoolkit=11.3 -c pytorch

SALIENT was developed against PyTorch 1.8.1 and has been tested up to PyTorch 1.10.0.

4. Install PyTorch-Geometric (PyG)

Option A: Latest PyG

To get the latest version of PyG, follow the instructions on the PyG Github page.

conda install pyg -c pyg -c conda-forge

Option B: Versions used in paper

To install the versions used in the paper, you must build from source. Make sure you have the nvcc compiler installed as part of the CUDA Toolkit. If you take this option, you can skip Step 5 below.

export FORCE_CUDA=1
pip install git+git://github.com/rusty1s/pytorch_scatter.git@2.0.7
pip install git+git://github.com/rusty1s/pytorch_cluster.git@1.5.9
pip install git+git://github.com/rusty1s/pytorch_spline_conv.git@1.2.1
pip install git+git://github.com/rusty1s/pytorch_sparse.git@master
pip install git+git://github.com/pyg-team/pytorch_geometric.git@1.7.0

SALIENT was developed against PyG 1.7.0 and has been tested up to PyG 2.0.3.

5. Install Modified PyTorch-Sparse

PyTorch-Sparse is usually installed together with PyG. We made a slight modification to support faster data transfer. We recently contributed this patch to the upstream pytorch_sparse repo that, at the time of releasing this artifact, has been merged but not yet distributed. Before a pytorch_sparse release with version greater than 0.6.12 is available, you will need to install this package from source before installing PyG. If you followed option B in Step 4, you can skip the current step.

To install from source, you will need the nvcc compiler. Uninstall the prior version if it exists and install the latest version from source:

pip uninstall torch_sparse
FORCE_CUDA=1
pip install git+git://github.com/rusty1s/pytorch_sparse.git@master

6. Install SALIENT's fast_sampler

Go to the fast_sampler folder and install it:

cd fast_sampler
python setup.py install

To check that it is properly installed, start python and type:

>>> import torch
>>> import fast_sampler
>>> help(fast_sampler)

You should see information of the package.

Note: Compilation requires a C++ compiler that supports C++17 (e.g., gcc >= 7).

7. Install Other Dependencies

conda install prettytable -c conda-forge