This comprehensive guide will walk you through the installation process for STGraph, along with its essential prerequisites, enabling you to train GNN models on your local machine. Let's get started.
To download STGraph on your local machine, execute the following command:
git clone https://github.com/bfGraph/STGraph.git
cd STGraph
It is highly recommended to create a dedicated Python virtual environment for running and developing with STGraph. To create a virtual environment named stgraph, you can use either conda or venv.
Using conda:
conda create --name stgraph
conda activate stgraph
To install the necessary Python packages for STGraph, run the following command. Ensure that you have activated the stgraph virtual environment before installing these packages.
pip install -r requirements.txt
You may encounter some errors or warnings, which you can ignore for now.
Installing PyTorch and PyG-T
It is recommeneded to install PyTorch and PyTorch Geometric Temporal separately. Execute the following commands
pip install torch torchvision torchaudio
pip install torch-scatter torch-sparse -f https://data.pyg.org/whl/torch-2.0.0+cu117.html
pip install torch-geometric-temporal
Before proceeding with the remaining installation steps, it is crucial to build the shared object (.so) file for the CUDA/C++ extension that implements certain dynamic and static graph representations in STGraph.
cd stgraph/graph/
./build_static.sh
./build_dynamic.sh gpma pcsr
Once the build process completes successfully, you can proceed with the remaining installation steps.
To complete the installation of the STGraph package, execute the following commands:
cd ../..
python3 -m build && pip uninstall stgraph -y && pip install dist/stgraph-1.0.0-py3-none-any.whl
After running the above commands, you can verify whether STGraph is installed correctly by executing pip show stgraph.
By executing the above instructions, you will have successfully installed STGraph along with all the required packages.
STGraph leverages CUDA Python, which encompasses a standardized collection of low-level interfaces, granting complete coverage of and access to the CUDA host APIs within Python. Prior to installing CUDA Python, ensure that your system meets the following requirements:
Note:
STGraph has undergone testing with Python CUDA version 11.7 and above. This installation guide is based on the latest release of Python CUDA, namely version 12.1.0. You can also install Python CUDA for versions 11.7 and above.
- Driver: Linux (450.80.02 or later) / Windows (456.38 or later)
- CUDA Toolkit: Versions 12.0 to 12.1
- Python: Versions 3.8 to 3.11
Execute the following command to install CUDA Python via PyPI:
pip install cuda-python
To validate the successful installation of Python CUDA, run the provided Python script located within the STGraph directory:
cd stgraph/compiler/code_gen/
python3 cuda_check.py
If the installation was successful, you should observe the following output. However, please note that the specific output may differ depending on the GPU present in your machine:
Device Property Value
───────────────────────────────────────────────────
Number of Devices 1
Name NVIDIA GeForce MX350
Compute Capability 6.1
Multiprocessor Count 5
Concurrent Threads 10240
GPU Clock 1468.0 MHz
Memory Clock 3504.0 MHz
Total Memory 2047.875 MiB
Free Memory 1641.7216796875 MiB
Note: If either the Total Memory or Free Memory shows 0,
it indicates that no context has been loaded into the device.
With this, you have now completed the installation of STGraph and verified the presence of CUDA Python on your system.
To ensure the successful installation of STGraph, let's proceed with running STGraph by training a T-GCN model on the EnglandCOVID dataset.
cd ../../..
cd benchmarking/tgcn/stgraph-dynamic/
python3 train.py --type naive --num_epochs 10
Upon executing the above command, you should observe the following output:
If you encounter any errors while attempting to train the T-GCN model, kindly raise an issue, and our team will promptly assist you.