Group project (Robert Veres, Ben Bullinger, Adam Suma, Simon Storf) for the course Deep Learning (263-3210-00L) at ETH Zurich.
The pdf for the whole report/paper will be there
This is a short introduciton how to run things:
-
Create a conda env with
CONDA_OVERRIDE_CUDA=12.4 conda create --name [env name] python=3.12.7 -
Install torch and its dependencies
conda install pytorch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 pytorch-cuda=12.4 -c pytorch -c nvidia -
Install PyG
pip install torch_geometric -
Install PyG dependencies
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.4.0+cu124.html -
Install other dependencies
pip install wandb pandas tabulate -
To run gamba install transformers
pip install transformers -
Then also install
pip install ray[tune]andpip install pynndescent
To run wandb logging, add your API key to wandb.key. (Wandb logging is currently unimplemented)
To run experiments run the python file
python run_multiple_configs.pyafter specifying the config files of the experiments to run in its main function
if __name__ == "__main__":
# List of configuration files
config_files = [
"data/configs/[config_name_1].json",
"data/configs/[config_name_2].json",
...
]
run_experiments(config_files, num_trials=3) Each experiment is repeated with different random seeds for num_trials trials. The results and aggregate statistics (mean and standard deviation) across the experimental trials are saved to the results directory.
Configs used in the paper are found in here
To easily make use of Google Colab GPUs we provide the notebook google_colab.ipynb. It allows to run experiments directly from a Google Colab GPU environment. It requires that this repository is installed into /content/drive/MyDrive/DL_Project/DL_Project/.
Too many gambas might collapse the ecosystem.