1. Assuming you have CUDA installed, check the version with nvidia-smi. Then install pytorch with a CUDA version <= whatever nvidia-smi reports (i.e. if your CUDA version is 11.0, install torch for CUDA 10.4, not 11.1). See the install instructions here for more details. Create and setup a new conda environment:

	conda create -n hgcal-ml python=3.9
	conda activate hgcal-ml
	conda install pytorch cudatoolkit=11.1 -c pytorch -c nvidia

Optionally, torchvision and torchaudio can also be installed.

2. Get torch geometric. This part used to be harder but nowadays the following seems to work quite well:

    conda install pytorch-geometric -c rusty1s -c conda-forge

3. Get the pytorch_cmspepr package:

    git clone https://github.com/cms-pepr/pytorch_cmspepr.git
    pip install -e pytorch_cmspepr

4. For now, clone this repo to use the latest gravnet model:

	git clone https://github.com/naodell/hgcal_gravnet_model.git