Environment Setup

We use conda to install required packages:

    conda env create -f env/equiformer_v2_torch1_ocp2.yml

This will create a new environment called equiformer_v2_torch1_ocp2.

We activate the environment:

    export PYTHONNOUSERSITE=True    # prevent using packages from base
    conda activate equiformer_v2_torch1_ocp2

Clone OC20 GitHub repository and install ocpmodels:

    git clone https://github.com/Open-Catalyst-Project/ocp.git
    cd ocp
    git checkout 5d0f59ebac5b2128142e75f1f1b72b276003b08c
    pip install -e .

The corresponding version of OCP (FAIR Chemistry) is here.

Since we reuse the original codebase of OCP (FAIR Chemistry) and provide a new trainer adding DeNS as an auxiliary task, we need to put this repository under the ocp codebase cloned above and rename to experimental:

    cd ocp
    git clone https://github.com/atomicarchitects/DeNS experimental

In this way, when we launch training under ocp, the OCP codebase will look into experimental directory and find the new trainer and models contained in this DeNS repository.