Accelerating Molecular Graph Neural Networks via Knowledge Distillation

Authors: Filip Ekström Kelvinius, Dimitar Georgiev, Artur Toshev, Johannes Gasteiger

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To evaluate our proposed methods, we perform comprehensive benchmarking experiments on the OC20-2M [17] dataset (structure to energy and forces (S2EF) task) a large and diverse catalyst dataset; and COLL [6] a challenging molecular dynamics dataset.
Researcher Affiliation Collaboration Filip Ekström Kelvinius Linköping University filip.ekstrom@liu.se Dimitar Georgiev Imperial College London d.georgiev21@imperial.ac.uk Artur Petrov Toshev Technical University of Munich artur.toshev@tum.de Johannes Gasteiger Google Research johannesg@google.com
Pseudocode No The paper does not contain any clearly labeled pseudocode or algorithm blocks. Methods are described in prose and mathematical equations.
Open Source Code Yes Associated code is available online2. 2https://github.com/gasteigerjo/ocp/blob/main/DISTILL.md
Open Datasets Yes To evaluate our proposed methods, we perform comprehensive benchmarking experiments on the OC20-2M [17] dataset (structure to energy and forces (S2EF) task) a large and diverse catalyst dataset; and COLL [6] a challenging molecular dynamics dataset.
Dataset Splits Yes Values represent the average across the four available validation sets. Results for individual validation datasets are provided in Appendix B.
Hardware Specification Yes Models were trained on NVIDIA A100 40 GB and NVIDIA RTX A6000 48 GB GPUs, except Gem Net-OC-small which were trained on NVIDIA A100 80 GB and NVIDIA RTX A6000 48 GB. All models were trained on single GPUs, except for Sch Net when trained on OC20-2M, which required 3 GPUs. Inference throughput was profiled on A100 40 GB GPUs, with reported values representing approximate numbers averaged across three evaluations.
Software Dependencies No The paper mentions using the 'Open Catalyst Project (OCP) codebase' but does not specify version numbers for this or any other software dependencies like programming languages or libraries.
Experiment Setup Yes We provide detailed information about the hyperparameters we used for each model in Tables 5, 6, and 7. Moreover, we summarize the KD weighting factors λ we used for each model configuration in Table 8.