Improving Graph Generation by Restricting Graph Bandwidth

Authors: Nathaniel Lee Diamant, Alex M Tseng, Kangway V. Chuang, Tommaso Biancalani, Gabriele Scalia

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

Reproducibility Variable Result LLM Response
Research Type Experimental We experimentally validate our method on both synthetic and real graphs, comparing bandwidth-constrained architectures and non-constrained baselines. ... We extensively validate our strategy on synthetic and real datasets, including molecular graphs.
Researcher Affiliation Industry 1Department of Artificial Intelligence and Machine Learning, Research and Early Development, Genentech, USA. Correspondence to: Nathaniel Diamant <diamant.nathaniel@gene.com>, Gabriele Scalia <scalia.gabriele@gene.com>. All authors are employees of Genentech, Inc. and shareholders of Roche.
Pseudocode Yes Algorithm 1 GINEStack. Algorithm 2 Graphite decoder. Algorithm 3 Modified EDP-GNN architecture.
Open Source Code Yes The implementation is made available1. ... 1https://github.com/Genentech/ bandwidth-graph-generation
Open Datasets Yes All datasets, except Peptides-func, are available through the TUDataset collection (Morris et al., 2020). Peptides-func is available in the Long Range Graph Benchmark (Dwivedi et al., 2022).
Dataset Splits Yes All models were trained for 100 epochs of 30 training batches and nine validation batches. The batch size was fixed at 32.
Hardware Specification No The paper mentions 'GPU utilization' in the context of memory usage, but does not provide specific details on the hardware used for experiments, such as GPU models, CPU types, or memory specifications.
Software Dependencies No The paper mentions 'PyTorch' and 'PyTorch Geometric' but does not provide specific version numbers for these or other software dependencies.
Experiment Setup Yes All models were trained for 100 epochs of 30 training batches and nine validation batches. The batch size was fixed at 32. The AdamW optimizer (Loshchilov & Hutter, 2019) was used with a cosine annealed learning rate (Loshchilov & Hutter, 2017).