DiGress: Discrete Denoising diffusion for graph generation
Authors: Clement Vignac, Igor Krawczuk, Antoine Siraudin, Bohan Wang, Volkan Cevher, Pascal Frossard
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experiments demonstrate that Di Gress achieve state-of-the-art performance, generating a high rate of realistic graphs while maintaining high degree of diversity and novelty. On the large MOSES and Guaca Mol molecular datasets, which were previously too large for one-shot models, it notably matches the performance of autoregressive models trained using expert knowledge. 7 EXPERIMENTS In our experiments, we compare the performance of Di Gress against several state-of-the-art one-shot graph generation methods on both molecular and non-molecular benchmarks. |
| Researcher Affiliation | Academia | Cl ement Vignac LTS4, EPFL Lausanne, Switzerland Igor Krawczuk LIONS, EPFL Lausanne, Switzerland Antoine Siraudin LTS4, EPFL Lausanne, Switzerland Bohan Wang LTS4, EPFL Lausanne, Switzerland Volkan Cevher LIONS, EPFL Lausanne, Switzerland Pascal Frossard LTS4, EPFL Lausanne, Switzerland |
| Pseudocode | Yes | Algorithm 1: Training Di Gress; Algorithm 2: Sampling from Di Gress; Algorithm 3: Sampling from Di Gress with discrete regressor guidance.; Algorithm 4: Training Con Gress; Algorithm 5: Sampling from Con Gress |
| Open Source Code | Yes | 1Code is available at github.com/cvignac/Di Gress. |
| Open Datasets | Yes | We then evaluate our model on the standard QM9 dataset (Wu et al., 2018) that contains molecules with up to 9 heavy atoms. ... MOSES (Polykovskiy et al., 2020), which contains small drug-like molecules, and Guaca Mol (Brown et al., 2019), which contains larger molecules. |
| Dataset Splits | Yes | We use a split of 100k molecules for training, 20k for validation and 13k for evaluating likelihood on a test set. |
| Hardware Specification | No | The paper mentions "GPUs" in the context of computation but does not provide any specific hardware details such as CPU/GPU models, processor types, or memory amounts used for running experiments. |
| Software Dependencies | No | We use Rd Kit (Landrum et al., 2006) to produce conformers of the generated graphs, and then Psi4 (Smith et al., 2020) to estimate the values of ยต and HOMO. (The paper mentions software tools like RDKit and Psi4, but does not provide specific version numbers for these dependencies.) |
| Experiment Setup | No | The paper describes the noise schedule and network architecture, but it does not provide specific experimental setup details such as learning rates, batch sizes, optimizers, or number of epochs in the main text. |