Clifford Group Equivariant Simplicial Message Passing Networks
Authors: Cong Liu, David Ruhe, Floor Eijkelboom, Patrick Forré
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experimental results show that our method is able to outperform both equivariant and simplicial graph neural networks on a variety of geometric tasks. Our implementation is available on Git Hub. |
| Researcher Affiliation | Academia | Cong Liu12, , David Ruhe123, , Floor Eijkelboom14, Patrick Forr e12 AMLab, University of Amsterdam {c.liu4,d.ruhe,f.eijkelboom,p.d.forre}@uva.nl 1AMLab. 2AI4Science Lab. 3Anton Pannekoek Institute. 4Uv A-Bosch Delta Lab. |
| Pseudocode | Yes | Algorithm 1 Shared Simplicial Message Passing |
| Open Source Code | Yes | Our implementation is available on Git Hub. We will make the complete codebase utilized in our experiments publicly available, encompassing aspects like model architectures, data preprocessing, training configurations, hyperparameters, and evaluation methodologies. This open-source approach is aimed at ensuring straightforward reproducibility. |
| Open Datasets | Yes | Currently, the QM9 (Ramakrishnan et al., 2014) and MD17 (Chmiela et al., 2017) datasets are highly popular benchmarks for equivariant graph neural networks. We run this experiment based on the convex hull volumetric experiment of Ruhe et al. (2023a). We evaluate our models on the CMU Human Motion Capture dataset (Gross & Shi, 2001). We turn to the molecular domain with the MD17 dataset Chmiela et al. (2017). Finally, we subject our CSMPN to testing on the STATS Sport VU NBA Dataset (STATS Perform, 2023). |
| Dataset Splits | Yes | The finalized dataset contains 16384 entries for training, validation, and test sets. The processed dataset has 200 entries in the training set and 600 entries both in the validation set and test set. The training set and test set of each molecule have 5000 and 2000 instances, respectively. For Aspirin, we have 1303 validation instances. For the other molecules, we have 2000 instances to validate and tune the model performance. 8420 entries are used to train the CSMPN. We have 2806 entries in the validation test sets, respectively. |
| Hardware Specification | Yes | Averaged inference time is measured for a batch of 100 samples running on a Ge Force RTX 3090 GPU. |
| Software Dependencies | No | The paper mentions the use of 'Scipy (Virtanen et al. (2020))' and 'Adam optimizer (Kingma & Ba, 2017)' but does not provide specific version numbers for these or other key software components like Python, PyTorch, or CUDA. |
| Experiment Setup | Yes | We use three simplicial message passing layers where the message and update functions are Clifford group-equivariant MLPs with 28 hidden features. Training of CSMPN is achieved through an Adam optimizer (Kingma & Ba, 2017) with a learning rate of 1 10 3. We train baselines with 105 steps with a batch size of 512. CSMPN uses a batch size of 16 in the training process. |