Improving Equivariant Model Training via Constraint Relaxation

Authors: Stefanos Pertigkiozoglou, Evangelos Chatzipantazis, Shubhendu Trivedi, Kostas Daniilidis

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We provide experimental results on different state-of-the-art network architectures, demonstrating how this training framework can result in equivariant models with improved generalization performance. Our code is available at https://github.com/StefanosPert/Equivariant_Optimization_CR
Researcher Affiliation Academia Stefanos Pertigkiozoglou University of Pennsylvania pstefano@seas.upenn.edu Evangelos Chatzipantazis University of Pennsylvania vaghat@seas.upenn.edu Shubhendu Trivedi Independent shubhendu@csail.mit.edu Kostas Daniilidis University of Pennsylvania Archimedes, Athena RC kostas@cis.upenn.edu
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Our code is available at https://github.com/StefanosPert/Equivariant_Optimization_CR
Open Datasets Yes We train on the Model Net40 dataset (Chang et al., 2015), which contains 12311 point clouds from 40 different classes. [...] We use the Steerable E(3) GNN (SEGNN) (Brandstetter et al., 2021) and we train it on the task of Nbody particle simulation (Kipf et al., 2018). [...] we train Equiformer (Liao & Smidt, 2023) on the task of molecular dynamics simulations for a set of molecules provided as part of the MD17 dataset (Chmiela et al., 2017). [...] We evaluate our method on the task of 2D smoke flow prediction described in Wang et al. (2022).
Dataset Splits Yes For the choice of λreg we perform hyperparameter grid search using cross-validation with an 80%-20% split of the original training set of Model Net40 into training and validation. [...] The other 20% of the data were held out as the validation set used to evaluate the model.
Hardware Specification Yes We run all the experiments on NVIDIA A40 GPUs.
Software Dependencies No The paper mentions optimizers like 'Adam optimizer (Kingma & Ba, 2015)' and 'Adam W optimizer (Loshchilov & Hutter, 2019)' and a framework 'Phi Flow (Holl et al., 2020)', but it does not specify version numbers for general software dependencies (e.g., Python, PyTorch, CUDA, specific library versions).
Experiment Setup Yes We fix the weight of the regularization term to be λreg = 0.01 for all the experiments. [...] For the relaxed version of VNPoit Net we train for 250 epochs using the Adam optimizer (Kingma & Ba, 2015), with an initial learning rate of 10-3, that we decrease every 20 epochs by a factor of 0.7, and weight decay equal to 10-4. For the relaxed version of VN-DGCNN we train for 250 epochs using stochastic gradient descent, with an initial learning rate of 10-3, that we decrease using cosine annealing, and weight decay equal to 10-4. The batch size used was 32.