VectorAdam for Rotation Equivariant Geometry Optimization

Authors: Selena Zihan Ling, Nicholas Sharp, Alec Jacobson

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We now compare the performance of Adam and Vector Adam in a range of optimization problems in machine learning and traditional geometric optimization context. These experiments demonstrate the undesirable coordinate-system bias and artifacts caused by Adam s per-coordinate update rules, and show how Vector Adam solves this problem while achieving equivalent or even improved quality and convergence rate. All experiments are performed in PyTorch on an NVIDIA RTX 2080 GPU.
Researcher Affiliation Collaboration Selena Ling University of Toronto selena.ling@mail.utoronto.ca Nicholas Sharp University of Toronto nsharp@cs.toronto.edu Alec Jacobson University of Toronto, Adobe Research jacobson@cs.toronto.edu
Pseudocode Yes Algorithm 1 Vector Adam: minimizing f over r parameters, each an n-dimensional vector.
Open Source Code No The paper does not provide any concrete access to source code, repository links, or explicit statements about code release for the methodology described.
Open Datasets Yes Here, we trained a point cloud classifier from the rotation-equivariant Vector Neuron architecture [4] on the Model Net40 dataset [2]... using the widely-adopted Point Net architecture [24], in which we train the network using both Vector Adam and conventional Adam on a subset of Shape Net[33] dataset.
Dataset Splits No The paper mentions training and testing on datasets like ModelNet40 and ShapeNet, but does not specify the exact training, validation, and test splits (e.g., percentages, sample counts, or explicit splitting methodology).
Hardware Specification Yes All experiments are performed in PyTorch on an NVIDIA RTX 2080 GPU.
Software Dependencies No The paper mentions 'PyTorch' as the software used ('All experiments are performed in PyTorch'), but does not specify a version number for it or any other key software dependencies.
Experiment Setup Yes For geometry reconstruction, we run the same experiments as in [19], which optimize a sphere to match against rendering of a Nerfertiti bust. More specifically, we run the optimization 8 times with different learning rates in the range of 9e-6 and 1e-5... On the left, we optimize the ARAP energy for a triangle mesh with 200 vertices for 100 iterations with 1000 randomly rotated initial configurations.