This project is carried out as part of the EDSGN-561 course at Penn State during the Spring 2021 semester. The outcomes presented within the scope of this research can be accessed here.
The project aims to investigate learning-based models trained on problem-specific datasets to facilitate creative inquiries by providing design suggestions. By doing so, we aim to address the gap between task-oriented generic solutions provided by current CAD systems and the reciprocal and problem-specific nature of the act of designing.
The repository retains the final epoch-500 checkpoint for the custom_single and custom_multi experiments, together with their configurations, final 500-epoch loss traces, labels, and 128-dimensional latent vectors. The models consume and reconstruct point clouds containing 2,048 XYZ vertices and use Chamfer reconstruction loss.
These checkpoints are research artifacts trained on project-specific procedural datasets without a separate held-out evaluation set. Exact dependency versions and training seeds were not recorded, so retraining is not expected to reproduce the checkpoint weights exactly.
The Autoencoder architecture used in this research is based on a previous research titled "Learning Representations and Generative Models for 3D Point Clouds" by Achlioptas et al. (arXiv | GitHub). Most files under src/ and the structural-loss sources under external/structural_losses/ derive from that MIT-licensed reference implementation.