rundiwu/DeepCAD

Encodes parametric CAD construction sequences into a learnable vectorized representation, enabling both autoencoding reconstruction and generative modeling via a latent GAN trained in the compressed space. Integrates OpenCASCADE (pythonocc) for CAD geometry processing and exports results to industry-standard `.step` file format compatible with commercial CAD software. Training pipeline separates the autoencoder phase from latent space generation, with evaluation metrics covering command/parameter accuracy, geometric fidelity (Chamfer distance), and coverage/MMD statistics.

706 stars. No commits in the last 6 months.