Hierarchical Neural Coding for Controllable CAD Model Generation

Authors: Xiang Xu, Pradeep Kumar Jayaraman, Joseph George Lambourne, Karl D.D. Willis, Yasutaka Furukawa

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate superior performance on conventional tasks such as unconditional generation while enabling novel interaction capabilities on conditional generation tasks.
Researcher Affiliation Collaboration 1Simon Fraser University, Canada 2Autodesk Research. Correspondence to: Xiang Xu <xuxiangx@sfu.ca>.
Pseudocode No The paper includes architectural diagrams (e.g., Figure 4) and descriptions of model components, but no formal pseudocode blocks or algorithms labeled as such.
Open Source Code Yes The code is available at https: //github.com/samxuxiang/hnc-cad.
Open Datasets Yes We use the large-scale Deep CAD dataset (Wu et al., 2021) with ground-truth sketch-and-extrude models. Deep CAD contains 178,238 sketch-and-extrude models with a split of 90% train, 5% validation, and 5% test samples.
Dataset Splits Yes Deep CAD contains 178,238 sketch-and-extrude models with a split of 90% train, 5% validation, and 5% test samples.
Hardware Specification Yes Models are trained on an Nvidia RTX A6000 GPU with a batch size of 256.
Software Dependencies No The paper mentions using the Adam W optimizer and Transformer backbones, but does not specify version numbers for any software, libraries, or frameworks (e.g., Python, PyTorch, TensorFlow, CUDA versions).
Experiment Setup Yes Input embedding dimension is 256. Feed-forward dimension is 512. Dropout rate is 0.1. Each Transformer network in the generation module has 6 layers with 8 attention heads. The codebook learning networks have 4 layers. We use the Adam W (Loshchilov & Hutter, 2018) optimizer with a learning rate of 0.001 after linear warm-up for 2000 steps.